Journal of Machine Learning Research 21 (2020) 1-67 Submitted 1/20; Revised 6/20; Published 6/20
Exploring the Limits of Transfer Learning with a Unified
Text-to-Text Transformer
Colin Raffel∗ craffel@gmail.com
Noam Shazeer∗ noam@google.com
Adam Roberts∗ adarob@google.com
Katherine Lee∗ katherinelee@google.com
arXiv:1910.10683v4 [cs.LG] 19 Sep 2023
Sharan Narang sharannarang@google.com
Michael Matena mmatena@google.com
Yanqi Zhou yanqiz@google.com
Wei Li mweili@google.com
Peter J. Liu peterjliu@google.com
Google, Mountain View, CA 94043, USA
Editor: Ivan Titov
Abstract
Transfer learning, where a model is first pre-trained on a data-rich task before being fine-
tuned on a downstream task, has emerged as a powerful technique in natural language
processing (NLP). The effectiveness of transfer learning has given rise to a diversity of
approaches, methodology, and practice. In this paper, we explore the landscape of transfer
learning techniques for NLP by introducing a unified framework that converts all text-based
language problems into a text-to-text format. Our systematic study compares pre-training
objectives, architectures, unlabeled data sets, transfer approaches, and other factors on
dozens of language understanding tasks. By combining the insights from our exploration
with scale and our new “Colossal Clean Crawled Corpus”, we achieve state-of-the-art results
on many benchmarks covering summarization, question answering, text classification, and
more. To facilitate future work on transfer learning for NLP, we release our data set,
pre-trained models, and code.1
Keywords: transfer learning, natural language processing, multi-task learning, attention-
based models, deep learning
1. Introduction
Training a machine learning model to perform natural language processing (NLP) tasks
often requires that the model can process text in a way that is amenable to downstream
learning. This can be loosely viewed as developing general-purpose knowledge that allows
the model to “understand” text. This knowledge can range from low-level (e.g. the spelling
∗. Equal contribution. A description of each author’s contribution is available in Appendix A. Correspondence
to craffel@gmail.com.
1. https://github.com/google-research/text-to-text-transfer-transformer
©2020 Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei
Li, and Peter J. Liu.
License: CC-BY 4.0, see https://creativecommons.org/licenses/by/4.0/. Attribution requirements are provided at
http://jmlr.org/papers/v21/20-074.html.
Raffel, Shazeer, Roberts, Lee, Narang, Matena, Zhou, Li and Liu
or meaning of words) to high-level (e.g. that a tuba is too large to fit in most backpacks).
In modern machine learning practice, providing this knowledge is rarely done explicitly;
instead, it is often learned as part of an auxiliary task. For example, a historically common
approach is to use word vectors (Mikolov et al., 2013b,a; Pennington et al., 2014) to map
word identities to a continuous representation where, ideally, similar words map to similar
vectors. These vectors are often learned through an objective that, for example, encourages
co-occurring words to be positioned nearby in the continuous space (Mikolov et al., 2013b).
Recently, it has become increasingly common to pre-train the entire model on a data-rich
task. Ideally, this pre-training causes the model to develop general-purpose abilities and
knowledge that can then be transferred to downstream tasks. In applications of transfer
learning to computer vision (Oquab et al., 2014; Jia et al., 2014; Huh et al., 2016; Yosinski
et al., 2014), pre-training is typically done via supervised learning on a large labeled data set
like ImageNet (Russakovsky et al., 2015; Deng et al., 2009). In contrast, modern techniques
for transfer learning in NLP often pre-train using unsupervised learning on unlabeled data.
This approach has recently been used to obtain state-of-the-art results in many of the most
common NLP benchmarks (Devlin et al., 2018; Yang et al., 2019; Dong et al., 2019; Liu
et al., 2019c; Lan et al., 2019). Beyond its empirical strength, unsupervised pre-training
for NLP is particularly attractive because unlabeled text data is available en masse thanks
to the Internet—for example, the Common Crawl project2 produces about 20TB of text
data extracted from web pages each month. This is a natural fit for neural networks, which
have been shown to exhibit remarkable scalability, i.e. it is often possible to achieve better
performance simply by training a larger model on a larger data set (Hestness et al., 2017;
Shazeer et al., 2017; Jozefowicz et al., 2016; Mahajan et al., 2018; Radford et al., 2019;
Shazeer et al., 2018; Huang et al., 2018b; Keskar et al., 2019a).
This synergy has resulted in a great deal of recent work developing transfer learning
methodology for NLP, which has produced a wide landscape of pre-training objectives
(Howard and Ruder, 2018; Devlin et al., 2018; Yang et al., 2019; Dong et al., 2019), unlabeled
data sets (Yang et al., 2019; Liu et al., 2019c; Zellers et al., 2019), benchmarks (Wang et al.,
2019b, 2018; Conneau and Kiela, 2018), fine-tuning methods (Howard and Ruder, 2018;
Houlsby et al., 2019; Peters et al., 2019), and more. The rapid rate of progress and diversity
of techniques in this burgeoning field can make it difficult to compare different algorithms,
tease apart the effects of new contributions, and understand the space of existing methods for
transfer learning. Motivated by a need for more rigorous understanding, we leverage a unified
approach to transfer learning that allows us to systematically study different approaches
and push the current limits of the field.
The basic idea underlying our work is to treat every text processing problem as a
“text-to-text” problem, i.e. taking text as input and producing new text as output. This
approach is inspired by previous unifying frameworks for NLP tasks, including casting all text
problems as question answering (McCann et al., 2018), language modeling (Radford et al.,
2019), or span extraction Keskar et al. (2019b) tasks. Crucially, the text-to-text framework
allows us to directly apply the same model, objective, training procedure, and decoding
process to every task we consider. We leverage this flexibility by evaluating performance
on a wide variety of English-based NLP problems, including question answering, document
2. http://commoncrawl.org
2
Exploring the Limits of Transfer Learning
"translate English to German: That is good."
"cola sentence: The "Das ist gut."
course is jumping well."
T5
"not acceptable"
"stsb sentence1: The rhino grazed
on the grass. sentence2: A rhino
is grazing in a field." "3.8"
"summarize: state authorities "six people hospitalized after
dispatched emergency crews tuesday to a storm in attala county."
survey the damage after an onslaught
of severe weather in mississippi…"
Figure 1: A diagram of our text-to-text framework. Every task we consider—including
translation, question answering, and classification—is cast as feeding our model
text as input and training it to generate some target text. This allows us to use the
same model, loss function, hyperparameters, etc. across our diverse set of tasks. It
also provides a standard testbed for the methods included in our empirical survey.
“T5” refers to our model, which we dub the “Text-to-Text Transfer Transformer”.
summarization, and sentiment classification, to name a few. With this unified approach,
we can compare the effectiveness of different transfer learning objectives, unlabeled data
sets, and other factors, while exploring the limits of transfer learning for NLP by scaling up
models and data sets beyond what has previously been considered.
We emphasize that our goal is not to propose new methods but instead to provide a
comprehensive perspective on where the field stands. As such, our work primarily comprises
a survey, exploration, and empirical comparison of existing techniques. We also explore the
limits of current approaches by scaling up the insights from our systematic study (training
models up to 11 billion parameters) to obtain state-of-the-art results in many of the tasks
we consider. In order to perform experiments at this scale, we introduce the “Colossal Clean
Crawled Corpus” (C4), a data set consisting of hundreds of gigabytes of clean English text
scraped from the web. Recognizing that the main utility of transfer learning is the possibility
of leveraging pre-trained models in data-scarce settings, we release our code, data sets, and
pre-trained models.1
The remainder of the paper is structured as follows: In the following section, we discuss
our base model and its implementation, our procedure for formulating every text processing
problem as a text-to-text task, and the suite of tasks we consider. In Section 3, we present a
large set of experiments that explore the field of transfer learning for NLP. At the end of the
section (Section 3.7), we combine insights from our systematic study to obtain state-of-the-art
results on a wide variety of benchmarks. Finally, we provide a summary of our results and
wrap up with a look towards the future in Section 4.
3
Raffel, Shazeer, Roberts, Lee, Narang, Matena, Zhou, Li and Liu
2. Setup
Before presenting the results from our large-scale empirical study, we review the necessary
background topics required to understand our results, including the Transformer model
architecture and the downstream tasks we evaluate on. We also introduce our approach
for treating every problem as a text-to-text task and describe our “Colossal Clean Crawled
Corpus” (C4), the Common Crawl-based data set we created as a source of unlabeled text
data. We refer to our model and framework as the “Text-to-Text Transfer Transformer”
(T5).
2.1 Model
Early results on transfer learning for NLP leveraged recurrent neural networks (Peters
et al., 2018; Howard and Ruder, 2018), but it has recently become more common to use
models based on the “Transformer” architecture (Vaswani et al., 2017). The Transformer
was initially shown to be effective for machine translation, but it has subsequently been
used in a wide variety of NLP settings (Radford et al., 2018; Devlin et al., 2018; McCann
et al., 2018; Yu et al., 2018). Due to its increasing ubiquity, all of the models we study are
based on the Transformer architecture. Apart from the details mentioned below and the
variants we explore in Section 3.2, we do not deviate significantly from this architecture as
originally proposed. Instead of providing a comprehensive definition of this model, we refer
the interested reader to the original paper (Vaswani et al., 2017) or follow-up tutorials3,4 for
a more detailed introduction.
The primary building block of the Transformer is self-attention (Cheng et al., 2016).
Self-attention is a variant of attention (Graves, 2013; Bahdanau et al., 2015) that processes
a sequence by replacing each element by a weighted average of the rest of the sequence.
The original Transformer consisted of an encoder-decoder architecture and was intended
for sequence-to-sequence (Sutskever et al., 2014; Kalchbrenner et al., 2014) tasks. It has
recently also become common to use models consisting of a single Transformer layer stack,
with varying forms of self-attention used to produce architectures appropriate for language
modeling (Radford et al., 2018; Al-Rfou et al., 2019) or classification and span prediction
tasks (Devlin et al., 2018; Yang et al., 2019). We empirically explore these architectural
variants in Section 3.2.
Overall, our encoder-decoder Transformer implementation closely follows its originally-
proposed form (Vaswani et al., 2017). First, an input sequence of tokens is mapped to
a sequence of embeddings, which is then passed into the encoder. The encoder consists
of a stack of “blocks”, each of which comprises two subcomponents: a self-attention layer
followed by a small feed-forward network. Layer normalization (Ba et al., 2016) is applied to
the input of each subcomponent. We use a simplified version of layer normalization where
the activations are only rescaled and no additive bias is applied. After layer normalization,
a residual skip connection (He et al., 2016) adds each subcomponent’s input to its output.
Dropout (Srivastava et al., 2014) is applied within the feed-forward network, on the skip
connection, on the attention weights, and at the input and output of the entire stack. The
decoder is similar in structure to the encoder except that it includes a standard attention
3. http://nlp.seas.harvard.edu/2018/04/03/attention.html
4. http://jalammar.github.io/illustrated-transformer/
4
Exploring the Limits of Transfer Learning
mechanism after each self-attention layer that attends to the output of the encoder. The
self-attention mechanism in the decoder also uses a form of autoregressive or causal self-
attention, which only allows the model to attend to past outputs. The output of the final
decoder block is fed into a dense layer with a softmax output, whose weights are shared with
the input embedding matrix. All attention mechanisms in the Transformer are split up into
independent “heads” whose outputs are concatenated before being further processed.
Since self-attention is order-independent (i.e. it is an operation on sets), it is common
to provide an explicit position signal to the Transformer. While the original Transformer
used a sinusoidal position signal or learned position embeddings, it has recently become
more common to use relative position embeddings (Shaw et al., 2018; Huang et al., 2018a).
Instead of using a fixed embedding for each position, relative position embeddings produce
a different learned embedding according to the offset between the “key” and “query” being
compared in the self-attention mechanism. We use a simplified form of position embeddings
where each “embedding” is simply a scalar that is added to the corresponding logit used
for computing the attention weights. For efficiency, we also share the position embedding
parameters across all layers in our model, though within a given layer each attention head
uses a different learned position embedding. Typically, a fixed number of embeddings are
learned, each corresponding to a range of possible key-query offsets. In this work, we use 32
embeddings for all of our models with ranges that increase in size logarithmically up to an
offset of 128 beyond which we assign all relative positions to the same embedding. Note
that a given layer is insensitive to relative position beyond 128 tokens, but subsequent layers
can build a sensitivity to larger offsets by combining local information from previous layers.
To summarize, our model is roughly equivalent to the original Transformer proposed by
Vaswani et al. (2017) with the exception of removing the Layer Norm bias, placing the layer
normalization outside the residual path, and using a different position embedding scheme.
Since these architectural changes are orthogonal to the experimental factors we consider in
our empirical survey of transfer learning, we leave the ablation of their impact for future
work.
As part of our study, we experiment with the scalability of these models, i.e. how their
performance changes as they are made to have more parameters or layers. Training large
models can be non-trivial since they might not fit on a single machine and require a great deal
of computation. As a result, we use a combination of model and data parallelism and train
models on “slices” of Cloud TPU Pods.5 TPU pods are are multi-rack ML supercomputers
that contain 1,024 TPU v3 chips connected via a high-speed 2D mesh interconnect with
supporting CPU host machines. We leverage the Mesh TensorFlow library (Shazeer et al.,
2018) for ease of implementation of both model parallelism and data parallelism (Krizhevsky,
2014).
2.2 The Colossal Clean Crawled Corpus
Much of the previous work on transfer learning for NLP makes use of large unlabeled data
sets for unsupervised learning. In this paper, we are interested in measuring the effect of the
quality, characteristics, and size of this unlabeled data. To generate data sets that satisfy
our needs, we leverage Common Crawl as a source of text scraped from the web. Common
5. https://cloud.google.com/tpu/
5
Raffel, Shazeer, Roberts, Lee, Narang, Matena, Zhou, Li and Liu
Crawl has previously been used as a source of text data for NLP, for example to train an
n-gram language model (Buck et al., 2014), as training data for commonsense reasoning
(Trinh and Le, 2018), for mining parallel texts for machine translation (Smith et al., 2013),
as a pre-training data set (Grave et al., 2018; Zellers et al., 2019; Liu et al., 2019c), and
even simply as a giant text corpus for testing optimizers (Anil et al., 2019).
Common Crawl is a publicly-available web archive that provides “web extracted text”
by removing markup and other non-text content from the scraped HTML files. This process
produces around 20TB of scraped text data each month. Unfortunately, the majority of the
resulting text is not natural language. Instead, it largely comprises gibberish or boiler-plate
text like menus, error messages, or duplicate text. Furthermore, a good deal of the scraped
text contains content that is unlikely to be helpful for any of the tasks we consider (offensive
language, placeholder text, source code, etc.). To address these issues, we used the following
heuristics for cleaning up Common Crawl’s web extracted text:
• We only retained lines that ended in a terminal punctuation mark (i.e. a period,
exclamation mark, question mark, or end quotation mark).
• We discarded any page with fewer than 3 sentences and only retained lines that
contained at least 5 words.
• We removed any page that contained any word on the “List of Dirty, Naughty, Obscene
or Otherwise Bad Words”.6
• Many of the scraped pages contained warnings stating that Javascript should be
enabled so we removed any line with the word Javascript.
• Some pages had placeholder “lorem ipsum” text; we removed any page where the
phrase “lorem ipsum” appeared.
• Some pages inadvertently contained code. Since the curly bracket “{” appears in
many programming languages (such as Javascript, widely used on the web) but not in
natural text, we removed any pages that contained a curly bracket.
• Since some of the scraped pages were sourced from Wikipedia and had citation markers
(e.g. [1], [citation needed], etc.), we removed any such markers.
• Many pages had boilerplate policy notices, so we removed any lines containing the
strings “terms of use”, “privacy policy”, “cookie policy”, “uses cookies”, “use of
cookies”, or “use cookies”.
• To deduplicate the data set, we discarded all but one of any three-sentence span
occurring more than once in the data set.
Additionally, since most of our downstream tasks are focused on English-language text,
we used langdetect7 to filter out any pages that were not classified as English with a
probability of at least 0.99. Our heuristics are inspired by past work on using Common
6. https://github.com/LDNOOBW/List-of-Dirty-Naughty-Obscene-and-Otherwise-Bad-Words
7. https://pypi.org/project/langdetect/
6
Exploring the Limits of Transfer Learning
Crawl as a source of data for NLP: For example, Grave et al. (2018) also filter text using an
automatic language detector and discard short lines and Smith et al. (2013); Grave et al.
(2018) both perform line-level deduplication. However, we opted to create a new data set
because prior data sets use a more limited set of filtering heuristics, are not publicly available,
and/or are different in scope (e.g. are limited to News data (Zellers et al., 2019; Liu et al.,
2019c), comprise only Creative Commons content (Habernal et al., 2016), or are focused on
parallel training data for machine translation (Smith et al., 2013)).
To assemble our base data set, we downloaded the web extracted text from April 2019
and applied the aforementioned filtering. This produces a collection of text that is not only
orders of magnitude larger than most data sets used for pre-training (about 750 GB) but also
comprises reasonably clean and natural English text. We dub this data set the “Colossal
Clean Crawled Corpus” (or C4 for short) and release it as part of TensorFlow Datasets.8
We consider the impact of using various alternative versions of this data set in Section 3.4.
2.3 Downstream Tasks
Our goal in this paper is to measure general language learning abilities. As such, we study
downstream performance on a diverse set of benchmarks, including machine translation,
question answering, abstractive summarization, and text classification. Specifically, we
measure performance on the GLUE and SuperGLUE text classification meta-benchmarks;
CNN/Daily Mail abstractive summarization; SQuAD question answering; and WMT English
to German, French, and Romanian translation. All data was sourced from TensorFlow
Datasets.9
GLUE (Wang et al., 2018) and SuperGLUE (Wang et al., 2019b) each comprise a
collection of text classification tasks meant to test general language understanding abilities:
• Sentence acceptability judgment (CoLA (Warstadt et al., 2018))
• Sentiment analysis (SST-2 (Socher et al., 2013))
• Paraphrasing/sentence similarity (MRPC (Dolan and Brockett, 2005), STS-B (Cer
et al., 2017), QQP (Iyer et al., 2017))
• Natural language inference (MNLI (Williams et al., 2017), QNLI (Rajpurkar et al.,
2016), RTE (Dagan et al., 2005), CB (De Marneff et al., 2019))
• Coreference resolution (WNLI and WSC (Levesque et al., 2012))
• Sentence completion (COPA (Roemmele et al., 2011))
• Word sense disambiguation (WIC (Pilehvar and Camacho-Collados, 2018))
• Question answering (MultiRC (Khashabi et al., 2018), ReCoRD (Zhang et al., 2018),
BoolQ (Clark et al., 2019))
8. https://www.tensorflow.org/datasets/catalog/c4
9. https://www.tensorflow.org/datasets
7
Raffel, Shazeer, Roberts, Lee, Narang, Matena, Zhou, Li and Liu
We use the data sets as distributed by the GLUE and SuperGLUE benchmarks. For
simplicity, when fine-tuning we treat all of the tasks in the GLUE benchmark (and similarly
for SuperGLUE) as a single task by concatenating all of the constituent data sets. As
suggested by Kocijan et al. (2019) we also include the Definite Pronoun Resolution (DPR)
data set (Rahman and Ng, 2012) in the combined SuperGLUE task.
The CNN/Daily Mail (Hermann et al., 2015) data set was introduced as a question-
answering task but was adapted for text summarization by Nallapati et al. (2016); we
use the non-anonymized version from See et al. (2017) as an abstractive summarization
task. SQuAD (Rajpurkar et al., 2016) is a common question-answering benchmark. In our
experiments, the model is fed the question and its context and asked to generate the answer
token-by-token. For WMT English to German, we use the same training data as (Vaswani
et al., 2017) (i.e. News Commentary v13, Common Crawl, Europarl v7) and newstest2013
as a validation set (Bojar et al., 2014). For English to French, we use the standard training
data from 2015 and newstest2014 as a validation set (Bojar et al., 2015). For English to
Romanian, which is a standard lower-resource machine translation benchmark, we use the
train and validation sets from WMT 2016 (Bojar et al., 2016). Note that we only pre-train
on English data, so in order to learn to translate a given model will need to learn to generate
text in a new language.
2.4 Input and Output Format
In order to train a single model on the diverse set of tasks described above, we cast all of
the tasks we consider into a “text-to-text” format—that is, a task where the model is fed
some text for context or conditioning and is then asked to produce some output text. This
framework provides a consistent training objective both for pre-training and fine-tuning.
Specifically, the model is trained with a maximum likelihood objective (using “teacher forcing”
(Williams and Zipser, 1989)) regardless of the task. To specify which task the model should
perform, we add a task-specific (text) prefix to the original input sequence before feeding it
to the model.
As an example, to ask the model to translate the sentence “That is good.” from English
to German, the model would be fed the sequence “translate English to German: That is
good.” and would be trained to output “Das ist gut.” For text classification tasks, the
model simply predicts a single word corresponding to the target label. For example, on the
MNLI benchmark (Williams et al., 2017) the goal is to predict whether a premise implies
(“entailment”), contradicts (“contradiction”), or neither (“neutral”) a hypothesis. With
our preprocessing, the input sequence becomes “mnli premise: I hate pigeons. hypothesis:
My feelings towards pigeons are filled with animosity.” with the corresponding target word
“entailment”. Note that an issue arises if our model outputs text on a text classification
task that does not correspond to any of the possible labels (for example if the model
outputs “hamburger” when the only possible labels for a task were “entailment”, “neutral”,
or “contradiction”). In this case, we always count the model’s output as wrong, though we
never observed this behavior in any of our trained models. Note that the choice of text prefix
used for a given task is essentially a hyperparameter; we found that changing the exact
wording of the prefix had limited impact and so did not perform extensive experiments into
different prefix choices. A diagram of our text-to-text framework with a few input/output
8
Exploring the Limits of Transfer Learning
examples is shown in Figure 1. We provide full examples of preprocessed inputs for every
task we studied in Appendix D.
Our text-to-text framework follows previous work that casts multiple NLP tasks into
a common format: McCann et al. (2018) propose the “Natural Language Decathlon”, a
benchmark that uses a consistent question-answering format for a suite of ten NLP tasks.
The Natural Language Decathlon also stipulates that all models must be multi-task, i.e.
are able to simultaneously tackle all of the tasks at once. We instead allow for separately
fine-tuning the model on each individual task and use short task prefixes instead of an explicit
question-answer format. Radford et al. (2019) evaluate the zero-shot learning capabilities of
language models by feeding some input to the model as a prefix and then autoregressively
sampling an output. For example, automatic summarization is done by feeding in a document
followed by the text “TL;DR:” (short for “too long, didn’t read”, a common abbreviation)
and then the summary is predicted via autoregressive decoding. We mainly consider models
that explicitly process an input with an encoder before generating an output with a separate
decoder and we focus on transfer learning rather than zero-shot learning. Finally, Keskar
et al. (2019b) unify many NLP tasks as “span extraction”, where text corresponding to
possible output choices are appended to the input and the model is trained to extract the
input span corresponding to the correct choice. In contrast, our framework also allows for
generative tasks like machine translation and abstractive summarization where it is not
possible to enumerate all possible output choices.
We were able to straightforwardly cast all of the tasks we considered into a text-to-text
format with the exception of STS-B, which is a regression task where the goal is to predict
a similarity score between 1 and 5. We found that most of these scores were annotated
in increments of 0.2, so we simply rounded any score to the nearest increment of 0.2 and
converted the result to a literal string representation of the number (e.g. the floating-point
value 2.57 would be mapped to the string “2.6”). At test time, if the model outputs a
string corresponding to a number between 1 and 5, we convert it to a floating-point value;
otherwise, we treat the model’s prediction as incorrect. This effectively recasts the STS-B
regression problem as a 21-class classification problem.
Separately, we also convert the Winograd tasks (WNLI from GLUE, WSC from Super-
GLUE, and the DPR data set we add to SuperGLUE) into a simpler format that is more
amenable to the text-to-text framework. Examples from the Winograd tasks consist of a
text passage containing an ambiguous pronoun that could refer to more than one of the noun
phrases in the passage. For example, the passage might be “The city councilmen refused
the demonstrators a permit because they feared violence.”, which contains the ambiguous
pronoun “they” that could refer to “city councilmen” or “demonstrators”. We cast the WNLI,
WSC, and DPR tasks as text-to-text problems by highlighting the ambiguous pronoun in
the text passage and asking the model to predict the noun that it refers to. The example
mentioned above would be transformed to the input “The city councilmen refused the
demonstrators a permit because *they* feared violence.” and the model would be trained to
predict the target text “The city councilmen”.
For WSC, examples contain the passage, the ambiguous pronoun, a candidate noun,
and a True/False label reflecting whether the candidate matches the pronoun (ignoring any
articles). We only train on examples with a “True” label since we do not know the correct
noun targets for examples with a “False” label. For evaluation, we assign a “True” label if
9
Raffel, Shazeer, Roberts, Lee, Narang, Matena, Zhou, Li and Liu
the words in the model’s output are a subset of the words in the candidate noun phrase
(or vice versa) and assign a “False” label otherwise. This removes roughly half of the WSC
training set, but the DPR data set adds about 1,000 pronoun resolution examples. Examples
from DPR are annotated with the correct referent noun, making it easy to use this data set
in the format listed above.
The WNLI training and validation sets have a significant overlap with the WSC training
set. To avoid leaking validation examples into our training data (a particular issue in the
multi-task experiments of Section 3.5.2), we therefore never train on WNLI and never report
results on the WNLI validation set. Omitting results on the WNLI validation set is standard
practice (Devlin et al., 2018) due to the fact that it is “adversarial” with respect to the
training set, i.e. validation examples are all slightly-perturbed versions of training examples
with the opposite label. As such, we do not include WNLI in the average GLUE score
whenever we report on the validation set (all sections except Section 3.7 where results
are presented on the test sets). Converting examples from WNLI to the “referent noun
prediction” variant described above is a little more involved; we describe this process in
Appendix B.
3. Experiments
Recent advances in transfer learning for NLP have come from a wide variety of developments,
such as new pre-training objectives, model architectures, unlabeled data sets, and more.
In this section, we carry out an empirical survey of these techniques in hopes of teasing
apart their contribution and significance. We then combine the insights gained to attain
state-of-the-art in many of the tasks we consider. Since transfer learning for NLP is a rapidly
growing area of research, it is not feasible for us to cover every possible technique or idea
in our empirical study. For a broader literature review, we recommend a recent survey by
Ruder et al. (2019).
We systematically study these contributions by taking a reasonable baseline (described
in Section 3.1) and altering one aspect of the setup at a time. For example, in Section 3.3
we measure the performance of different unsupervised objectives while keeping the rest of
our experimental pipeline fixed. This “coordinate ascent” approach might miss second-order
effects (for example, some particular unsupervised objective may work best on a model
larger than our baseline setting), but performing a combinatorial exploration of all of the
factors in our study would be prohibitively expensive. In future work, we expect it could be
fruitful to more thoroughly consider combinations of the approaches we study.
Our goal is to compare a variety of different approaches on a diverse set of tasks while
keeping as many factors fixed as possible. In order to satisfy this aim, in some cases we do
not exactly replicate existing approaches. For example, “encoder-only” models like BERT
(Devlin et al., 2018) are designed to produce a single prediction per input token or a single
prediction for an entire input sequence. This makes them applicable for classification or span
prediction tasks but not for generative tasks like translation or abstractive summarization.
As such, none of the model architectures we consider are identical to BERT or consist of an
encoder-only structure. Instead, we test approaches that are similar in spirit—for example,
we consider an analogous objective to BERT’s “masked language modeling” objective in
10
Exploring the Limits of Transfer Learning
Section 3.3 and we consider a model architecture that behaves similarly to BERT on text
classification tasks in Section 3.2.
After outlining our baseline experimental setup in the following subsection, we undertake
an empirical comparison of model architectures (Section 3.2), unsupervised objectives
(Section 3.3), pre-training data sets (Section 3.4), transfer approaches (Section 3.5), and
scaling (Section 3.6). At the culmination of this section, we combine insights from our study
with scale to obtain state-of-the-art results in many tasks we consider (Section 3.7).
3.1 Baseline
Our goal for our baseline is to reflect typical, modern practice. We pre-train a standard
Transformer (described in Section 2.1) using a simple denoising objective and then separately
fine-tune on each of our downstream tasks. We describe the details of this experimental
setup in the following subsections.
3.1.1 Model
For our model, we use a standard encoder-decoder Transformer as proposed by Vaswani et al.
(2017). While many modern approaches to transfer learning for NLP use a Transformer
architecture consisting of only a single “stack” (e.g. for language modeling (Radford et al.,
2018; Dong et al., 2019) or classification and span prediction (Devlin et al., 2018; Yang et al.,
2019)), we found that using a standard encoder-decoder structure achieved good results
on both generative and classification tasks. We explore the performance of different model
architectures in Section 3.2.
Our baseline model is designed so that the encoder and decoder are each similar in
size and configuration to a “BERTBASE ” (Devlin et al., 2018) stack. Specifically, both the
encoder and decoder consist of 12 blocks (each block comprising self-attention, optional
encoder-decoder attention, and a feed-forward network). The feed-forward networks in each
block consist of a dense layer with an output dimensionality of dff = 3072 followed by a
ReLU nonlinearity and another dense layer. The “key” and “value” matrices of all attention
mechanisms have an inner dimensionality of dkv = 64 and all attention mechanisms have 12
heads. All other sub-layers and embeddings have a dimensionality of dmodel = 768. In total,
this results in a model with about 220 million parameters. This is roughly twice the number
of parameters of BERTBASE since our baseline model contains two layer stacks instead of
one. For regularization, we use a dropout probability of 0.1 everywhere dropout is applied
in the model.
3.1.2 Training
As described in Section 2.4, all tasks are formulated as text-to-text tasks. This allows us to
always train using standard maximum likelihood, i.e. using teacher forcing (Williams and
Zipser, 1989) and a cross-entropy loss. For optimization, we use AdaFactor (Shazeer and
Stern, 2018). At test time, we use greedy decoding (i.e. choosing the highest-probability
logit at every timestep).
We pre-train each model for 219 = 524,288 steps on C4 before fine-tuning. We use a
maximum sequence length of 512 and a batch size of 128 sequences. Whenever possible,
11
Raffel, Shazeer, Roberts, Lee, Narang, Matena, Zhou, Li and Liu
we “pack” multiple sequences into each entry of the batch10 so that our batches contain
roughly 216 = 65,536 tokens. In total, this batch size and number of steps corresponds
to pre-training on 235 ≈ 34B tokens. This is considerably less than BERT (Devlin et al.,
2018), which used roughly 137B tokens, or RoBERTa (Liu et al., 2019c), which used roughly
2.2T tokens. Using only 235 tokens results in a reasonable computational budget while still
providing a sufficient amount of pre-training for acceptable performance. We consider the
effect of pre-training for more steps in Sections 3.6 and 3.7. Note that 235 tokens only covers
a fraction of the entire C4 data set, so we never repeat any data during pre-training.
During pre-training, we use an “inverse square root” learning rate schedule: 1 max(n, k)
p
where n is the current training iteration and k is the number of warm-up steps (set to 104
in all of our experiments). This sets a constant learning rate of 0.01 for the first 104 steps,
then exponentially decays the learning rate until pre-training is over. We also experimented
with using a triangular learning rate (Howard and Ruder, 2018), which produced slightly
better results but requires knowing the total number of training steps ahead of time. Since
we will be varying the number of training steps in some of our experiments, we opt for the
more generic inverse square root schedule.
Our models are fine-tuned for 218 = 262,144 steps on all tasks. This value was chosen
as a trade-off between the high-resource tasks (i.e. those with large data sets), which
benefit from additional fine-tuning, and low-resource tasks (smaller data sets), which overfit
quickly. During fine-tuning, we continue using batches with 128 length-512 sequences (i.e.
216 tokens per batch). We use a constant learning rate of 0.001 when fine-tuning. We save
a checkpoint every 5,000 steps and report results on the model checkpoint corresponding
to the highest validation performance. For models fine-tuned on multiple tasks, we choose
the best checkpoint for each task independently. For all of the experiments except those in
Section 3.7, we report results in the validation set to avoid performing model selection on
the test set.
3.1.3 Vocabulary
We use SentencePiece (Kudo and Richardson, 2018) to encode text as WordPiece tokens
(Sennrich et al., 2015; Kudo, 2018). For all experiments, we use a vocabulary of 32,000
wordpieces. Since we ultimately fine-tune our model on English to German, French, and
Romanian translation, we also require that our vocabulary covers these non-English languages.
To address this, we classified pages from the Common Crawl scrape used in C4 as German,
French, and Romanian. Then, we trained our SentencePiece model on a mixture of 10 parts
of English C4 data with 1 part each of data classified as German, French or Romanian.
This vocabulary was shared across both the input and output of our model. Note that
our vocabulary makes it so that our model can only process a predetermined, fixed set of
languages.
3.1.4 Unsupervised Objective
Leveraging unlabeled data to pre-train our model necessitates an objective that does not
require labels but (loosely speaking) teaches the model generalizable knowledge that will be
10. https://www.pydoc.io/pypi/tensor2tensor-1.5.7/autoapi/data_generators/generator_utils/
index.html#data_generators.generator_utils.pack_examples
12
Exploring the Limits of Transfer Learning
<X> <Y>
<X> <Y> <Z>
Figure 2: Schematic of the objective we use in our baseline model. In this example, we
process the sentence “Thank you for inviting me to your party last week.” The
words “for”, “inviting” and “last” (marked with an ×) are randomly chosen for
corruption. Each consecutive span of corrupted tokens is replaced by a sentinel
token (shown as <X> and <Y>) that is unique over the example. Since “for” and
“inviting” occur consecutively, they are replaced by a single sentinel <X>. The
output sequence then consists of the dropped-out spans, delimited by the sentinel
tokens used to replace them in the input plus a final sentinel token <Z>.
useful in downstream tasks. Preliminary work that applied the transfer learning paradigm
of pre-training and fine-tuning all of the model’s parameters to NLP problems used a
causal language modeling objective for pre-training (Dai and Le, 2015; Peters et al., 2018;
Radford et al., 2018; Howard and Ruder, 2018). However, it has recently been shown that
“denoising” objectives (Devlin et al., 2018; Taylor, 1953) (also called “masked language
modeling”) produce better performance and as a result they have quickly become standard.
In a denoising objective, the model is trained to predict missing or otherwise corrupted
tokens in the input. Inspired by BERT’s “masked language modeling” objective and the
“word dropout” regularization technique (Bowman et al., 2015), we design an objective that
randomly samples and then drops out 15% of tokens in the input sequence. All consecutive
spans of dropped-out tokens are replaced by a single sentinel token. Each sentinel token
is assigned a token ID that is unique to the sequence. The sentinel IDs are special tokens
which are added to our vocabulary and do not correspond to any wordpiece. The target
then corresponds to all of the dropped-out spans of tokens, delimited by the same sentinel
tokens used in the input sequence plus a final sentinel token to mark the end of the target
sequence. Our choices to mask consecutive spans of tokens and only predict dropped-out
tokens were made to reduce the computational cost of pre-training. We perform thorough
investigation into pre-training objectives in Section 3.3. An example of the transformation
resulting from applying this objective is shown in Figure 2. We empirically compare this
objective to many other variants in Section 3.3.
3.1.5 Baseline Performance
In this section, we present results using the baseline experimental procedure described above
to get a sense of what kind of performance to expect on our suite of downstream tasks.
Ideally, we would repeat every experiment in our study multiple times to get a confidence
interval on our results. Unfortunately, this would be prohibitively expensive due to the large
13
Raffel, Shazeer, Roberts, Lee, Narang, Matena, Zhou, Li and Liu
GLUE CNNDM SQuAD SGLUE EnDe EnFr EnRo
⋆ Baseline average 83.28 19.24 80.88 71.36 26.98 39.82 27.65
Baseline standard deviation 0.235 0.065 0.343 0.416 0.112 0.090 0.108
No pre-training 66.22 17.60 50.31 53.04 25.86 39.77 24.04
Table 1: Average and standard deviation of scores achieved by our baseline model and
training procedure. For comparison, we also report performance when training on
each task from scratch (i.e. without any pre-training) for the same number of steps
used to fine-tune the baseline model. All scores in this table (and every table in
our paper except Table 14) are reported on the validation sets of each data set.
number of experiments we run. As a cheaper alternative, we train our baseline model 10
times from scratch (i.e. with different random initializations and data set shuffling) and
assume that the variance over these runs of the base model also applies to each experimental
variant. We don’t expect most of the changes we make to have a dramatic effect on the
inter-run variance, so this should provide a reasonable indication of the significance of
different changes. Separately, we also measure the performance of training our model for 218
steps (the same number we use for fine-tuning) on all downstream tasks without pre-training.
This gives us an idea of how much pre-training benefits our model in the baseline setting.
When reporting results in the main text, we only report a subset of the scores across all
the benchmarks to conserve space and ease interpretation. For GLUE and SuperGLUE, we
report the average score across all subtasks (as stipulated by the official benchmarks) under
the headings “GLUE” and “SGLUE”. For all translation tasks, we report the BLEU score
(Papineni et al., 2002) as provided by SacreBLEU v1.3.0 (Post, 2018) with “exp” smoothing
and “intl” tokenization. We refer to scores for WMT English to German, English to French,
and English to Romanian as EnDe, EnFr, and EnRo, respectively. For CNN/Daily Mail,
we find the performance of models on the ROUGE-1-F, ROUGE-2-F, and ROUGE-L-F
metrics (Lin, 2004) to be highly correlated so we report the ROUGE-2-F score alone under
the heading “CNNDM”. Similarly, for SQuAD we find the performance of the “exact match”
and “F1” scores to be highly correlated so we report the “exact match” score alone. We
provide every score achieved on every task for all experiments in Table 16, Appendix E.
Our results tables are all formatted so that each row corresponds to a particular experi-
mental configuration with columns giving the scores for each benchmark. We will include
the mean performance of the baseline configuration in most tables. Wherever a baseline
configuration appears, we will mark it with a ⋆ (as in the first row of Table 1). We also
will boldface any score that is within two standard deviations of the maximum (best) in a
given experiment.
Our baseline results are shown in Table 1. Overall, our results are comparable to existing
models of similar size. For example, BERTBASE achieved an exact match score of 80.8
on SQuAD and an accuracy of 84.4 on MNLI-matched, whereas we achieve 80.88 and
84.24, respectively (see Table 16). Note that we cannot directly compare our baseline to
BERTBASE because ours is an encoder-decoder model and was pre-trained for roughly 1⁄4
as many steps. Unsurprisingly, we find that pre-training provides significant gains across
almost all benchmarks. The only exception is WMT English to French, which is a large
14
Exploring the Limits of Transfer Learning
enough data set that gains from pre-training tend to be marginal. We include this task in
our experiments to test the behavior of transfer learning in the high-resource regime. Since
we perform early stopping by selecting the best-performing checkpoint, the large disparity
between our baseline and “no pre-training” emphasize how much pre-training improves
performance on tasks with limited data. While we do not explicitly measure improvements
in data efficiency in this paper, we emphasize that this is one of the primary benefits of the
transfer learning paradigm.
As for inter-run variance, we find that for most tasks the standard deviation across runs
is smaller than 1% of the task’s baseline score. Exceptions to this rule include CoLA, CB,
and COPA, which are all low-resource tasks from the GLUE and SuperGLUE benchmarks.
For example, on CB our baseline model had an average F1 score of 91.22 with a standard
deviation of 3.237 (see Table 16), which may be partly due to the fact that CB’s validation
set contains only 56 examples. Note that the GLUE and SuperGLUE scores are computed
as the average of scores across the tasks comprising each benchmark. As a result, we caution
that the high inter-run variance of CoLA, CB, and COPA can make it harder to compare
models using the GLUE and SuperGLUE scores alone.
3.2 Architectures
While the Transformer was originally introduced with an encoder-decoder architecture, much
modern work on transfer learning for NLP uses alternative architectures. In this section, we
review and compare these architectural variants.
3.2.1 Model Structures
A major distinguishing factor for different architectures is the “mask” used by different
attention mechanisms in the model. Recall that the self-attention operation in a Transformer
takes a sequence as input and outputs a new sequence of the same length. Each entry of
the output sequence is produced by computing a weighted average of entries of the input
sequence. Specifically, let yi refer to the ith element of the output sequence and xj refer to
the jth entry of the input sequence. yi is computed as j wi,j xj , where wi,j is the scalar
P
weight produced by the self-attention mechanism as a function of xi and xj . The attention
mask is then used to zero out certain weights in order to constrain which entries of the input
can be attended to at a given output timestep. Diagrams of the masks we will consider are
shown in Figure 3. For example, the causal mask (Figure 3, middle) sets any wi,j to zero if
j > i.
The first model structure we consider is an an encoder-decoder Transformer, which
consists of two layer stacks: The encoder, which is fed an input sequence, and the decoder,
which produces a new output sequence. A schematic of this architectural variant is shown
in the left panel of Figure 4.
The encoder uses a “fully-visible” attention mask. Fully-visible masking allows a self-
attention mechanism to attend to any entry of the input when producing each entry of
its output. We visualize this masking pattern in Figure 3, left. This form of masking is
appropriate when attending over a “prefix”, i.e. some context provided to the model that
is later used when making predictions. BERT (Devlin et al., 2018) also uses a fully-visible
masking pattern and appends a special “classification” token to the input. BERT’s output
15
Raffel, Shazeer, Roberts, Lee, Narang, Matena, Zhou, Li and Liu
y5 y5 y5
y4 y4 y4
y3 y3 y3
y2 y2 y2
y1 y1 y1
x1 x2 x3 x4 x5 x1 x2 x3 x4 x5 x1 x2 x3 x4 x5
Figure 3: Matrices representing different attention mask patterns. The input and output
of the self-attention mechanism are denoted x and y respectively. A dark cell
at row i and column j indicates that the self-attention mechanism is allowed to
attend to input element j at output timestep i. A light cell indicates that the
self-attention mechanism is not allowed to attend to the corresponding i and j
combination. Left: A fully-visible mask allows the self-attention mechanism to
attend to the full input at every output timestep. Middle: A causal mask prevents
the ith output element from depending on any input elements from “the future”.
Right: Causal masking with a prefix allows the self-attention mechanism to use
fully-visible masking on a portion of the input sequence.
at the timestep corresponding to the classification token is then used to make a prediction
for classifying the input sequence.
The self-attention operations in the Transformer’s decoder use a “causal” masking pattern.
When producing the ith entry of the output sequence, causal masking prevents the model
from attending to the jth entry of the input sequence for j > i. This is used during training
so that the model can’t “see into the future” as it produces its output. An attention matrix
for this masking pattern is shown in Figure 3, middle.
The decoder in an encoder-decoder Transformer is used to autoregressively produce an
output sequence. That is, at each output timestep, a token is sampled from the model’s
predicted distribution and the sample is fed back into the model to produce a prediction for
the next output timestep, and so on. As such, a Transformer decoder (without an encoder)
can be used as a language model (LM), i.e. a model trained solely for next-step prediction
(Liu et al., 2018; Radford et al., 2018; Al-Rfou et al., 2019). This constitutes the second
model structure we consider. A schematic of this architecture is shown in Figure 4, middle.
In fact, early work on transfer learning for NLP used this architecture with a language
modeling objective as a pre-training method (Radford et al., 2018).
Language models are typically used for compression or sequence generation (Graves,
2013). However, they can also be used in the text-to-text framework simply by concatenating
the inputs and targets. As an example, consider the case of English to German translation:
If we have a training datapoint with input sentence “That is good.” and target “Das ist
gut.”, we would simply train the model on next-step prediction over the concatenated input
sequence “translate English to German: That is good. target: Das ist gut.” If we wanted to
16
Exploring the Limits of Transfer Learning
y1 y2 .
Language model Prefix LM
Decoder
x2 x3 y1 y2 . x2 x3 y1 y2 .
Encoder
x1 x2 x3 y1 y2 x1 x2 x3 y1 y2
x1 x2 x3 x4
Figure 4: Schematics of the Transformer architecture variants we consider. In this diagram,
blocks represent elements of a sequence and lines represent attention visibility.
Different colored groups of blocks indicate different Transformer layer stacks. Dark
grey lines correspond to fully-visible masking and light grey lines correspond
to causal masking. We use “.” to denote a special end-of-sequence token that
represents the end of a prediction. The input and output sequences are represented
as x and y respectively. Left: A standard encoder-decoder architecture uses fully-
visible masking in the encoder and the encoder-decoder attention, with causal
masking in the decoder. Middle: A language model consists of a single Transformer
layer stack and is fed the concatenation of the input and target, using a causal
mask throughout. Right: Adding a prefix to a language model corresponds to
allowing fully-visible masking over the input.
obtain the model’s prediction for this example, the model would be fed the prefix “translate
English to German: That is good. target:” and would be asked to generate the remainder
of the sequence autoregressively. In this way, the model can predict an output sequence
given an input, which satisfies the needs of text-to-text tasks. This approach was recently
used to show that language models can learn to perform some text-to-text tasks without
supervision (Radford et al., 2019).
A fundamental and frequently cited drawback of using a language model in the text-
to-text setting is that causal masking forces the model’s representation of the ith entry of
the input sequence to only depend on the entries up until i. To see why this is potentially
disadvantageous, consider the text-to-text framework where the model is provided with a
prefix/context before being asked to make predictions (e.g., the prefix is an English sentence
and the model is asked to predict the German translation). With fully causal masking, the
model’s representation of a prefix state can only depend on prior entries of the prefix. So,
when predicting an entry of the output, the model will attend to a representation of the
prefix that is unnecessarily limited. Similar arguments have been made against using a
unidirectional recurrent neural network encoder in sequence-to-sequence models (Bahdanau
et al., 2015).
17
Raffel, Shazeer, Roberts, Lee, Narang, Matena, Zhou, Li and Liu
This issue can be avoided in a Transformer-based language model simply by changing
the masking pattern. Instead of using a causal mask, we use fully-visible masking during
the prefix portion of the sequence. This masking pattern and a schematic of the resulting
“prefix LM” (the third model structure we consider) are illustrated in the rightmost panels of
Figures 3 and 4, respectively. In the English to German translation example mentioned above,
fully-visible masking would be applied to the prefix “translate English to German: That is
good. target:” and causal masking would be used during training for predicting the target
“Das ist gut.” Using a prefix LM in the text-to-text framework was originally proposed by
Liu et al. (2018). More recently, Dong et al. (2019) showed that this architecture is effective
on a wide variety of text-to-text tasks. This architecture is similar to an encoder-decoder
model with parameters shared across the encoder and decoder and with the encoder-decoder
attention replaced with full attention across the input and target sequence.
We note that when following our text-to-text framework, the prefix LM architecture
closely resembles BERT (Devlin et al., 2018) for classification tasks. To see why, consider an
example from the MNLI benchmark where the premise is “I hate pigeons.”, the hypothesis is
“My feelings towards pigeons are filled with animosity.” and the correct label is “entailment”.
To feed this example into a language model, we would transform it into the sequence “mnli
premise: I hate pigeons. hypothesis: My feelings towards pigeons are filled with animosity.
target: entailment”. In this case, the fully-visible prefix would correspond to the entire input
sequence up to the word “target:”, which can be seen as being analogous to the “classification”
token used in BERT. So, our model would have full visibility over the entire input, and then
would be tasked with making a classification by outputting the word “entailment”. It is easy
for the model to learn to output one of the valid class labels given the task prefix (“mnli” in
this case). As such, the main difference between a prefix LM and the BERT architecture is
that the classifier is simply integrated into the output layer of the Transformer decoder in
the prefix LM.
3.2.2 Comparing Different Model Structures
In the interest of experimentally comparing these architectural variants, we would like each
model we consider to be equivalent in some meaningful way. We might say that two models
are equivalent if they either have the same number of parameters or they require roughly
the same amount of computation to process a given (input-sequence, target-sequence) pair.
Unfortunately, it is not possible to compare an encoder-decoder model to a language model
architecture (comprising a single Transformer stack) according to both of these criteria
at the same time. To see why, first note an encoder-decoder model with L layers in the
encoder and L layers in the decoder has approximately the same number of parameters as a
language model with 2L layers. However, the same L + L encoder-decoder model will have
approximately the same computational cost as a language model with only L layers. This
is a consequence of the fact that the L layers in the language model must be applied to
both the input and output sequence, while the encoder is only applied to the input sequence
and the decoder is only applied to the output sequence. Note that these equivalences are
approximate—there are some extra parameters in the decoder due to the encoder-decoder
attention and there are also some computational costs in the attention layers that are
quadratic in the sequence lengths. In practice, however, we observed nearly identical step
18
Exploring the Limits of Transfer Learning
times for L-layer language models versus L + L-layer encoder-decoder models, suggesting a
roughly equivalent computational cost. Further, for the model sizes we consider, the number
of parameters in the encoder-decoder attention layers is about 10% of the total parameter
count, so we make the simplifying assumption that an L + L-layer encoder-decoder model
has the same number of parameters as an 2L-layer language model.
To provide a reasonable means of comparison, we consider multiple configurations for
our encoder-decoder model. We will refer to the number of layers and parameters in a
BERTBASE -sized layer stack as L and P , respectively. We will use M to refer to the number
of FLOPs required for an L + L-layer encoder-decoder model or L-layer decoder-only model
to process a given input-target pair. In total, we will compare:
• An encoder-decoder model with L layers in the encoder and L layers in the decoder.
This model has 2P parameters and a computation cost of M FLOPs.
• An equivalent model, but with parameters shared across the encoder and decoder,
resulting in P parameters and an M -FLOP computational cost.
• An encoder-decoder model with L/2 layers each in the encoder and decoder, giving P
parameters and an M/2-FLOP cost.
• A decoder-only language model with L layers and P parameters and a resulting
computational cost of M FLOPs.
• A decoder-only prefix LM with the same architecture (and thus the same number
of parameters and computational cost), but with fully-visible self-attention over the
input.
3.2.3 Objectives
As an unsupervised objective, we will consider both a basic language modeling objective as
well as our baseline denoising objective described in Section 3.1.4. We include the language
modeling objective due to its historic use as a pre-training objective (Dai and Le, 2015;
Ramachandran et al., 2016; Howard and Ruder, 2018; Radford et al., 2018; Peters et al.,
2018) as well as its natural fit for the language model architectures we consider. For models
that ingest a prefix before making predictions (the encoder-decoder model and prefix LM),
we sample a span of text from our unlabeled data set and choose a random point to split
it into prefix and target portions. For the standard language model, we train the model
to predict the entire span from beginning to end. Our unsupervised denoising objective is
designed for text-to-text models; to adapt it for use with a language model we concatenate
the inputs and targets as described in Section 3.2.1.
3.2.4 Results
The scores achieved by each of the architectures we compare are shown in Table 2. For
all tasks, the encoder-decoder architecture with the denoising objective performed best.
This variant has the highest parameter count (2P ) but the same computational cost as the
P -parameter decoder-only models. Surprisingly, we found that sharing parameters across the
encoder and decoder performed nearly as well. In contrast, halving the number of layers in
19
Raffel, Shazeer, Roberts, Lee, Narang, Matena, Zhou, Li and Liu
Architecture Objective Params Cost GLUE CNNDM SQuAD SGLUE EnDe EnFr EnRo
⋆ Encoder-decoder Denoising 2P M 83.28 19.24 80.88 71.36 26.98 39.82 27.65
Enc-dec, shared Denoising P M 82.81 18.78 80.63 70.73 26.72 39.03 27.46
Enc-dec, 6 layers Denoising P M/2 80.88 18.97 77.59 68.42 26.38 38.40 26.95
Language model Denoising P M 74.70 17.93 61.14 55.02 25.09 35.28 25.86
Prefix LM Denoising P M 81.82 18.61 78.94 68.11 26.43 37.98 27.39
Encoder-decoder LM 2P M 79.56 18.59 76.02 64.29 26.27 39.17 26.86
Enc-dec, shared LM P M 79.60 18.13 76.35 63.50 26.62 39.17 27.05
Enc-dec, 6 layers LM P M/2 78.67 18.26 75.32 64.06 26.13 38.42 26.89
Language model LM P M 73.78 17.54 53.81 56.51 25.23 34.31 25.38
Prefix LM LM P M 79.68 17.84 76.87 64.86 26.28 37.51 26.76
Table 2: Performance of the different architectural variants described in Section 3.2.2. We
use P to refer to the number of parameters in a 12-layer base Transformer layer
stack and M to refer to the FLOPs required to process a sequence using the encoder-
decoder model. We evaluate each architectural variant using a denoising objective
(described in Section 3.1.4) and an autoregressive objective (as is commonly used
to train language models).
the encoder and decoder stacks significantly hurt performance. Concurrent work (Lan et al.,
2019) also found that sharing parameters across Transformer blocks can be an effective means
of lowering the total parameter count without sacrificing much performance. XLNet also
bears some resemblance to the shared encoder-decoder approach with a denoising objective
(Yang et al., 2019). We also note that the shared parameter encoder-decoder outperforms
the decoder-only prefix LM, suggesting that the addition of an explicit encoder-decoder
attention is beneficial. Finally, we confirm the widely-held conception that using a denoising
objective always results in better downstream task performance compared to a language
modeling objective. This observation has been previously made by Devlin et al. (2018),
Voita et al. (2019), and Lample and Conneau (2019) among others. We undertake a more
detailed exploration of unsupervised objectives in the following section.
3.3 Unsupervised Objectives
The choice of unsupervised objective is of central importance as it provides the mechanism
through which the model gains general-purpose knowledge to apply to downstream tasks.
This has led to the development of a wide variety of pre-training objectives (Dai and Le,
2015; Ramachandran et al., 2016; Radford et al., 2018; Devlin et al., 2018; Yang et al., 2019;
Liu et al., 2019b; Wang et al., 2019a; Song et al., 2019; Dong et al., 2019; Joshi et al., 2019).
In this section, we perform a procedural exploration of the space of unsupervised objectives.
In many cases, we will not replicate an existing objective exactly—some will be modified to
fit our text-to-text encoder-decoder framework and, in other cases, we will use objectives
that combine concepts from multiple common approaches.
Overall, all of our objectives ingest a sequence of token IDs corresponding to a tokenized
span of text from our unlabeled text data set. The token sequence is processed to produce a
(corrupted) input sequence and a corresponding target. Then, the model is trained as usual
20
Exploring the Limits of Transfer Learning
Objective Inputs Targets
Prefix language modeling Thank you for inviting me to your party last week .
BERT-style Devlin et al. (2018) Thank you <M> <M> me to your party apple week . (original text)
Deshuffling party me for your to . last fun you inviting week Thank (original text)
MASS-style Song et al. (2019) Thank you <M> <M> me to your party <M> week . (original text)
I.i.d. noise, replace spans Thank you <X> me to your party <Y> week . <X> for inviting <Y> last <Z>
I.i.d. noise, drop tokens Thank you me to your party week . for inviting last
Random spans Thank you <X> to <Y> week . <X> for inviting me <Y> your party last <Z>
Table 3: Examples of inputs and targets produced by some of the unsupervised objectives
we consider applied to the input text “Thank you for inviting me to your party last
week .” Note that all of our objectives process tokenized text. For this particular
sentence, all words were mapped to a single token by our vocabulary. We write
(original text) as a target to denote that the model is tasked with reconstructing the
entire input text. <M> denotes a shared mask token and <X>, <Y>, and <Z> denote
sentinel tokens that are assigned unique token IDs. The BERT-style objective
(second row) includes a corruption where some tokens are replaced by a random
token ID; we show this via the greyed-out word apple.
with maximum likelihood to predict the target sequence. We provide illustrative examples
of many of the objectives we consider in Table 3.
3.3.1 Disparate High-Level Approaches
To begin with, we compare three techniques that are inspired by commonly-used objectives
but differ significantly in their approach. First, we include a basic “prefix language modeling”
objective as was used in Section 3.2.3. This technique splits a span of text into two
components, one to use as inputs to the encoder and the other to use as a target sequence
to be predicted by the decoder. Second, we consider an objective inspired by the “masked
language modeling” (MLM) objective used in BERT (Devlin et al., 2018). MLM takes a
span of text and corrupts 15% of the tokens. 90% of the corrupted tokens are replaced
with a special mask token and 10% are replaced with a random token. Since BERT is an
encoder-only model, its goal during pre-training is to reconstruct masked tokens at the
output of the encoder. In the encoder-decoder case, we simply use the entire uncorrupted
sequence as the target. Note that this differs from our baseline objective, which uses only
the corrupted tokens as targets; we compare these two approaches in Section 3.3.2. Finally,
we also consider a basic deshuffling objective as used e.g. in (Liu et al., 2019a) where it was
applied to a denoising sequential autoencoder. This approach takes a sequence of tokens,
shuffles it, and then uses the original deshuffled sequence as a target. We provide examples
of the inputs and targets for these three methods in the first three rows of Table 3.
The performance of these three objectives is shown in Table 4. Overall, we find that the
BERT-style objective performs best, though the prefix language modeling objective attains
similar performance on the translation tasks. Indeed, the motivation for the BERT objective
was to outperform language model-based pre-training. The deshuffling objective performs
considerably worse than both prefix language modeling and the BERT-style objective.
21
Raffel, Shazeer, Roberts, Lee, Narang, Matena, Zhou, Li and Liu
Objective GLUE CNNDM SQuAD SGLUE EnDe EnFr EnRo
Prefix language modeling 80.69 18.94 77.99 65.27 26.86 39.73 27.49
BERT-style (Devlin et al., 2018) 82.96 19.17 80.65 69.85 26.78 40.03 27.41
Deshuffling 73.17 18.59 67.61 58.47 26.11 39.30 25.62
Table 4: Performance of the three disparate pre-training objectives described in Section 3.3.1.
Objective GLUE CNNDM SQuAD SGLUE EnDe EnFr EnRo
BERT-style (Devlin et al., 2018) 82.96 19.17 80.65 69.85 26.78 40.03 27.41
MASS-style (Song et al., 2019) 82.32 19.16 80.10 69.28 26.79 39.89 27.55
⋆ Replace corrupted spans 83.28 19.24 80.88 71.36 26.98 39.82 27.65
Drop corrupted tokens 84.44 19.31 80.52 68.67 27.07 39.76 27.82
Table 5: Comparison of variants of the BERT-style pre-training objective. In the first two
variants, the model is trained to reconstruct the original uncorrupted text segment.
In the latter two, the model only predicts the sequence of corrupted tokens.
3.3.2 Simplifying the BERT Objective
Based on the results in the prior section, we will now focus on exploring modifications to
the BERT-style denoising objective. This objective was originally proposed as a pre-training
technique for an encoder-only model trained for classification and span prediction. As
such, it may be possible to modify it so that it performs better or is more efficient in our
encoder-decoder text-to-text setup.
First, we consider a simple variant of the BERT-style objective where we don’t include the
random token swapping step. The resulting objective simply replaces 15% of the tokens in
the input with a mask token and the model is trained to reconstruct the original uncorrupted
sequence. A similar masking objective was used by Song et al. (2019) where it was referred to
as “MASS”, so we call this variant the “MASS-style” objective. Second, we were interested
to see if it was possible to avoid predicting the entire uncorrupted text span since this
requires self-attention over long sequences in the decoder. We consider two strategies to
achieve this: First, instead of replacing each corrupted token with a mask token, we replace
the entirety of each consecutive span of corrupted tokens with a unique mask token. Then,
the target sequence becomes the concatenation of the “corrupted” spans, each prefixed by
the mask token used to replace it in the input. This is the pre-training objective we use in
our baseline, described in Section 3.1.4. Second, we also consider a variant where we simply
drop the corrupted tokens from the input sequence completely and task the model with
reconstructing the dropped tokens in order. Examples of these approaches are shown in the
fifth and sixth rows of Table 3.
An empirical comparison of the original BERT-style objective to these three alternatives
is shown in Table 5. We find that in our setting, all of these variants perform similarly. The
only exception was that dropping corrupted tokens completely produced a small improvement
in the GLUE score thanks to a significantly higher score on CoLA (60.04, compared to our
22
Exploring the Limits of Transfer Learning
Corruption rate GLUE CNNDM SQuAD SGLUE EnDe EnFr EnRo
10% 82.82 19.00 80.38 69.55 26.87 39.28 27.44
⋆ 15% 83.28 19.24 80.88 71.36 26.98 39.82 27.65
25% 83.00 19.54 80.96 70.48 27.04 39.83 27.47
50% 81.27 19.32 79.80 70.33 27.01 39.90 27.49
Table 6: Performance of the i.i.d. corruption objective with different corruption rates.
baseline average of 53.84, see Table 16). This may be due to the fact that CoLA involves
classifying whether a given sentence is grammatically and syntactically acceptable, and
being able to determine when tokens are missing is closely related to detecting acceptability.
However, dropping tokens completely performed worse than replacing them with sentinel
tokens on SuperGLUE. The two variants that do not require predicting the full original
sequence (“replace corrupted spans” and “drop corrupted spans”) are both potentially
attractive since they make the target sequences shorter and consequently make training
faster. Going forward, we will explore variants where we replace corrupted spans with
sentinel tokens and only predict the corrupted tokens (as in our baseline objective).
3.3.3 Varying the Corruption Rate
So far, we have been corrupting 15% of the tokens, the value used in BERT (Devlin et al.,
2018). Again, since our text-to-text framework differs from BERT’s, we are interested to
see if a different corruption rate works better for us. We compare corruption rates of 10%,
15%, 25%, and 50% in Table 6. Overall, we find that the corruption rate had a limited
effect on the model’s performance. The only exception is that the largest corruption rate we
consider (50%) results in a significant degradation of performance on GLUE and SQuAD.
Using a larger corruption rate also results in longer targets, which can potentially slow down
training. Based on these results and the historical precedent set by BERT, we will use a
corruption rate of 15% going forward.
3.3.4 Corrupting Spans
We now turn towards the goal of speeding up training by predicting shorter targets. The
approach we have used so far makes an i.i.d. decision for each input token as to whether
to corrupt it or not. When multiple consecutive tokens have been corrupted, they are
treated as a “span” and a single unique mask token is used to replace the entire span.
Replacing entire spans with a single token results in unlabeled text data being processed into
shorter sequences. Since we are using an i.i.d. corruption strategy, it is not always the case
that a significant number of corrupted tokens appear consecutively. As a result, we might
obtain additional speedup by specifically corrupting spans of tokens rather than corrupting
individual tokens in an i.i.d. manner. Corrupting spans was also previously considered as a
pre-training objective for BERT, where it was found to improve performance (Joshi et al.,
2019).
To test this idea, we consider an objective that specifically corrupts contiguous, randomly-
spaced spans of tokens. This objective can be parametrized by the proportion of tokens to
be corrupted and the total number of corrupted spans. The span lengths are then chosen
23
Raffel, Shazeer, Roberts, Lee, Narang, Matena, Zhou, Li and Liu
Span length GLUE CNNDM SQuAD SGLUE EnDe EnFr EnRo
⋆ Baseline (i.i.d.) 83.28 19.24 80.88 71.36 26.98 39.82 27.65
2 83.54 19.39 82.09 72.20 26.76 39.99 27.63
3 83.49 19.62 81.84 72.53 26.86 39.65 27.62
5 83.40 19.24 82.05 72.23 26.88 39.40 27.53
10 82.85 19.33 81.84 70.44 26.79 39.49 27.69
Table 7: Performance of the span-corruption objective (inspired by Joshi et al. (2019)) for
different average span lengths. In all cases, we corrupt 15% of the original text
sequence.
randomly to satisfy these specified parameters. For example, if we are processing a sequence
of 500 tokens and we have specified that 15% of tokens should be corrupted and that there
should be 25 total spans, then the total number of corrupted tokens would be 500 × 0.15 = 75
and the average span length would be 75/25 = 3. Note that given the original sequence
length and corruption rate, we can equivalently parametrize this objective either by the
average span length or the total number of spans.
We compare the span-corruption objective to the i.i.d-corruption objective in Table 7.
We use a corruption rate of 15% in all cases and compare using average span lengths of 2, 3,
5 and 10. Again, we find a limited difference between these objectives, though the version
with an average span length of 10 slightly underperforms the other values in some cases.
We also find in particular that using an average span length of 3 slightly (but significantly)
outperforms the i.i.d. objective on most non-translation benchmarks. Fortunately, the
span-corruption objective also provides some speedup during training compared to the i.i.d.
noise approach because span corruption produces shorter sequences on average.
3.3.5 Discussion
Figure 5 shows a flow chart of the choices made during our exploration of unsupervised
objectives. Overall, the most significant difference in performance we observed was that
denoising objectives outperformed language modeling and deshuffling for pre-training. We
did not observe a remarkable difference across the many variants of the denoising objectives
we explored. However, different objectives (or parameterizations of objectives) can lead to
different sequence lengths and thus different training speeds. This implies that choosing
among the denoising objectives we considered here should mainly be done according to
their computational cost. Our results also suggest that additional exploration of objectives
similar to the ones we consider here may not lead to significant gains for the tasks and model
we consider. Instead, it may be fortuitous to explore entirely different ways of leveraging
unlabeled data.
3.4 Pre-training Data set
Like the unsupervised objective, the pre-training data set itself is a crucial component of
the transfer learning pipeline. However, unlike objectives and benchmarks, new pre-training
data sets are usually not treated as significant contributions on their own and are often not
24
Exploring the Limits of Transfer Learning
Figure 5: A flow chart of our exploration of unsupervised objectives. We first consider a
few disparate approaches in Section 3.3.1 and find that a BERT-style denoising
objective performs best. Then, we consider various methods for simplifying the
BERT objective so that it produces shorter target sequences in Section 3.3.2.
Given that replacing dropped-out spans with sentinel tokens performs well and
results in short target sequences, in Section 3.3.3 we experiment with different
corruption rates. Finally, we evaluate an objective that intentionally corrupts
contiguous spans of tokens in Section 3.3.4.
released alongside pre-trained models and code. Instead, they are typically introduced in
the course of presenting a new method or model. As a result, there has been relatively little
comparison of different pre-training data sets as well as a lack of a “standard” data set used
for pre-training. Some recent notable exceptions (Baevski et al., 2019; Liu et al., 2019c;
Yang et al., 2019) have compared pre-training on a new large (often Common Crawl-sourced)
data set to using a smaller preexisting data set (often Wikipedia). To probe more deeply
into the impact of the pre-training data set on performance, in this section we compare
variants of our C4 data set and other potential sources of pre-training data. We release all
of the C4 data set variants we consider as part of TensorFlow Datasets.11
3.4.1 Unlabeled Data Sets
In creating C4, we developed various heuristics to filter the web-extracted text from Common
Crawl (see Section 2.2 for a description). We are interested in measuring whether this
filtering results in improved performance on downstream tasks, in addition to comparing
it to other filtering approaches and common pre-training data sets. Towards this end, we
compare the performance of our baseline model after pre-training on the following data sets:
C4 As a baseline, we first consider pre-training on our proposed unlabeled data set as
described in Section 2.2.
Unfiltered C4 To measure the effect of the heuristic filtering we used in creating C4
(deduplication, removing bad words, only retaining sentences, etc.), we also generate
an alternate version of C4 that forgoes this filtering. Note that we still use langdetect
11. https://www.tensorflow.org/datasets/catalog/c4
25
Raffel, Shazeer, Roberts, Lee, Narang, Matena, Zhou, Li and Liu
to extract English text. As a result, our “unfiltered” variant still includes some filtering
because langdetect sometimes assigns a low probability to non-natural English text.
RealNews-like Recent work has used text data extracted from news websites (Zellers
et al., 2019; Baevski et al., 2019). To compare to this approach, we generate another
unlabeled data set by additionally filtering C4 to only include content from one of the
domains used in the “RealNews” data set (Zellers et al., 2019). Note that for ease of
comparison, we retain the heuristic filtering methods used in C4; the only difference is
that we have ostensibly omitted any non-news content.
WebText-like Similarly, the WebText data set (Radford et al., 2019) only uses content
from webpages that were submitted to the content aggregation website Reddit and
received a “score” of at least 3. The score for a webpage submitted to Reddit is
computed based on the proportion of users who endorse (upvote) or oppose (downvote)
the webpage. The idea behind using the Reddit score as a quality signal is that users
of the site would only upvote high-quality text content. To generate a comparable data
set, we first tried removing all content from C4 that did not originate from a URL that
appeared in the list prepared by the OpenWebText effort.12 However, this resulted in
comparatively little content—only about 2 GB—because most pages never appear on
Reddit. Recall that C4 was created based on a single month of Common Crawl data.
To avoid using a prohibitively small data set, we therefore downloaded 12 months
of data from Common Crawl from August 2018 to July 2019, applied our heuristic
filtering for C4, then applied the Reddit filter. This produced a 17 GB WebText-like
data set, which is of comparable size to the original 40GB WebText data set (Radford
et al., 2019).
Wikipedia The website Wikipedia consists of millions of encyclopedia articles written
collaboratively. The content on the site is subject to strict quality guidelines and
therefore has been used as a reliable source of clean and natural text. We use the
English Wikipedia text data from TensorFlow Datasets,13 which omits any markup or
reference sections from the articles.
Wikipedia + Toronto Books Corpus A drawback of using pre-training data from Wikipedia
is that it represents only one possible domain of natural text (encyclopedia articles).
To mitigate this, BERT (Devlin et al., 2018) combined data from Wikipedia with the
Toronto Books Corpus (TBC) (Zhu et al., 2015). TBC contains text extracted from
eBooks, which represents a different domain of natural language. BERT’s popularity
has led to the Wikipedia + TBC combination being used in many subsequent works.
The results achieved after pre-training on each of these data sets is shown in Table 8. A
first obvious takeaway is that removing the heuristic filtering from C4 uniformly degrades
performance and makes the unfiltered variant perform the worst in every task. Beyond
this, we found that in some cases a pre-training data set with a more constrained domain
outperformed the diverse C4 data set. For example, using the Wikipedia + TBC corpus
12. https://github.com/jcpeterson/openwebtext
13. https://www.tensorflow.org/datasets/catalog/wikipedia
26
Exploring the Limits of Transfer Learning
Data set Size GLUE CNNDM SQuAD SGLUE EnDe EnFr EnRo
⋆ C4 745GB 83.28 19.24 80.88 71.36 26.98 39.82 27.65
C4, unfiltered 6.1TB 81.46 19.14 78.78 68.04 26.55 39.34 27.21
RealNews-like 35GB 83.83 19.23 80.39 72.38 26.75 39.90 27.48
WebText-like 17GB 84.03 19.31 81.42 71.40 26.80 39.74 27.59
Wikipedia 16GB 81.85 19.31 81.29 68.01 26.94 39.69 27.67
Wikipedia + TBC 20GB 83.65 19.28 82.08 73.24 26.77 39.63 27.57
Table 8: Performance resulting from pre-training on different data sets. The first four
variants are based on our new C4 data set.
produced a SuperGLUE score of 73.24, beating our baseline’s score (using C4) of 71.36.
This is almost entirely attributable to a boost in performance from 25.78 (baseline, C4) to
50.93 (Wikipedia + TBC) on the Exact Match score for MultiRC (see Table 16). MultiRC
is a reading comprehension data set whose largest source of data comes from fiction books,
which is exactly the domain covered by TBC. Similarly, using the RealNews-like data set
for pre-training conferred an increase from 68.16 to 73.72 on the Exact Match score for
ReCoRD, a data set that measures reading comprehension on news articles. As a final
example, using data from Wikipedia produced significant (but less dramatic) gains on
SQuAD, which is a question-answering data set with passages sourced from Wikipedia.
Similar observations have been made in prior work, e.g. Beltagy et al. (2019) found that
pre-training BERT on text from research papers improved its performance on scientific tasks.
The main lesson behind these findings is that pre-training on in-domain unlabeled data can
improve performance on downstream tasks. This is unsurprising but also unsatisfying if
our goal is to pre-train a model that can rapidly adapt to language tasks from arbitrary
domains. Liu et al. (2019c) also observed that pre-training on a more diverse data set yielded
improvements on downstream tasks. This observation also motivates the parallel line of
research on domain adaptation for natural language processing; for surveys of this field see
e.g. Ruder (2019); Li (2012).
A drawback to only pre-training on a single domain is that the resulting data sets are
often substantially smaller. Similarly, while the WebText-like variant performed as well or
better than the C4 data set in our baseline setting, the Reddit-based filtering produced a
data set that was about 40× smaller than C4 despite being based on 12× more data from
Common Crawl. Note, however, that in our baseline setup we only pre-train on 235 ≈ 34B
tokens, which is only about 8 times larger than the smallest pre-training data set we consider.
We investigate at what point using a smaller pre-training data sets poses an issue in the
following section.
3.4.2 Pre-training Data set Size
The pipeline we use to create C4 was designed to be able to create extremely large pre-
training data sets. The access to so much data allows us to pre-train our models without
repeating examples. It is not clear whether repeating examples during pre-training would
be helpful or harmful to downstream performance because our pre-training objective is itself
stochastic and can help prevent the model from seeing the same exact data multiple times.
27
Raffel, Shazeer, Roberts, Lee, Narang, Matena, Zhou, Li and Liu
Number of tokens Repeats GLUE CNNDM SQuAD SGLUE EnDe EnFr EnRo
⋆ Full data set 0 83.28 19.24 80.88 71.36 26.98 39.82 27.65
229 64 82.87 19.19 80.97 72.03 26.83 39.74 27.63
227 256 82.62 19.20 79.78 69.97 27.02 39.71 27.33
225 1,024 79.55 18.57 76.27 64.76 26.38 39.56 26.80
223 4,096 76.34 18.33 70.92 59.29 26.37 38.84 25.81
Table 9: Measuring the effect of repeating data during pre-training. In these experiments,
we only use the first N tokens from C4 (with varying values of N shown in the
first column) but still pre-train over 235 tokens. This results in the data set being
repeated over the course of pre-training (with the number of repeats for each
experiment shown in the second column), which may result in memorization (see
Figure 6).
To test the effect of limited unlabeled data set sizes, we pre-trained our baseline model
on artificially truncated versions of C4. Recall that we pre-train our baseline model on
235 ≈ 34B tokens (a small fraction of the total size of C4). We consider training on truncated
variants of C4 consisting of 229 , 227 , 225 and 223 tokens. These sizes correspond to repeating
the data set 64, 256, 1,024, and 4,096 times respectively over the course of pre-training.
The resulting downstream performance is shown in Table 9. As expected, performance
degrades as the data set size shrinks. We suspect this may be due to the fact that the model
begins to memorize the pre-training data set. To measure if this is true, we plot the training
loss for each of these data set sizes in Figure 6. Indeed, the model attains significantly
smaller training losses as the size of the pre-training data set shrinks, suggesting possible
memorization. Baevski et al. (2019) similarly observed that truncating the pre-training data
set size can degrade downstream task performance.
We note that these effects are limited when the pre-training data set is repeated only
64 times. This suggests that some amount of repetition of pre-training data might not be
harmful. However, given that additional pre-training can be beneficial (as we will show in
Section 3.6) and that obtaining additional unlabeled data is cheap and easy, we suggest
using large pre-training data sets whenever possible. We also note that this effect may be
more pronounced for larger model sizes, i.e. a bigger model may be more prone to overfitting
to a smaller pre-training data set.
3.5 Training Strategy
So far we have considered the setting where all parameters of a model are pre-trained on
an unsupervised task before being fine-tuned on individual supervised tasks. While this
approach is straightforward, various alternative methods for training the model on down-
stream/supervised tasks have been proposed. In this section, we compare different schemes
for fine-tuning the model in addition to the approach of training the model simultaneously
on multiple tasks.
28
Exploring the Limits of Transfer Learning
8 V E M R M R K P S W W
( E X E W I X W M ^ I
* Y P P H E X E W I X
7 X I T
Figure 6: Pre-training loss for our original C4 data set as well as 4 artificially truncated
versions. The sizes listed refer to the number of tokens in each data set. The four
sizes considered correspond to repeating the data set between 64 and 4,096 times
over the course of pre-training. Using a smaller data set size results in smaller
training loss values, which may suggest some memorization of the unlabeled data
set.
3.5.1 Fine-tuning Methods
It has been argued that fine-tuning all of the model’s parameters can lead to suboptimal
results, particularly on low-resource tasks (Peters et al., 2019). Early results on transfer
learning for text classification tasks advocated fine-tuning only the parameters of a small
classifier that was fed sentence embeddings produced by a fixed pre-trained model (Subra-
manian et al., 2018; Kiros et al., 2015; Logeswaran and Lee, 2018; Hill et al., 2016; Conneau
et al., 2017). This approach is less applicable to our encoder-decoder model because the
entire decoder must be trained to output the target sequences for a given task. Instead, we
focus on two alternative fine-tuning approaches that update only a subset of the parameters
of our encoder-decoder model.
The first, “adapter layers” (Houlsby et al., 2019; Bapna et al., 2019), is motivated by
the goal of keeping most of the original model fixed while fine-tuning. Adapter layers are
additional dense-ReLU-dense blocks that are added after each of the preexisting feed-forward
networks in each block of the Transformer. These new feed-forward networks are designed
so that their output dimensionality matches their input. This allows them to be inserted
into the network with no additional changes to the structure or parameters. When fine-
tuning, only the adapter layer and layer normalization parameters are updated. The main
hyperparameter of this approach is the inner dimensionality d of the feed-forward network,
which changes the number of new parameters added to the model. We experiment with
various values for d.
The second alternative fine-tuning method we consider is “gradual unfreezing” (Howard
and Ruder, 2018). In gradual unfreezing, more and more of the model’s parameters are fine-
tuned over time. Gradual unfreezing was originally applied to a language model architecture
consisting of a single stack of layers. In this setting, at the start of fine-tuning only the
29
Raffel, Shazeer, Roberts, Lee, Narang, Matena, Zhou, Li and Liu
Fine-tuning method GLUE CNNDM SQuAD SGLUE EnDe EnFr EnRo
⋆ All parameters 83.28 19.24 80.88 71.36 26.98 39.82 27.65
Adapter layers, d = 32 80.52 15.08 79.32 60.40 13.84 17.88 15.54
Adapter layers, d = 128 81.51 16.62 79.47 63.03 19.83 27.50 22.63
Adapter layers, d = 512 81.54 17.78 79.18 64.30 23.45 33.98 25.81
Adapter layers, d = 2048 81.51 16.62 79.47 63.03 19.83 27.50 22.63
Gradual unfreezing 82.50 18.95 79.17 70.79 26.71 39.02 26.93
Table 10: Comparison of different alternative fine-tuning methods that only update a subset
of the model’s parameters. For adapter layers, d refers to the inner dimensionality
of the adapters.
parameters of the final layer are updated, then after training for a certain number of updates
the parameters of the second-to-last layer are also included, and so on until the entire
network’s parameters are being fine-tuned. To adapt this approach to our encoder-decoder
model, we gradually unfreeze layers in the encoder and decoder in parallel, starting from
the top in both cases. Since the parameters of our input embedding matrix and output
classification matrix are shared, we update them throughout fine-tuning. Recall that our
baseline model consists of 12 layers each in the encoder and decoder and is fine-tuned for
218 steps. As such, we subdivide the fine-tuning process into 12 episodes of 218/12 steps each
and train from layers 12 − n to 12 in the nth episode. We note that Howard and Ruder
(2018) suggested fine-tuning an additional layer after each epoch of training. However, since
our supervised data sets vary so much in size and since some of our downstream tasks are
actually mixtures of many tasks (GLUE and SuperGLUE), we instead adopt the simpler
strategy of fine-tuning an additional layer after every 218/12 steps.
A comparison of the performance of these fine-tuning approaches is shown in Table 10.
For adapter layers, we report the performance using an inner dimensionality d of 32, 128,
512, 2048. Pursuant with past results (Houlsby et al., 2019; Bapna et al., 2019) we find that
lower-resource tasks like SQuAD work well with a small value of d whereas higher resource
tasks require a large dimensionality to achieve reasonable performance. This suggests that
adapter layers could be a promising technique for fine-tuning on fewer parameters as long as
the dimensionality is scaled appropriately to the task size. Note that in our case we treat
GLUE and SuperGLUE each as a single “task” by concatenating their constituent data
sets, so although they comprise some low-resource data sets the combined data set is large
enough that it necessitates a large value of d. We found that gradual unfreezing caused
a minor degradation in performance across all tasks, though it did provide some speedup
during fine-tuning. Better results may be attainable by more carefully tuning the unfreezing
schedule.
3.5.2 Multi-task Learning
So far, we have been pre-training our model on a single unsupervised learning task before
fine-tuning it individually on each downstream task. An alternative approach, called “multi-
task learning” (Ruder, 2017; Caruana, 1997), is to train the model on multiple tasks at a
time. This approach typically has the goal of training a single model that can simultaneously
30
Exploring the Limits of Transfer Learning
perform many tasks at once, i.e. the model and most of its parameters are shared across all
tasks. We relax this goal somewhat and instead investigate methods for training on multiple
tasks at once in order to eventually produce separate parameter settings that perform well
on each individual task. For example, we might train a single model on many tasks, but
when reporting performance we are allowed to select a different checkpoint for each task.
This loosens the multi-task learning framework and puts it on more even footing compared
to the pre-train-then-fine-tune approach we have considered so far. We also note that in our
unified text-to-text framework, “multi-task learning” simply corresponds to mixing data sets
together. It follows that we can still train on unlabeled data when using multi-task learning
by treating the unsupervised task as one of the tasks being mixed together. In contrast,
most applications of multi-task learning to NLP add task-specific classification networks or
use different loss functions for each task (Liu et al., 2019b).
As pointed out by Arivazhagan et al. (2019), an extremely important factor in multi-task
learning is how much data from each task the model should be trained on. Our goal is to not
under- or over-train the model—that is, we want the model to see enough data from a given
task that it can perform the task well, but not to see so much data that it memorizes the
training set. How exactly to set the proportion of data coming from each task can depend on
various factors including data set sizes, the “difficulty” of learning the task (i.e. how much
data the model must see before being able to perform the task effectively), regularization,
etc. An additional issue is the potential for “task interference” or “negative transfer”, where
achieving good performance on one task can hinder performance on another. Given these
concerns, we begin by exploring various strategies for setting the proportion of data coming
from each task. A similar exploration was performed by Wang et al. (2019a).
Examples-proportional mixing A major factor in how quickly a model will overfit to
a given task is the task’s data set size. As such, a natural way to set the mixing
proportions is to sample in proportion to the size of each task’s data set. This is
equivalent to concatenating the data sets for all tasks and randomly sampling examples
from the combined data set. Note, however, that we are including our unsupervised
denoising task, which uses a data set that is orders of magnitude larger than every
other task’s. It follows that if we simply sample in proportion to each data set’s size,
the vast majority of the data the model sees will be unlabeled, and it will undertrain
on all of the supervised tasks. Even without the unsupervised task, some tasks (e.g.
WMT English to French) are so large that they would similarly crowd out most of
the batches. To get around this issue, we set an artificial “limit” on the data set sizes
before computing the proportions. Specifically, if the number of examples in each of
our N task’s data sets is en , n ∈ {1, . . . , N } then we set probability of sampling an
example from the mth task during training to rm = min(em , K)/ min(en , K) where
P
K is the artificial data set size limit.
Temperature-scaled mixing An alternative way of mitigating the huge disparity between
data set sizes is to adjust the “temperature” of the mixing rates. This approach was
used by multilingual BERT to ensure that the model was sufficiently trained on low-
resource languages.14 To implement temperature scaling with temperature T , we raise
14. https://github.com/google-research/bert/blob/master/multilingual.md
31
Raffel, Shazeer, Roberts, Lee, Narang, Matena, Zhou, Li and Liu
each task’s mixing rate rm to the power of 1⁄T and renormalize the rates so that they
sum to 1. When T = 1, this approach is equivalent to examples-proportional mixing
and as T increases the proportions become closer to equal mixing. We retain the data
set size limit K (applied to obtain rm before temperature scaling) but set it to a large
value of K = 221 . We use a large value of K because increasing the temperature will
decrease the mixing rate of the largest data sets.
Equal mixing In this case, we sample examples from each task with equal probability.
Specifically, each example in each batch is sampled uniformly at random from one of
the data sets we train on. This is most likely a suboptimal strategy, as the model will
overfit quickly on low-resource tasks and underfit on high-resource tasks. We mainly
include it as a point of reference of what might go wrong when the proportions are set
suboptimally.
To compare these mixing strategies on equal footing with our baseline pre-train-then-
fine-tune results, we train multi-task models for the same total number of steps: 219 + 218 =
786,432. The results are shown in Table 11.
In general, we find that multi-task training underperforms pre-training followed by
fine-tuning on most tasks. The “equal” mixing strategy in particular results in dramatically
degraded performance, which may be because the low-resource tasks have overfit, the high-
resource tasks have not seen enough data, or the model has not seen enough unlabeled data to
learn general-purpose language capabilities. For examples-proportional mixing, we find that
for most tasks there is a “sweet spot” for K where the model obtains the best performance,
and larger or smaller values of K tend to result in worse performance. The exception (for the
range of K values we considered) was WMT English to French translation, which is such a
high-resource task that it always benefits from a higher mixing proportion. Finally, we note
that temperature-scaled mixing also provides a means of obtaining reasonable performance
from most tasks, with T = 2 performing the best in most cases. The finding that a multi-task
model is outperformed by separate models trained on each individual task has previously
been observed e.g. by Arivazhagan et al. (2019) and McCann et al. (2018), though it has
been shown that the multi-task setup can confer benefits across very similar tasks Liu et al.
(2019b); Ratner et al. (2018). In the following section, we explore ways to close the gap
between multi-task training and the pre-train-then-fine-tune approach.
3.5.3 Combining Multi-Task Learning with Fine-Tuning
Recall that we are studying a relaxed version of multi-task learning where we train a single
model on a mixture of tasks but are allowed to evaluate performance using different parameter
settings (checkpoints) for the model. We can extend this approach by considering the case
where the model is pre-trained on all tasks at once but is then fine-tuned on the individual
supervised tasks. This is the method used by the “MT-DNN” (Liu et al., 2015, 2019b),
which achieved state-of-the-art performance on GLUE and other benchmarks when it was
introduced. We consider three variants of this approach: In the first, we simply pre-train the
model on an examples-proportional mixture with an artificial data set size limit of K = 219
before fine-tuning it on each individual downstream task. This helps us measure whether
including the supervised tasks alongside the unsupervised objective during pre-training
32
Exploring the Limits of Transfer Learning
Mixing strategy GLUE CNNDM SQuAD SGLUE EnDe EnFr EnRo
⋆ Baseline (pre-train/fine-tune) 83.28 19.24 80.88 71.36 26.98 39.82 27.65
Equal 76.13 19.02 76.51 63.37 23.89 34.31 26.78
Examples-proportional, K = 216 80.45 19.04 77.25 69.95 24.35 34.99 27.10
Examples-proportional, K = 217 81.56 19.12 77.00 67.91 24.36 35.00 27.25
Examples-proportional, K = 218 81.67 19.07 78.17 67.94 24.57 35.19 27.39
Examples-proportional, K = 219 81.42 19.24 79.78 67.30 25.21 36.30 27.76
Examples-proportional, K = 220 80.80 19.24 80.36 67.38 25.66 36.93 27.68
Examples-proportional, K = 221 79.83 18.79 79.50 65.10 25.82 37.22 27.13
Temperature-scaled, T = 2 81.90 19.28 79.42 69.92 25.42 36.72 27.20
Temperature-scaled, T = 4 80.56 19.22 77.99 69.54 25.04 35.82 27.45
Temperature-scaled, T = 8 77.21 19.10 77.14 66.07 24.55 35.35 27.17
Table 11: Comparison of multi-task training using different mixing strategies. Examples-
proportional mixing refers to sampling examples from each data set according to
the total size of each data set, with an artificial limit (K) on the maximum data set
size. Temperature-scaled mixing re-scales the sampling rates by a temperature T .
For temperature-scaled mixing, we use an artificial data set size limit of K = 221 .
gives the model some beneficial early exposure to the downstream tasks. We might also
hope that mixing in many sources of supervision could help the pre-trained model obtain a
more general set of “skills” (loosely speaking) before it is adapted to an individual task. To
measure this directly, we consider a second variant where we pre-train the model on the same
examples-proportional mixture (with K = 219 ) except that we omit one of the downstream
tasks from this pre-training mixture. Then, we fine-tune the model on the task that was
left out during pre-training. We repeat this for each of the downstream tasks we consider.
We call this approach “leave-one-out” multi-task training. This simulates the real-world
setting where a pre-trained model is fine-tuned on a task it had not seen during pre-training.
Note that multi-task pre-training provides a diverse mixture of supervised tasks. Since other
fields (e.g. computer vision (Oquab et al., 2014; Jia et al., 2014; Huh et al., 2016; Yosinski
et al., 2014)) use a supervised data set for pre-training, we were interested to see whether
omitting the unsupervised task from the multi-task pre-training mixture still produced good
results. For our third variant we therefore pre-train on an examples-proportional mixture of
all of the supervised tasks we consider with K = 219 . In all of these variants, we follow our
standard procedure of pre-training for 219 steps before fine-tuning for 218 steps.
We compare the results of these approaches in Table 12. For comparison, we also include
results for our baseline (pre-train then fine-tune) and for standard multi-task learning
(without fine-tuning) on an examples-proportional mixture with K = 219 . We find that
fine-tuning after multi-task pre-training results in comparable performance to our baseline.
This suggests that using fine-tuning after multi-task learning can help mitigate some of
the trade-offs between different mixing rates described in Section 3.5.2. Interestingly, the
performance of “leave-one-out” training was only slightly worse, suggesting that a model
that was trained on a variety of tasks can still adapt to new tasks (i.e. multi-task pre-
training might not result in a dramatic task interference). Finally, supervised multi-task
pre-training performed significantly worse in every case except for the translation tasks. This
33
Raffel, Shazeer, Roberts, Lee, Narang, Matena, Zhou, Li and Liu
Training strategy GLUE CNNDM SQuAD SGLUE EnDe EnFr EnRo
⋆ Unsupervised pre-training + fine-tuning 83.28 19.24 80.88 71.36 26.98 39.82 27.65
Multi-task training 81.42 19.24 79.78 67.30 25.21 36.30 27.76
Multi-task pre-training + fine-tuning 83.11 19.12 80.26 71.03 27.08 39.80 28.07
Leave-one-out multi-task training 81.98 19.05 79.97 71.68 26.93 39.79 27.87
Supervised multi-task pre-training 79.93 18.96 77.38 65.36 26.81 40.13 28.04
Table 12: Comparison of unsupervised pre-training, multi-task learning, and various forms
of multi-task pre-training.
could suggest that the translation tasks benefit less from (English) pre-training, whereas
unsupervised pre-training is an important factor in the other tasks.
3.6 Scaling
The “bitter lesson” of machine learning research argues that general methods that can
leverage additional computation ultimately win out against methods that rely on human
expertise (Sutton, 2019; Hestness et al., 2017; Shazeer et al., 2017; Jozefowicz et al., 2016;
Mahajan et al., 2018; Shazeer et al., 2018, 2017; Huang et al., 2018b; Keskar et al., 2019a).
Recent results suggest that this may hold true for transfer learning in NLP (Liu et al., 2019c;
Radford et al., 2019; Yang et al., 2019; Lan et al., 2019), i.e. it has repeatedly been shown
that scaling up produces improved performance compared to more carefully-engineered
methods. However, there are a variety of possible ways to scale, including using a bigger
model, training the model for more steps, and ensembling. In this section, we compare these
different approaches by addressing the following premise: “You were just given 4× more
compute. How should you use it?”
We start with our baseline model, which has 220M parameters and is pre-trained and
fine-tuned for 219 and 218 steps respectively. The encoder and decoder are both sized
similarly to “BERTBASE ”. To experiment with increased model size, we follow the guidelines
of “BERTLARGE ” Devlin et al. (2018) and use dff = 4096, dmodel = 1024, dkv = 64 and
16-head attention mechanisms. We then generate two variants with 16 and 32 layers each in
the encoder and decoder, producing models with 2× and 4× as many parameters as our
original model. These two variants also have a roughly 2× and 4× the computational cost.
Using our baseline and these two larger models, we consider three ways of using 4× as much
computation: Training for 4× as many steps, training for 2× as many steps with the 2×
bigger model, and training the 4× bigger model for the “baseline” number of training steps.
When we increase the training steps, we scale both the pre-train and fine-tune steps for
simplicity. Note that when increasing the number of pre-training steps, we are effectively
including more pre-training data as C4 is so large that we do not complete one pass over
the data even when training for 223 steps.
An alternative way for the model to see 4× as much data is to increase the batch size by a
factor of 4. This can potentially result in faster training due to more efficient parallelization.
However, training with a 4× larger batch size can yield a different outcome than training
34
Exploring the Limits of Transfer Learning
Scaling strategy GLUE CNNDM SQuAD SGLUE EnDe EnFr EnRo
⋆ Baseline 83.28 19.24 80.88 71.36 26.98 39.82 27.65
1× size, 4× training steps 85.33 19.33 82.45 74.72 27.08 40.66 27.93
1× size, 4× batch size 84.60 19.42 82.52 74.64 27.07 40.60 27.84
2× size, 2× training steps 86.18 19.66 84.18 77.18 27.52 41.03 28.19
4× size, 1× training steps 85.91 19.73 83.86 78.04 27.47 40.71 28.10
4× ensembled 84.77 20.10 83.09 71.74 28.05 40.53 28.57
4× ensembled, fine-tune only 84.05 19.57 82.36 71.55 27.55 40.22 28.09
Table 13: Comparison of different methods of scaling up our baseline model. All methods
except ensembling fine-tuned models use 4× the computation as the baseline.
“Size” refers to the number of parameters in the model and “training time” refers
to the number of steps used for both pre-training and fine-tuning.
for 4× as many steps (Shallue et al., 2018). We include an additional experiment where we
train our baseline model with a 4× larger batch size to compare these two cases.
It is common practice on many of the benchmarks we consider to eke out additional
performance by training and evaluating using an ensemble of models. This provides an
orthogonal way of using additional computation. To compare other scaling methods to
ensembling, we also measure the performance of an ensemble of 4 separately pre-trained and
fine-tuned models. We average the logits across the ensemble before feeding them into the
output softmax nonlinearity to obtain an aggregate prediction. Instead of pre-training 4
separate models, a cheaper alternative is to take a single pre-trained model and produce 4
separate fine-tuned versions. While this does not use our entire 4× computational budget,
we also include this method to see if it produces competitive performance to the other scaling
methods.
The performance achieved after applying these various scaling methods is shown in
Table 13. Unsurprisingly, increasing the training time and/or model size consistently
improves the baseline. There was no clear winner between training for 4× as many steps
or using a 4× larger batch size, though both were beneficial. In general, increasing the
model size resulted in an additional bump in performance compared to solely increasing
the training time or batch size. We did not observe a large difference between training a
2× bigger model for 2× as long and training a 4× bigger model on any of the tasks we
studied. This suggests that increasing the training time and increasing the model size can be
complementary means of improving performance. Our results also suggest that ensembling
provides an orthogonal and effective means of improving performance through scale. In some
tasks (CNN/DM, WMT English to German, and WMT English to Romanian), ensembling 4
completely separately trained models significantly outperformed every other scaling approach.
Ensembling models that were pre-trained together but fine-tuned separately also gave a
substantial performance increase over the baseline, which suggests a cheaper means of
improving performance. The only exception was SuperGLUE, where neither ensembling
approach significantly improved over the baseline.
We note that different scaling methods have different trade-offs that are separate from
their performance. For example, using a larger model can make downstream fine-tuning and
35
Raffel, Shazeer, Roberts, Lee, Narang, Matena, Zhou, Li and Liu
inference more expensive. In contrast, the cost of pre-training a small model for longer is
effectively amortized if it is applied to many downstream tasks. Separately, we note that
ensembling N separate models has a similar cost to using a model that has an N × higher
computational cost. As a result, some consideration for the eventual use of the model is
important when choosing between scaling methods.
3.7 Putting It All Together
We now leverage the insights from our systematic study to determine how far we can push
performance on popular NLP benchmarks. We are also interested in exploring the current
limits of transfer learning for NLP by training larger models on large amounts of data. We
start with our baseline training approach and make the following changes:
Objective We swap out the i.i.d. denoising objective in our baseline for the span-corruption
objective described in Section 3.3.4, which was loosely inspired by SpanBERT (Joshi
et al., 2019). Specifically, we use a mean span length of 3 and corrupt 15% of the
original sequence. We found that this objective produced marginally better performance
(Table 7) while being slightly more computationally efficient due to shorter target
sequence lengths.
Longer training Our baseline model uses a relatively small amount of pre-training (1⁄4 as
much as BERT (Devlin et al., 2018), 1⁄16 as much as XLNet (Yang et al., 2019), 1⁄64 as
much as RoBERTa (Liu et al., 2019c), etc.). Fortunately, C4 is big enough that we
can train for substantially longer without repeating data (which can be detrimental,
as shown in Section 3.4.2). We found in Section 3.6 that additional pre-training can
indeed be helpful, and that both increasing the batch size and increasing the number of
training steps can confer this benefit. We therefore pre-train our models for 1 million
steps on a batch size of 211 sequences of length 512, corresponding to a total of about
1 trillion pre-training tokens (about 32× as many as our baseline). In Section 3.4.1, we
showed that pre-training on the RealNews-like, WebText-like, and Wikipedia + TBC
data sets outperformed pre-training on C4 on a few downstream tasks. However, these
data set variants are sufficiently small that they would be repeated hundreds of times
over the course of pre-training on 1 trillion tokens. Since we showed in Section 3.4.2
that this repetition could be harmful, we opted instead to continue using the C4 data
set.
Model sizes In Section 3.6 we also showed how scaling up the baseline model size improved
performance. However, using smaller models can be helpful in settings where limited
computational resources are available for fine-tuning or inference. Based on these
factors, we train models with a wide range of sizes:
• Base. This is our baseline model, whose hyperparameters are described in
Section 3.1.1. It has roughly 220 million parameters.
• Small. We consider a smaller model, which scales the baseline down by using
dmodel = 512, dff = 2,048, 8-headed attention, and only 6 layers each in the
encoder and decoder. This variant has about 60 million parameters.
36
Exploring the Limits of Transfer Learning
• Large. Since our baseline uses a BERTBASE -sized encoder and decoder, we
also consider a variant where the encoder and decoder are both similar in size
and structure to BERTLARGE . Specifically, this variant uses dmodel = 1,024,
dff = 4,096, dkv = 64, 16-headed attention, and 24 layers each in the encoder and
decoder, resulting in around 770 million parameters.
• 3B and 11B. To further explore what kind of performance is possible when
using larger models, we consider two additional variants. In both cases, we use
dmodel = 1024, a 24 layer encoder and decoder, and dkv = 128. For the “3B”
variant, we use dff = 16,384 with 32-headed attention, which results in around
2.8 billion parameters; for “11B” we use dff = 65,536 with 128-headed attention
producing a model with about 11 billion parameters. We chose to scale up dff
specifically because modern accelerators (such as the TPUs we train our models
on) are most efficient for large dense matrix multiplications like those in the
Transformer’s feed-forward networks.
Multi-task pre-training In Section 3.5.3, we showed that pre-training on a multi-task
mixture of unsupervised and supervised tasks before fine-tuning worked as well as
pre-training on the unsupervised task alone. This is the approach advocated by the
“MT-DNN” (Liu et al., 2015, 2019b). It also has the practical benefit of being able to
monitor “downstream” performance for the entire duration of training, rather than
just during fine-tuning. We therefore used multi-task pre-training in our final set of
experiments. We hypothesize that larger models trained for longer might benefit from
a larger proportion of unlabeled data because they are more likely to overfit to smaller
training data sets. However, we also note that the results of Section 3.5.3 suggest that
fine-tuning after multi-task pre-training can mitigate some of the issues that might
arise from choosing a suboptimal proportion of unlabeled data. Based on these ideas,
we substitute the following artificial data set sizes for our unlabeled data before using
standard example-proportional mixing (described in Section 3.5.2): 710,000 for Small,
2,620,000 for Base, 8,660,000 for Large, 33,500,000 for 3B, and 133,000,000 for 11B.
For all model variants, we also capped the effective data set size of the WMT English
to French and WMT English to German data sets to 1M examples during pre-training.
Fine-tuning on individual GLUE and SuperGLUE tasks So far, when fine-tuning
on GLUE and SuperGLUE, we have concatenated all of the data sets in each benchmark
so that we only fine-tune models once for GLUE and once for SuperGLUE. This
approach makes our study logistically simpler, but we found that this sacrifices a small
amount of performance on some tasks compared to fine-tuning on the task separately. A
potential issue with fine-tuning on individual tasks, which would otherwise be mitigated
by training on all tasks at once, is that we might overfit quickly to low-resource tasks.
For example, our large batch size of 211 length-512 sequences would result in the entire
data set appearing multiple times in each batch for many of the low-resource GLUE
and SuperGLUE tasks. We therefore use a smaller batch size of 8 length-512 sequences
during fine-tuning for each GLUE and SuperGLUE task. We also save checkpoints
every 1,000 steps rather than every 5,000 steps to ensure we have access to the model’s
parameters before it overfits.
37
Raffel, Shazeer, Roberts, Lee, Narang, Matena, Zhou, Li and Liu
Beam search All of our previous results were reported using greedy decoding. For tasks
with long output sequences, we found improved performance from using beam search
(Sutskever et al., 2014). Specifically, we use a beam width of 4 and a length penalty
of α = 0.6 (Wu et al., 2016) for the WMT translation and CNN/DM summarization
tasks.
Test set Since this is our final set of experiments, we report results on the test set rather
than the validation set. For CNN/Daily Mail, we use the standard test set distributed
with the data set. For the WMT tasks, this corresponds to using newstest2014 for
English-German, newstest2015 for English-French, and newstest2016 for English-
Romanian. For GLUE and SuperGLUE, we used the benchmark evaluation servers to
compute official test set scores.15,16 For SQuAD, evaluating on the test set requires
running inference on a benchmark server. Unfortunately, the computational resources
on this server are insufficient for obtaining predictions from our largest models. As
a result, we instead continue to report performance on the SQuAD validation set.
Fortunately, the model with the highest performance on the SQuAD test set also
reported results on the validation set, so we can still compare to what is ostensibly
the state-of-the-art.
Apart from those changes mentioned above, we use the same training procedure and
hyperparameters as our baseline (AdaFactor optimizer, inverse square root learning rate
schedule for pre-training, constant learning rate for fine-tuning, dropout regularization,
vocabulary, etc.). For reference, these details are described in Section 2.
The results of this final set of experiments are shown in Table 14. Overall, we achieved
state-of-the-art performance on 18 out of the 24 tasks we consider. As expected, our largest
(11 billion parameter) model performed best among our model size variants across all tasks.
Our T5-3B model variant did beat the previous state of the art in a few tasks, but scaling
the model size to 11 billion parameters was the most important ingredient for achieving our
best performance. We now analyze the results for each individual benchmark.
We achieved a state-of-the-art average GLUE score of 90.3. Notably, our performance was
substantially better than the previous state-of-the-art for the natural language inference tasks
MNLI, RTE, and WNLI. RTE and WNLI are two of the tasks where machine performance
has historically lagged behind human performance, which is 93.6 and 95.9 respectively (Wang
et al., 2018). In terms of parameter count, our 11B model variant is the largest model that
has been submitted to the GLUE benchmark. However, most of the best-scoring submissions
use a large amount of ensembling and computation to produce predictions. For example,
the best-performing variant of ALBERT (Lan et al., 2019) uses a model similar in size and
architecture to our 3B variant (though it has dramatically fewer parameters due to clever
parameter sharing). To produce its impressive performance on GLUE, the ALBERT authors
ensembled “from 6 to 17” models depending on the task. This likely results in it being more
computationally expensive to produce predictions with the ALBERT ensemble than it is
with T5-11B.
For SQuAD, we outperformed the previous state-of-the-art (ALBERT (Lan et al., 2019))
by over one point on the Exact Match score. SQuAD is a long-standing benchmark that
15. http://gluebenchmark.com
16. http://super.gluebenchmark.com
38
Exploring the Limits of Transfer Learning
GLUE CoLA SST-2 MRPC MRPC STS-B STS-B
Model Average Matthew’s Accuracy F1 Accuracy Pearson Spearman
Previous best 89.4a 69.2b 97.1a 93.6b 91.5b 92.7b 92.3b
T5-Small 77.4 41.0 91.8 89.7 86.6 85.6 85.0
T5-Base 82.7 51.1 95.2 90.7 87.5 89.4 88.6
T5-Large 86.4 61.2 96.3 92.4 89.9 89.9 89.2
T5-3B 88.5 67.1 97.4 92.5 90.0 90.6 89.8
T5-11B 90.3 71.6 97.5 92.8 90.4 93.1 92.8
QQP QQP MNLI-m MNLI-mm QNLI RTE WNLI
Model F1 Accuracy Accuracy Accuracy Accuracy Accuracy Accuracy
Previous best 74.8c 90.7b 91.3a 91.0a 99.2a 89.2a 91.8a
T5-Small 70.0 88.0 82.4 82.3 90.3 69.9 69.2
T5-Base 72.6 89.4 87.1 86.2 93.7 80.1 78.8
T5-Large 73.9 89.9 89.9 89.6 94.8 87.2 85.6
T5-3B 74.4 89.7 91.4 91.2 96.3 91.1 89.7
T5-11B 75.1 90.6 92.2 91.9 96.9 92.8 94.5
SQuAD SQuAD SuperGLUE BoolQ CB CB COPA
Model EM F1 Average Accuracy F1 Accuracy Accuracy
Previous best 90.1a 95.5a 84.6d 87.1d 90.5d 95.2d 90.6d
T5-Small 79.10 87.24 63.3 76.4 56.9 81.6 46.0
T5-Base 85.44 92.08 76.2 81.4 86.2 94.0 71.2
T5-Large 86.66 93.79 82.3 85.4 91.6 94.8 83.4
T5-3B 88.53 94.95 86.4 89.9 90.3 94.4 92.0
T5-11B 91.26 96.22 88.9 91.2 93.9 96.8 94.8
MultiRC MultiRC ReCoRD ReCoRD RTE WiC WSC
Model F1a EM F1 Accuracy Accuracy Accuracy Accuracy
Previous best 84.4d 52.5d 90.6d 90.0d 88.2d 69.9d 89.0d
T5-Small 69.3 26.3 56.3 55.4 73.3 66.9 70.5
T5-Base 79.7 43.1 75.0 74.2 81.5 68.3 80.8
T5-Large 83.3 50.7 86.8 85.9 87.8 69.3 86.3
T5-3B 86.8 58.3 91.2 90.4 90.7 72.1 90.4
T5-11B 88.1 63.3 94.1 93.4 92.5 76.9 93.8
WMT EnDe WMT EnFr WMT EnRo CNN/DM CNN/DM CNN/DM
Model BLEU BLEU BLEU ROUGE-1 ROUGE-2 ROUGE-L
Previous best 33.8e 43.8e 38.5f 43.47g 20.30g 40.63g
T5-Small 26.7 36.0 26.8 41.12 19.56 38.35
T5-Base 30.9 41.2 28.0 42.05 20.34 39.40
T5-Large 32.0 41.5 28.1 42.50 20.68 39.75
T5-3B 31.8 42.6 28.2 42.72 21.02 39.94
T5-11B 32.1 43.4 28.1 43.52 21.55 40.69
Table 14: Performance of our T5 variants on every task we study. Small, Base, Large, 3B,
and 11B refer to model configurations with 60 million, 220 million, 770 million,
3 billion, and 11 billion parameters, respectively. In the first row of each table,
we report the state-of-the-art for the task (as of October 24th, 2019), with the
superscript denoting its source with references listed at the end of this caption. All
results are reported on the test set except for SQuAD where we use the validation
set. a (Lan et al., 2019) b (Wang et al., 2019c) c (Zhu et al., 2019) d (Liu et al.,
2019c) e (Edunov et al., 2018) f (Lample and Conneau, 2019) g (Dong et al., 2019)
39
Raffel, Shazeer, Roberts, Lee, Narang, Matena, Zhou, Li and Liu
was created over three years ago, and most recent improvements have only increased the
state-of-the-art by a fraction of a percentage point. We note that when results are reported
on the test set, they are typically based on an ensemble of models and/or leverage external
data sets (e.g. TriviaQA (Joshi et al., 2017) or NewsQA (Trischler et al., 2016)) to augment
the small SQuAD training set. Human performance on SQuAD is estimated at 82.30 and
91.22 for the Exact Match and F1 metric respectively (Rajpurkar et al., 2016), so it is not
clear if further improvements on this benchmark are meaningful.
For SuperGLUE, we improved upon the state-of-the-art by a large margin (from an
average score of 84.6 (Liu et al., 2019c) to 88.9). SuperGLUE was designed to include
tasks that were “beyond the scope of current state-of-the-art systems, but solvable by most
college-educated English speakers” (Wang et al., 2019b). We nearly match the human
performance of 89.8 (Wang et al., 2019b). Interestingly, on the reading comprehension tasks
(MultiRC and ReCoRD) we exceed human performance by a large margin, suggesting the
evaluation metrics used for these tasks may be biased towards machine-made predictions.
On the other hand, humans achieve 100% accuracy on both COPA and WSC, which is
significantly better than our model’s performance. This suggests that there remain linguistic
tasks that are hard for our model to perfect, particularly in the low-resource setting.
We did not achieve state-of-the-art performance on any of the WMT translation tasks.
This may be in part due to our use of an English-only unlabeled data set. We also note that
most of the best results on these tasks use backtranslation (Edunov et al., 2018; Lample and
Conneau, 2019), which is a sophisticated data augmentation scheme. The state of the art on
the low-resource English to Romanian benchmark also uses additional forms of cross-lingual
unsupervised training (Lample and Conneau, 2019). Our results suggest that scale and
English-language pre-training may be insufficient to match the performance of these more
sophisticated methods. On a more specific note, the best results on English to German
newstest2014 set use the much larger training set from WMT 2018 (Edunov et al., 2018),
making direct comparison to our results difficult.
Finally, on CNN/Daily Mail we attain state-of-the-art performance, though only by
a significant amount on the ROUGE-2-F score. It has been shown that improvements
to the ROUGE score do not necessarily correspond to more coherent summaries (Paulus
et al., 2017). Furthermore, while CNN/Daily Mail is posed as an abstractive summarization
benchmark, purely extractive approaches have been shown to work well (Liu, 2019). It has
also been argued that generative models trained with maximum likelihood are prone to
producing repetitive summaries (See et al., 2017). Despite these potential issues, we find
that our models do generate coherent and largely correct summaries. We provide some
non-cherry-picked validation set examples in Appendix C.
To achieve its strong results, T5 combines insights from our experimental study with
unprecedented scale. Note that in Section 3.6 we found that scaling up the pre-training
amount or size of our baseline model produced substantial gains. Given this, we were
interested to measure how much the “non-scaling” changes we introduced into T5 contributed
to its strong performance. We therefore carried out a final experiment where we compared
the following three configurations: First, the standard baseline model, which was pre-trained
on 235 ≈ 34B tokens; second, the baseline trained instead for about 1 trillion tokens (i.e.
the same amount of pre-training used for T5), which we refer to as “baseline-1T”; and
third, T5-Base. Note that the differences between baseline-1T and T5-Base comprise the
40
Exploring the Limits of Transfer Learning
Model GLUE CNNDM SQuAD SGLUE EnDe EnFr EnRo
⋆ Baseline 83.28 19.24 80.88 71.36 26.98 39.82 27.65
Baseline-1T 84.80 19.62 83.01 73.90 27.46 40.30 28.34
T5-Base 85.97 20.90 85.44 75.64 28.37 41.37 28.98
Table 15: Performance comparison of T5-Base to our baseline experimental setup used in
the rest of the paper. Results are reported on the validation set. “Baseline-1T”
refers to the performance achieved by pre-training the baseline model on 1 trillion
tokens (the same number used for the T5 model variants) instead of 235 ≈ 34B
tokens (as was used for the baseline).
“non-scaling” changes we made when designing T5. As such, comparing the performance of
these two models gives us a concrete measurement of the impact of the insights from our
systematic study.
The performance of these three model configurations is shown in Table 15. Consistent
with the findings in Section 3.6, we find that additional pre-training improves performance
over the baseline. Nevertheless, T5-Base substantially outperforms baseline-1T on all
downstream tasks. This suggests that scale is not the only factor that contributes to T5’s
success. We hypothesize that the larger models benefit not only from their increased size
but also from these non-scaling factors.
4. Reflection
Having completed our systematic study, we wrap up by first recapping some of our most
significant findings. Our results provide some high-level perspective on which avenues of
research might be more or less promising. To conclude, we outline some topics we think
might provide effective approaches for further progressing the field.
4.1 Takeaways
Text-to-text Our text-to-text framework provides a simple way to train a single model
on a wide variety of text tasks using the same loss function and decoding procedure.
We showed how this approach can be successfully applied to generative tasks like
abstractive summarization, classification tasks like natural language inference, and
even regression tasks like STS-B. In spite of its simplicity, we found the text-to-
text framework obtained comparable performance to task-specific architectures and
ultimately produced state-of-the-art results when combined with scale.
Architectures While some work on transfer learning for NLP has considered architectural
variants of the Transformer, we found the original encoder-decoder form worked
best in our text-to-text framework. Though an encoder-decoder model uses twice as
many parameters as “encoder-only” (e.g. BERT) or “decoder-only” (language model)
architectures, it has a similar computational cost. We also showed that sharing the
parameters in the encoder and decoder did not result in a substantial performance
drop while halving the total parameter count.
41
Raffel, Shazeer, Roberts, Lee, Narang, Matena, Zhou, Li and Liu
Unsupervised objectives Overall, we found that most “denoising” objectives, which train
the model to reconstruct randomly corrupted text, performed similarly in the text-to-
text setup. As a result, we suggest using objectives that produce short target sequences
so that unsupervised pre-training is more computationally efficient.
Data sets We introduced the “Colossal Clean Crawled Corpus” (C4), which comprises
heuristically-cleaned text from the Common Crawl web dump. When comparing C4 to
data sets that use additional filtering, we found that training on in-domain unlabeled
data could boost performance in a few downstream tasks. However, constraining to
a single domain typically results in a smaller data set. We separately showed that
performance can degrade when an unlabeled data set is small enough that it is repeated
many times over the course of pre-training. This motivates the use of a large and
diverse data set like C4 for generic language understanding tasks.
Training strategies We found that the basic approach of updating all of a pre-trained
model’s parameters during fine-tuning outperformed methods that are designed to
update fewer parameters, although updating all parameters is most expensive. We also
experimented with various approaches for training the model on multiple tasks at once,
which in our text-to-text setting simply corresponds to mixing examples from different
data sets when constructing batches. The primary concern in multi-task learning is
setting the proportion of each task to train on. We ultimately did not find a strategy
for setting mixing proportions that matched the performance of the basic approach of
unsupervised pre-training followed by supervised fine-tuning. However, we found that
fine-tuning after pre-training on a mixture of tasks produced comparable performance
to unsupervised pre-training.
Scaling We compared various strategies for taking advantage of additional compute, includ-
ing training the model on more data, training a larger model, and using an ensemble
of models. We found each approach conferred a significant boost in performance,
though training a smaller model on more data was often outperformed by training
a larger model for fewer steps. We also showed an ensemble of models can provide
substantially better results than a single model, which provides an orthogonal means
of leveraging additional computation. Ensembling models that were fine-tuned from
the same base pre-trained model performed worse than pre-training and fine-tuning
all models completely separately, though fine-tune-only ensembling still substantially
outperformed a single model.
Pushing the limits We combined our above insights and trained substantially larger
models (up to 11 billion parameters) to achieve state-of-the-art results across many of
the benchmarks we considered. For unsupervised training, we extracted text from our
C4 data set and applied a denoising objective that corrupts contiguous spans of tokens.
We pre-trained on a multi-task mixture before fine-tuning on individual tasks. Overall,
our models were trained on over 1 trillion tokens. In the interest of facilitating the
replication, extension, and application of our results, we release our code, the C4 data
set, and pre-trained model weights for each T5 variant.1
42
Exploring the Limits of Transfer Learning
4.2 Outlook
The inconvenience of large models An unsurprising but important result from our
study is that larger models tend to perform better. The fact that the hardware used for
running these models is continually getting cheaper and more powerful suggests that
scaling up may continue to be a promising way to achieve better performance (Sutton,
2019). However, it will always be the case that there are applications and scenarios
where using a smaller or less expensive model is helpful, for example when performing
client-side inference or federated learning (Konečnỳ et al., 2015, 2016). Relatedly, one
beneficial use of transfer learning is the possibility of attaining good performance on
low-resource tasks. Low-resource tasks often occur (by definition) in settings where
one lacks the assets to label more data. It follows that low-resource applications often
also have limited access to computational resources which can incur additional costs.
As a result, we advocate for research on methods that achieve stronger performance
with cheaper models so that transfer learning can be applied where it will have the
most impact. Some current work along these lines include distillation (Hinton et al.,
2015; Sanh et al., 2019; Jiao et al., 2019), parameter sharing (Lan et al., 2019), and
conditional computation (Shazeer et al., 2017).
More efficient knowledge extraction Recall that one of the goals of pre-training is
(loosely speaking) to provide the model with general-purpose “knowledge” that improves
its performance on downstream tasks. The method we use in this work, which is
currently common practice, is to train the model to denoise corrupted spans of text.
We suspect that this simplistic technique may not be a very efficient way to teach the
model general-purpose knowledge. More concretely, it would be useful to be able to
attain good fine-tuning performance without needing to train our models on 1 trillion
tokens of text first. Some concurrent work along these lines improves efficiency by
pre-training a model to distinguish between real and machine-generated text (Clark
et al., 2020).
Formalizing the similarity between tasks We observed that pre-training on unlabeled
in-domain data can improve performance on downstream tasks (Section 3.4). This
finding mostly relies on basic observations like the fact that SQuAD was created using
data from Wikipedia. It would be useful to formulate a more rigorous notion of the
“similarity” between the pre-training and downstream tasks, so that we could make
more principled choices about what source of unlabeled data to use. There is some
early empirical work along these lines in the field of computer vision (Huh et al., 2016;
Kornblith et al., 2018; He et al., 2018). A better notion of the relatedness of tasks could
also help choose supervised pre-training tasks, which has been shown to be helpful for
the GLUE benchmark (Phang et al., 2018).
Language-agnostic models We were disappointed to find that English-only pre-training
did not achieve state-of-the-art results on the translation tasks we studied. We also
are interested in avoiding the logistical difficulty of needing to specify which languages
a vocabulary can encode ahead of time. To address these issues, we are interested in
further investigating language-agnostic models, i.e. models that can perform a given
NLP task with good performance regardless of the text’s language. This is an especially
43
Raffel, Shazeer, Roberts, Lee, Narang, Matena, Zhou, Li and Liu
pertinent issue given that English is not the native language for the majority of the
world’s population.
The motivation for this paper was the flurry of recent work on transfer learning for
NLP. Before we began this work, these advances had already enabled breakthroughs in
settings where learning-based methods had not yet been shown to be effective. We are
happy to be able to continue this trend, for example by nearly matching human-level
performance on the SuperGLUE benchmark, a task specifically designed to be difficult
for modern transfer-learning pipelines. Our results stem from the combination of a
straightforward and unified text-to-text framework, our new C4 data set, and insights
from our systematic study. Additionally, we provided an empirical overview of the
field and a perspective on where it stands. We are excited to see continued work using
transfer learning towards the goal of general language understanding.
Acknowledgments
We thank Grady Simon, Noah Fiedel, Samuel R. Bowman, Augustus Odena, Daphne Ippolito,
Noah Constant, Orhan Firat, Ankur Bapna, and Sebastian Ruder for their comments on
this manuscript; Zak Stone and the TFRC team for their support; Austin Tarango for
his guidance on data set creation; Melvin Johnson, Dima Lepikhin, Katrin Tomanek, Jeff
Klingner, and Naveen Arivazhagan for insight into multi-task machine translation; Neil
Houlsby for comments on adapter layers; Olga Wichowska, Ola Spyra, Michael Banfield,
Yi Lin, and Frank Chen for assistance with infrastructure; Etienne Pot, Ryan Sepassi, and
Pierre Ruyssen for collaboration on TensorFlow Datasets; Rohan Anil for help with our
download pipeline for Common Crawl; Robby Neale and Taku Kudo for their work on
SentencePiece; Jeffrey Li for pointing out missing details about the creation of C4; and
many other members of the Google Brain team for their discussion and insight.
44
Exploring the Limits of Transfer Learning
Appendix A. Contributions
Colin designed the scope of this project and wrote this paper, ran all the experiments in
Sections 3.1 to 3.6, and contributed a large portion of our codebase. Noam contributed
many of the ideas, including the text-to-text framework, unsupervised objectives, and
data set mixing strategies; implemented our base Transformer model and its architectural
variants; and ran the experiments in Section 3.7. Adam oversaw all engineering aspects
for this project, created the C4 data set, implemented our data set pipeline, and added
various benchmark data sets. Katherine coordinated experiments, wrote and updated
documentation, ran experiments to help design our baseline, and contributed to many parts
of our codebase. Sharan contributed some of the required data sets and preprocessors, and
ran assorted preliminary experiments, in addition to co-leading the open-sourcing of our
codebase. Michael owned all aspects of the Winograd data sets, ingested many of the data
sets we used, contributed various improvements and fixes to our infrastructure, and ran some
preliminary experiments. Yanqi ran experiments and implemented methods to help settle on
a reasonable baseline and helped with the final fine-tuning of the models in Section 3.7. Wei
also helped with final fine-tuning and improved some of our preprocessors. Peter prototyped
an early version of the pre-training data set and resolved issues pertaining to the SQuAD
and CNN/DM tasks. All authors helped set the scope and research direction we followed in
this work.
Appendix B. Converting WNLI to Our Text-to-Text Format
Note that as discussed in Section 2.4, we do not train on any of the data from WNLI. Instead,
when evaluating on the WNLI test set (for the results in Section 3.7), we convert the WNLI
test set to the “referent noun prediction” text-to-text format so that we can evaluate using a
model trained on WSC and DPR. Our WNLI preprocessor is inspired by the one proposed
by He et al. (2019). Recall that examples from WNLI consist of a premise, a hypothesis,
and a label that indicates whether the hypothesis is True or False. Using the example from
Section 2.4, the hypothesis would be “The city councilmen refused the demonstrators a
permit because they feared violence.” with the premise “The demonstrators feared violence.”
and the label False. We first find the location of all pronouns in the premise (“they” in
our example). Then, we find the maximum number of words that precede or follow each
pronoun that are a substring in the hypothesis (“feared violence” in our example), ignoring
case and punctuation. When the premise contains multiple candidate pronouns, we choose
the pronoun that is preceded or followed by the largest substring of the hypothesis. We then
highlight the pronoun in the premise by surrounding it with asterisks. For the candidate
noun (which is compared to our model’s prediction to obtain a True or False label), we
remove the matching substring from the hypothesis and optionally make it non-possessive
(resulting in “the demonstrators”).
Appendix C. Example Predictions on CNN/Daily Mail
To show that our model is generating fluent summaries, we include a few example decodes
from our best model (T5-11B) on the validation set along with the ground-truth summaries.
These examples selected at random and were not cherry-picked.
45
Raffel, Shazeer, Roberts, Lee, Narang, Matena, Zhou, Li and Liu
1. Ground-truth: leopard gave up after spiky creature refused to back down
in fight in kruger national park, south africa . wildlife enthusiast lisl
moolman, 41, caught the bizarre battle while out on the road and armed with
her camera . the leopard and porcupine tumbled out of the bushes and began
to fight by roadside - watched by ms moolman .
Prediction: leopard tried to make lunch out of a plucky porcupine in kruger
national park, south africa . but the predator was put firmly in its place
after the spiky creature refused to back down during a fight . wildlife
enthusiast lisl moolman, 41, caught the bizarre battle while out on the
road and armed with her camera .
2. Ground-truth: researchers say homes are a hotbed of undiscovered species
. study of 40 north carolina homes found 8,000 bacterial and archaeal taxa
. 11 houses in california revealed a great variety of fungus among us .
Prediction: researchers estimate that the indoor biome makes up about 0.5
percent of ice-free land, or about 247,000 square miles, almost the size
of texas . a study of 40 houses in north carolina used genomic technologies
to document more than 8,000 bacterial and archaeal taxa . another study
of 50 houses in the same state described more than 750 types of arthropods,
including insects and spiders . an examination of 11 houses in california
found a great variety of fungus among us .
3. Ground-truth: executions have been temporarily halted in the state after
the drug used for lethal injections appeared faulty on monday night . kelly
gissendaner had her death delayed for a second time when it was found that
the lethal pentobarbital dose appeared cloudy . state officials dithered
over the decision , changing their mind three times before deciding not
to take the risk that it wouldn’t work . it is ’essential that executions
are carried out in a constitutional manner,’ said georgia attorney general
sam olens .
Prediction: the execution of kelly gissendaner, the only woman on georgia’s
death row, was delayed for a second time on monday . the execution team
found that the lethal pentobarbital dose appeared cloudy . the cloudy drug
bolstered death penalty opponents, who have been vocal in their opposition
after three botched executions in other parts of the country .
4. Ground-truth: dani alves was not selected for the brazil squad to face
france and chile . barcelona defender was filmed serving up hot dogs to
people on saturday . this week alves released a charity single with former
team-mat jose pinto . alves looks set to leave barcelona on a free transfer
this summer . reports in spanish press claim he will sign for paris saint-germain
.
Prediction: dani alves was not selected for brazil’s friendlies against
france and chile . the barcelona right back has released a charity single
with jose pinto . alves was filmed serving up snacks from behind the counter
46
Exploring the Limits of Transfer Learning
of a hot dog bar on saturday . alves is yet to be offered a new deal at
the nou camp . click here for all the latest barcelona news .
Appendix D. Preprocessed Examples
In this section, we provide examples of our preprocessing for each of the data sets we consider.
D.1 CoLA
Original input:
Sentence: John made Bill master of himself.
Processed input: cola sentence: John made Bill master of himself.
Original target: 1
Processed target: acceptable
D.2 RTE
Original input:
Sentence 1: A smaller proportion of Yugoslavia’s Italians were settled in Slovenia
(at the 1991 national census, some 3000 inhabitants of Slovenia declared
themselves as ethnic Italians).
Sentence 2: Slovenia has 3,000 inhabitants.
Processed input: rte sentence1: A smaller proportion of Yugoslavia’s Italians
were settled in Slovenia (at the 1991 national census, some 3000 inhabitants
of Slovenia declared themselves as ethnic Italians). sentence2: Slovenia
has 3,000 inhabitants.
Original target: 1
Processed target: not_entailment
D.3 MNLI
Original input:
Hypothesis: The St. Louis Cardinals have always won.
Premise: yeah well losing is i mean i’m i’m originally from Saint Louis and
Saint Louis Cardinals when they were there were uh a mostly a losing team
but
Processed input: mnli hypothesis: The St. Louis Cardinals have always won. premise:
yeah well losing is i mean i’m i’m originally from Saint Louis and Saint Louis
Cardinals when they were there were uh a mostly a losing team but
47
Raffel, Shazeer, Roberts, Lee, Narang, Matena, Zhou, Li and Liu
Original target: 2
Processed target: contradiction
D.4 MRPC
Original input:
Sentence 1: We acted because we saw the existing evidence in a new light ,
through the prism of our experience on 11 September , " Rumsfeld said .
Sentence 2: Rather , the US acted because the administration saw " existing
evidence in a new light , through the prism of our experience on September
11 " .
Processed input: mrpc sentence1: We acted because we saw the existing evidence
in a new light , through the prism of our experience on 11 September , " Rumsfeld
said . sentence2: Rather , the US acted because the administration saw "
existing evidence in a new light , through the prism of our experience on
September 11 " .
Original target: 1
Processed target: equivalent
D.5 QNLI
Original input:
Question: Where did Jebe die?
Sentence: Genghis Khan recalled Subutai back to Mongolia soon afterwards, and
Jebe died on the road back to Samarkand.
Processed input: qnli question: Where did Jebe die? sentence: Genghis Khan recalled
Subutai back to Mongolia soon afterwards, and Jebe died on the road back to
Samarkand.
Original target: 0
Processed target: entailment
D.6 QQP
Original input:
Question 1: What attributes would have made you highly desirable in ancient
Rome?
Question 2: How I GET OPPERTINUTY TO JOIN IT COMPANY AS A FRESHER?
Processed input: qqp question1: What attributes would have made you highly desirable
in ancient Rome? question2: How I GET OPPERTINUTY TO JOIN IT COMPANY AS A
FRESHER?
48
Exploring the Limits of Transfer Learning
Original target: 0
Processed target: not_duplicate
D.7 SST2
Original input:
Sentence: it confirms fincher ’s status as a film maker who artfully bends
technical know-how to the service of psychological insight .
Processed input: sst2 sentence: it confirms fincher ’s status as a film maker
who artfully bends technical know-how to the service of psychological insight
.
Original target: 1
Processed target: positive
D.8 STSB
Original input:
Sentence 1: Representatives for Puretunes could not immediately be reached
for comment Wednesday.
Sentence 2: Puretunes representatives could not be located Thursday to comment
on the suit.
Processed input: stsb sentence1: Representatives for Puretunes could not immediately
be reached for comment Wednesday. sentence2: Puretunes representatives could
not be located Thursday to comment on the suit.
Original target: 3.25
Processed target: 3.2
D.9 CB
Original input:
Hypothesis: Valence was helping
Premise: Valence the void-brain, Valence the virtuous valet. Why couldn’t
the figger choose his own portion of titanic anatomy to shaft? Did he think
he was helping?
Processed input: cb hypothesis: Valence was helping premise: Valence the void-brain,
Valence the virtuous valet. Why couldn’t the figger choose his own portion
of titanic anatomy to shaft? Did he think he was helping?
Original target: 1
Processed target: contradiction
49
Raffel, Shazeer, Roberts, Lee, Narang, Matena, Zhou, Li and Liu
D.10 COPA
Original input:
Question: effect
Premise: Political violence broke out in the nation.
Choice 1: Many citizens relocated to the capitol.
Choice 2: Many citizens took refuge in other territories.
Processed input: copa choice1: Many citizens relocated to the capitol. choice2:
Many citizens took refuge in other territories. premise: Political violence
broke out in the nation. question: effect
Original target: 1
Processed target: True
D.11 MultiRC
Original input:
Answer: There was only pie to eat, rather than traditional breakfast foods
Paragraph: <b>Sent 1: </b>Once upon a time, there was a squirrel named Joey.<br><b>Sent
2: </b>Joey loved to go outside and play with his cousin Jimmy.<br><b>Sent
3: </b>Joey and Jimmy played silly games together, and were always laughing.<br><b>Sent
4: </b>One day, Joey and Jimmy went swimming together at their Aunt Julie’s
pond.<br><b>Sent 5: </b>Joey woke up early in the morning to eat some food
before they left.<br><b>Sent 6: </b>He couldn’t find anything to eat except
for pie!<br><b>Sent 7: </b>Usually, Joey would eat cereal, fruit (a pear),
or oatmeal for breakfast.<br><b>Sent 8: </b>After he ate, he and Jimmy went
to the pond.<br><b>Sent 9: </b>On their way there they saw their friend
Jack Rabbit.<br><b>Sent 10: </b>They dove into the water and swam for several
hours.<br><b>Sent 11: </b>The sun was out, but the breeze was cold.<br><b>Sent
12: </b>Joey and Jimmy got out of the water and started walking home.<br><b>Sent
13: </b>Their fur was wet, and the breeze chilled them.<br><b>Sent 14: </b>When
they got home, they dried off, and Jimmy put on his favorite purple shirt.<br><b>Sent
15: </b>Joey put on a blue shirt with red and green dots.<br><b>Sent 16:
</b>The two squirrels ate some food that Joey’s mom, Jasmine, made and went
off to bed.<br>
Question: Why was Joey surprised the morning he woke up for breakfast?
Processed input: multirc question: Why was Joey surprised the morning he woke
up for breakfast? answer: There was only pie to eat, rather than traditional
breakfast foods paragraph: <b>Sent 1: </b>Once upon a time, there was a squirrel
named Joey.<br><b>Sent 2: </b>Joey loved to go outside and play with his cousin
Jimmy.<br><b>Sent 3: </b>Joey and Jimmy played silly games together, and were
always laughing.<br><b>Sent 4: </b>One day, Joey and Jimmy went swimming together
50
Exploring the Limits of Transfer Learning
at their Aunt Julie’s pond.<br><b>Sent 5: </b>Joey woke up early in the morning
to eat some food before they left.<br><b>Sent 6: </b>He couldn’t find anything
to eat except for pie!<br><b>Sent 7: </b>Usually, Joey would eat cereal, fruit
(a pear), or oatmeal for breakfast.<br><b>Sent 8: </b>After he ate, he and
Jimmy went to the pond.<br><b>Sent 9: </b>On their way there they saw their
friend Jack Rabbit.<br><b>Sent 10: </b>They dove into the water and swam for
several hours.<br><b>Sent 11: </b>The sun was out, but the breeze was cold.<br><b>Sent
12: </b>Joey and Jimmy got out of the water and started walking home.<br><b>Sent
13: </b>Their fur was wet, and the breeze chilled them.<br><b>Sent 14: </b>When
they got home, they dried off, and Jimmy put on his favorite purple shirt.<br><b>Sent
15: </b>Joey put on a blue shirt with red and green dots.<br><b>Sent 16: </b>The
two squirrels ate some food that Joey’s mom, Jasmine, made and went off to
bed.<br>
Original target: 1
Processed target: True
D.12 WiC
Original input:
POS: N
Sentence 1: It was the deliberation of his act that was insulting .
Sentence 2: The deliberations of the jury .
Word: deliberation
Processed input: wic pos: N sentence1: It was the deliberation of his act that
was insulting . sentence2: The deliberations of the jury . word: deliberation
Original target: 0
Processed target: False
D.13 WSC and DPR
Original input:
Span 2 text: it
Span 1 text: stable
Span 2 index: 20
Span 1 index: 1
Text: The stable was very roomy, with four good stalls; a large swinging window
opened into the yard , which made it pleasant and airy.
Processed input: wsc: The stable was very roomy, with four good stalls; a large
swinging window opened into the yard , which made *it* pleasant and airy.
51
Raffel, Shazeer, Roberts, Lee, Narang, Matena, Zhou, Li and Liu
Original target: 1
Processed target: stable
D.14 CNN/Daily Mail
Original input: marouane fellaini and adnan januzaj continue to show the world
they are not just teammates but also best mates. the manchester united and
belgium duo both posted pictures of themselves out at a restaurant on monday
night ahead of their game against newcastle on wednesday . januzaj poses
in the middle of fellaini and a friend looking like somebody who failed to
receive the memo about it being a jackson 5 themed night. premier league
duo adnan januzaj and marouane fellaini pose with a friend on the dance floor
. manchester united and belgium duo fellaini and januzaj are good friends
both on and off the pitch . manchester united ace fellaini runs over to the
bench to celebrate his goal against qpr with friend januzaj . the disco effect
in the background adds to the theory, but januzaj doesn’t seem to mind as
they later pose on the dance floor with other friends. united haven’t had
too many reasons to have a song and dance this season so it seems they may
be hitting the discotheques as another form of release. however, victory against
newcastle on wednesday would leave manager louis van gaal at least tapping
his toes as they continue to fight for a champions league spot this season.
januzaj and robin van persie join fellaini in celebrating in front of the
manchester united fans at west brom . januzaj receives some words of wisdom
from manchester united’s dutch manager louis van gaal . januzaj and fellaini
are joined by some friends as they take to the dance floor ahead of the newcastle
game .
Processed input: summarize: marouane fellaini and adnan januzaj continue to show
the world they are not just teammates but also best mates. the manchester
united and belgium duo both posted pictures of themselves out at a restaurant
on monday night ahead of their game against newcastle on wednesday . januzaj
poses in the middle of fellaini and a friend looking like somebody who failed
to receive the memo about it being a jackson 5 themed night. premier league
duo adnan januzaj and marouane fellaini pose with a friend on the dance floor
. manchester united and belgium duo fellaini and januzaj are good friends
both on and off the pitch . manchester united ace fellaini runs over to the
bench to celebrate his goal against qpr with friend januzaj . the disco effect
in the background adds to the theory, but januzaj doesn’t seem to mind as
they later pose on the dance floor with other friends. united haven’t had
too many reasons to have a song and dance this season so it seems they may
be hitting the discotheques as another form of release. however, victory against
newcastle on wednesday would leave manager louis van gaal at least tapping
his toes as they continue to fight for a champions league spot this season.
januzaj and robin van persie join fellaini in celebrating in front of the
manchester united fans at west brom . januzaj receives some words of wisdom
52
Exploring the Limits of Transfer Learning
from manchester united’s dutch manager louis van gaal . januzaj and fellaini
are joined by some friends as they take to the dance floor ahead of the newcastle
game .
Original target: the belgian duo took to the dance floor on monday night with
some friends . manchester united face newcastle in the premier league on
wednesday . red devils will be looking for just their second league away win
in seven . louis van gaal’s side currently sit two points clear of liverpool
in fourth .
Processed target: the belgian duo took to the dance floor on monday night with
some friends . manchester united face newcastle in the premier league on
wednesday . red devils will be looking for just their second league away win
in seven . louis van gaal’s side currently sit two points clear of liverpool
in fourth .
D.15 SQuAD
Original input:
Question: What does increased oxygen concentrations in the patient’s lungs
displace?
Context: Hyperbaric (high-pressure) medicine uses special oxygen chambers
to increase the partial pressure of O 2 around the patient and, when needed,
the medical staff. Carbon monoxide poisoning, gas gangrene, and decompression
sickness (the ’bends’) are sometimes treated using these devices. Increased
O 2 concentration in the lungs helps to displace carbon monoxide from the
heme group of hemoglobin. Oxygen gas is poisonous to the anaerobic bacteria
that cause gas gangrene, so increasing its partial pressure helps kill them.
Decompression sickness occurs in divers who decompress too quickly after
a dive, resulting in bubbles of inert gas, mostly nitrogen and helium, forming
in their blood. Increasing the pressure of O 2 as soon as possible is part
of the treatment.
Processed input: question: What does increased oxygen concentrations in the patient’s
lungs displace? context: Hyperbaric (high-pressure) medicine uses special
oxygen chambers to increase the partial pressure of O 2 around the patient
and, when needed, the medical staff. Carbon monoxide poisoning, gas gangrene,
and decompression sickness (the ’bends’) are sometimes treated using these
devices. Increased O 2 concentration in the lungs helps to displace carbon
monoxide from the heme group of hemoglobin. Oxygen gas is poisonous to the
anaerobic bacteria that cause gas gangrene, so increasing its partial pressure
helps kill them. Decompression sickness occurs in divers who decompress too
quickly after a dive, resulting in bubbles of inert gas, mostly nitrogen and
helium, forming in their blood. Increasing the pressure of O 2 as soon as
possible is part of the treatment.
53
Raffel, Shazeer, Roberts, Lee, Narang, Matena, Zhou, Li and Liu
Original target: carbon monoxide
Processed target: carbon monoxide
D.16 WMT English to German
Original input: "Luigi often said to me that he never wanted the brothers to end
up in court," she wrote.
Processed input: translate English to German: "Luigi often said to me that he
never wanted the brothers to end up in court," she wrote.
Original target: "Luigi sagte oft zu mir, dass er nie wollte, dass die Brüder
vor Gericht landen", schrieb sie.
Processed target: "Luigi sagte oft zu mir, dass er nie wollte, dass die Brüder
vor Gericht landen", schrieb sie.
D.17 WMT English to French
Original input: This image section from an infrared recording by the Spitzer telescope
shows a "family portrait" of countless generations of stars: the oldest stars
are seen as blue dots, while more difficult to identify are the pink-coloured
"new-borns" in the star delivery room.
Processed input: translate English to French: This image section from an infrared
recording by the Spitzer telescope shows a "family portrait" of countless
generations of stars: the oldest stars are seen as blue dots, while more difficult
to identify are the pink-coloured "new-borns" in the star delivery room.
Original target: Ce détail d’une photographie infrarouge prise par le télescope
Spitzer montre un "portrait de famille" des innombrables générations d’étoiles:
les plus vieilles étoiles sont en bleu et les points roses, plus difficiles
à identifier, sont les "nouveau-nés" dans la salle d’accouchement de l’univers.
Processed target: Ce détail d’une photographie infrarouge prise par le télescope
Spitzer montre un "portrait de famille" des innombrables générations d’étoiles:
les plus vieilles étoiles sont en bleu et les points roses, plus difficiles
à identifier, sont les "nouveau-nés" dans la salle d’accouchement de l’univers.
D.18 WMT English to Romanian
Original input: Taco Bell said it plans to add 2,000 locations in the US by 2022.
Processed input: translate English to Romanian: Taco Bell said it plans to add
2,000 locations in the US by 2022.
Original target: Taco Bell a afirmat că, până în 2022, intent, ionează să deschidă
2000 de restaurante în SUA.
54
Exploring the Limits of Transfer Learning
Processed target: Taco Bell a afirmat că, până în 2022, intent, ionează să deschidă
2000 de restaurante în SUA.
55
Raffel, Shazeer, Roberts, Lee, Narang, Matena, Zhou, Li and Liu
Appendix E. Scores on Every Task for All Experiments
The following table lists the scores achieved on every task in the experiments described in
Sections 3.2 to 3.6.
56
GLUE SuperGLUE WMT
Score CoLA SST-2 MRPC MRPC STSB STSB QQP QQP MNLIm MNLImm QNLI RTE CNN/DM SQuAD Score BoolQ CB CB COPA MultiRC MultiRC ReCoRD ReCoRD RTE WiC WSC EnDe EnFr EnRo
Table Experiment Average MCC Acc F1 Acc PCC SCC F1 Acc Acc Acc Acc Acc R-1-F R-2-F R-L-F EM F1 Average Acc F1 Acc Acc F1 EM F1 EM Acc Acc Acc BLEU BLEU BLEU
1 ⋆ Baseline average 83.28 53.84 92.68 92.07 88.92 88.02 87.94 88.67 91.56 84.24 84.57 90.48 76.28 41.33 19.24 38.77 80.88 88.81 71.36 76.62 91.22 91.96 66.20 66.13 25.78 69.05 68.16 75.34 68.04 78.56 26.98 39.82 27.65
1 Baseline standard deviation 0.235 1.111 0.569 0.729 1.019 0.374 0.418 0.108 0.070 0.291 0.231 0.361 1.393 0.065 0.065 0.058 0.343 0.226 0.416 0.365 3.237 2.560 2.741 0.716 1.011 0.370 0.379 1.228 0.850 2.029 0.112 0.090 0.108
1 No pre-training 66.22 12.29 80.62 81.42 73.04 72.58 72.97 81.94 86.62 68.02 67.98 75.69 58.84 39.19 17.60 36.69 50.31 61.97 53.04 65.38 71.61 76.79 62.00 59.10 0.84 20.33 17.95 54.15 54.08 65.38 25.86 39.77 24.04
2 ⋆ Enc/dec, denoising 83.28 53.84 92.68 92.07 88.92 88.02 87.94 88.67 91.56 84.24 84.57 90.48 76.28 41.33 19.24 38.77 80.88 88.81 71.36 76.62 91.22 91.96 66.20 66.13 25.78 69.05 68.16 75.34 68.04 78.56 26.98 39.82 27.65
2 Enc/dec, shared, denoising 82.81 55.24 91.86 91.58 88.24 87.43 87.58 88.69 91.60 83.88 84.01 90.23 73.65 41.11 18.78 38.48 80.63 88.49 70.73 77.13 95.04 96.43 65.00 66.16 22.98 68.95 68.09 70.76 68.18 75.96 26.72 39.03 27.46
2 Enc/dec, 6 layers, denoising 80.88 46.26 92.09 91.51 87.99 87.01 86.76 87.93 90.97 82.20 82.41 88.83 71.48 40.83 18.97 38.31 77.59 86.07 68.42 73.79 91.70 92.86 67.00 61.02 19.62 61.26 60.33 72.20 65.99 75.00 26.38 38.40 26.95
2 Language model, denoising 74.70 24.50 90.60 86.08 78.92 85.22 85.42 85.40 88.99 76.72 77.05 86.02 64.62 39.49 17.93 36.91 61.14 71.37 55.02 65.47 60.08 71.43 58.00 43.03 2.94 53.35 52.31 53.07 58.62 63.46 25.09 35.28 25.86
2 Prefix LM, denoising 81.82 49.99 92.43 91.43 88.24 87.20 86.98 88.41 91.39 82.32 82.93 88.71 74.01 40.46 18.61 37.90 78.94 87.31 68.11 75.50 93.37 91.07 60.00 63.43 21.20 65.03 64.11 71.48 65.67 73.08 26.43 37.98 27.39
2 Enc/dec, LM 79.56 42.03 91.86 91.64 88.24 87.13 87.00 88.21 91.15 81.68 81.66 88.54 65.70 40.67 18.59 38.13 76.02 84.85 64.29 72.23 85.74 89.29 57.00 60.53 16.26 59.28 58.30 65.34 64.89 70.19 26.27 39.17 26.86
2 Enc/dec, shared, LM 79.60 44.83 92.09 90.20 85.78 86.03 85.87 87.77 91.02 81.74 82.29 89.16 65.34 40.16 18.13 37.59 76.35 84.86 63.50 70.49 91.41 87.50 55.00 60.21 16.89 57.83 56.73 63.54 63.48 70.19 26.62 39.17 27.05
2 Enc/dec, 6 layers, LM 78.67 38.72 91.40 90.40 86.52 86.82 86.49 87.87 91.03 80.99 80.92 88.05 65.70 40.29 18.26 37.70 75.32 84.06 64.06 71.38 85.25 89.29 60.00 57.56 16.79 55.22 54.30 66.79 63.95 71.15 26.13 38.42 26.89
2 Language model, LM 73.78 28.53 89.79 85.23 78.68 84.22 84.00 84.88 88.70 74.94 75.77 84.84 58.84 38.97 17.54 36.37 53.81 64.55 56.51 64.22 59.92 71.43 64.00 53.04 1.05 46.81 45.78 58.84 56.74 69.23 25.23 34.31 25.38
2 Prefix LM, LM 79.68 41.26 92.09 90.11 86.27 86.82 86.32 88.35 91.35 81.71 82.02 89.04 68.59 39.66 17.84 37.13 76.87 85.39 64.86 71.47 93.37 91.07 57.00 58.67 16.89 59.25 58.16 64.26 66.30 71.15 26.28 37.51 26.76
4 Language modeling with prefix 80.69 44.22 93.00 91.68 88.48 87.20 87.18 88.39 91.41 82.66 83.09 89.29 68.95 40.71 18.94 38.15 77.99 86.43 65.27 73.55 83.95 87.50 55.00 59.65 18.89 61.76 60.76 68.59 65.67 73.08 26.86 39.73 27.49
4 BERT-style (Devlin et al., 2018) 82.96 52.49 92.55 92.79 89.95 87.68 87.66 88.47 91.44 83.60 84.05 90.33 75.45 41.27 19.17 38.72 80.65 88.24 69.85 76.48 94.37 94.64 61.00 63.29 25.08 66.76 65.85 72.20 69.12 75.00 26.78 40.03 27.41
4 Deshuffling 73.17 22.82 87.16 86.88 81.13 84.03 83.82 86.38 89.90 76.30 76.34 84.18 58.84 40.75 18.59 38.10 67.61 76.76 58.47 69.17 63.70 78.57 56.00 59.85 12.70 45.52 44.36 57.04 64.89 68.27 26.11 39.30 25.62
5 BERT-style (Devlin et al., 2018) 82.96 52.49 92.55 92.79 89.95 87.68 87.66 88.47 91.44 83.60 84.05 90.33 75.45 41.27 19.17 38.72 80.65 88.24 69.85 76.48 94.37 94.64 61.00 63.29 25.08 66.76 65.85 72.20 69.12 75.00 26.78 40.03 27.41
5 MASS-style (Song et al., 2019) 82.32 47.01 91.63 92.53 89.71 88.21 88.18 88.58 91.44 82.96 83.67 90.02 77.26 41.16 19.16 38.55 80.10 88.07 69.28 75.08 84.98 89.29 63.00 64.46 23.50 66.71 65.91 72.20 67.71 78.85 26.79 39.89 27.55
5 ⋆ Replace corrupted spans 83.28 53.84 92.68 92.07 88.92 88.02 87.94 88.67 91.56 84.24 84.57 90.48 76.28 41.33 19.24 38.77 80.88 88.81 71.36 76.62 91.22 91.96 66.20 66.13 25.78 69.05 68.16 75.34 68.04 78.56 26.98 39.82 27.65
5 Drop corrupted tokens 84.44 60.04 92.89 92.79 89.95 87.28 86.85 88.56 91.54 83.94 83.92 90.74 79.42 41.27 19.31 38.70 80.52 88.28 68.67 75.90 96.02 94.64 56.00 65.06 23.92 65.54 64.60 71.12 67.40 74.04 27.07 39.76 27.82
6 Corruption rate = 10% 82.82 52.71 92.09 91.55 88.24 88.19 88.15 88.47 91.40 83.50 84.51 90.33 75.45 41.05 19.00 38.53 80.38 88.36 69.55 74.98 92.37 92.86 62.00 66.04 24.66 67.93 67.09 70.76 67.24 75.96 26.87 39.28 27.44
6 ⋆ Corruption rate = 15% 83.28 53.84 92.68 92.07 88.92 88.02 87.94 88.67 91.56 84.24 84.57 90.48 76.28 41.33 19.24 38.77 80.88 88.81 71.36 76.62 91.22 91.96 66.20 66.13 25.78 69.05 68.16 75.34 68.04 78.56 26.98 39.82 27.65
6 Corruption rate = 25% 83.00 53.47 93.00 92.44 89.46 87.36 87.36 88.68 91.53 84.44 84.15 90.77 74.01 41.69 19.54 39.14 80.96 88.61 70.48 76.39 93.02 92.86 68.00 65.46 24.66 68.20 67.39 73.65 67.87 72.12 27.04 39.83 27.47
6 Corruption rate = 50% 81.27 46.26 91.63 91.11 87.99 87.87 87.64 88.70 91.57 83.64 84.10 90.24 70.76 41.51 19.32 38.89 79.80 87.76 70.33 75.02 93.05 92.86 68.00 62.97 24.13 64.94 64.13 72.20 68.50 77.88 27.01 39.90 27.49
7 ⋆ Baseline (i.i.d.) 83.28 53.84 92.68 92.07 88.92 88.02 87.94 88.67 91.56 84.24 84.57 90.48 76.28 41.33 19.24 38.77 80.88 88.81 71.36 76.62 91.22 91.96 66.20 66.13 25.78 69.05 68.16 75.34 68.04 78.56 26.98 39.82 27.65
7 Average span length = 2 83.54 53.82 92.20 93.05 90.44 87.85 87.71 88.42 91.40 84.28 84.46 90.88 77.62 41.23 19.39 38.69 82.09 89.69 72.20 77.06 90.43 91.07 70.00 66.28 26.13 71.34 70.61 75.45 68.34 78.85 26.76 39.99 27.63
7 Average span length = 3 83.49 53.90 92.43 92.25 89.46 87.49 87.53 88.72 91.51 84.85 84.84 90.99 77.26 41.50 19.62 38.94 81.84 89.66 72.53 76.85 94.37 94.64 70.00 67.64 28.75 70.84 69.90 74.73 67.71 77.88 26.86 39.65 27.62
7 Average span length = 5 83.40 52.12 93.12 92.63 89.71 88.70 88.47 88.84 91.64 84.32 84.29 90.79 76.90 41.39 19.24 38.82 82.05 89.79 72.23 77.06 83.06 89.29 69.00 68.16 30.12 71.36 70.53 75.81 69.91 79.81 26.88 39.40 27.53
7 Average span length = 10 82.85 50.11 92.09 91.95 88.97 88.45 88.22 88.86 91.63 84.34 84.28 91.07 76.17 41.38 19.33 38.80 81.84 89.39 70.44 76.45 87.40 89.29 65.00 66.87 29.59 69.82 68.94 72.56 67.55 75.96 26.79 39.49 27.69
8 ⋆ C4 83.28 53.84 92.68 92.07 88.92 88.02 87.94 88.67 91.56 84.24 84.57 90.48 76.28 41.33 19.24 38.77 80.88 88.81 71.36 76.62 91.22 91.96 66.20 66.13 25.78 69.05 68.16 75.34 68.04 78.56 26.98 39.82 27.65
8 C4, unfiltered 81.46 48.01 91.63 92.72 89.95 87.79 87.60 88.31 91.27 82.30 82.34 88.71 72.20 41.09 19.14 38.54 78.78 87.04 68.04 75.75 89.17 91.07 62.00 65.52 25.60 62.42 61.58 69.68 67.08 72.12 26.55 39.34 27.21
8 RealNews-like 83.83 56.55 92.66 92.06 88.97 87.71 87.37 88.51 91.49 84.35 84.46 90.61 78.34 41.38 19.23 38.84 80.39 88.50 72.38 77.00 93.09 94.64 66.00 65.92 23.82 74.56 73.72 75.81 66.61 80.77 26.75 39.90 27.48
8 WebText-like 84.03 56.38 93.12 92.31 89.22 88.69 88.68 88.65 91.56 84.70 84.84 90.83 77.62 41.23 19.31 38.70 81.42 89.15 71.40 76.88 83.08 89.29 66.00 64.10 24.24 72.24 71.36 75.45 68.03 82.69 26.80 39.74 27.59
8 Wikipedia 81.85 45.53 92.32 91.67 88.24 85.62 86.40 88.37 91.34 82.61 83.25 90.96 77.26 41.39 19.31 38.81 81.29 89.18 68.01 76.12 56.03 80.36 67.00 65.01 25.92 69.03 68.06 74.73 67.08 76.92 26.94 39.69 27.67
8 Wikipedia + TBC 83.65 55.53 92.78 92.41 89.22 86.67 86.27 89.47 92.29 84.38 83.45 91.94 76.90 41.22 19.28 38.67 82.08 89.70 73.24 76.22 95.40 92.86 69.00 51.59 50.93 69.53 68.51 77.62 66.93 81.73 26.77 39.63 27.57
9 ⋆ Full data set 83.28 53.84 92.68 92.07 88.92 88.02 87.94 88.67 91.56 84.24 84.57 90.48 76.28 41.33 19.24 38.77 80.88 88.81 71.36 76.62 91.22 91.96 66.20 66.13 25.78 69.05 68.16 75.34 68.04 78.56 26.98 39.82 27.65
9 229 (64 repeats) 82.87 53.82 92.78 91.79 88.73 87.56 87.58 88.73 91.54 84.07 84.21 90.59 73.65 41.18 19.19 38.67 80.97 88.90 72.03 76.76 92.96 92.86 66.00 65.11 26.76 69.35 68.49 75.81 67.24 82.69 26.83 39.74 27.63
9 227 (256 repeats) 82.62 50.60 92.32 92.07 88.73 87.83 87.60 88.65 91.54 83.43 84.37 90.12 75.81 41.24 19.20 38.70 79.78 87.63 69.97 75.29 93.42 91.07 63.00 61.82 23.61 66.27 65.39 73.65 66.30 80.77 27.02 39.71 27.33
9 225 (1,024 repeats) 79.55 43.84 91.28 89.32 85.05 85.92 85.74 88.05 91.09 81.29 81.72 87.90 69.31 40.66 18.57 38.13 76.27 84.58 64.76 72.63 83.97 82.14 64.00 59.39 17.94 56.94 56.04 64.98 65.20 73.08 26.38 39.56 26.80
9 223 (4,096 repeats) 76.34 32.68 89.45 89.84 86.03 83.49 83.42 87.18 90.61 77.80 78.69 85.47 64.62 40.16 18.33 37.66 70.92 80.20 59.29 69.85 73.48 73.21 56.00 57.66 14.38 46.69 45.79 59.57 65.05 68.27 26.37 38.84 25.81
10 ⋆ All parameters 83.28 53.84 92.68 92.07 88.92 88.02 87.94 88.67 91.56 84.24 84.57 90.48 76.28 41.33 19.24 38.77 80.88 88.81 71.36 76.62 91.22 91.96 66.20 66.13 25.78 69.05 68.16 75.34 68.04 78.56 26.98 39.82 27.65
10 Adapter layers, d = 32 80.52 45.33 91.63 90.59 86.76 88.38 88.06 86.99 90.26 83.63 83.94 90.72 67.15 34.50 15.08 32.15 79.32 87.70 60.40 65.32 50.87 73.21 52.00 58.61 19.41 65.50 64.58 62.09 64.58 73.08 13.84 17.88 15.54
10 Adapter layers, d = 128 81.51 45.35 92.89 91.49 88.24 87.73 87.65 87.73 90.93 83.64 84.09 90.52 72.56 36.71 16.62 34.37 79.47 87.61 63.03 69.20 52.21 75.00 56.00 61.08 18.05 67.94 66.97 68.59 66.77 73.08 19.83 27.50 22.63
10 Adapter layers, d = 512 81.54 44.25 93.35 91.00 87.25 88.74 88.44 88.02 91.15 83.08 83.80 89.62 74.37 38.63 17.78 36.25 79.18 87.32 64.30 73.18 59.86 71.43 56.00 62.94 18.57 66.56 65.74 70.76 67.87 74.04 23.45 33.98 25.81
10 Adapter layers, d = 2048 82.62 49.86 92.55 91.30 87.99 88.46 88.35 88.36 91.40 83.63 83.18 90.66 76.53 39.44 18.30 37.06 79.40 87.36 68.61 74.53 88.00 91.07 58.00 61.10 18.89 66.73 66.06 73.29 71.16 75.96 25.64 36.92 26.93
10 Gradual Unfreezing 82.50 51.74 91.97 92.61 89.71 87.27 86.90 88.26 91.35 83.42 83.49 89.71 75.09 40.88 18.95 38.40 79.17 87.30 70.79 75.51 93.09 94.64 70.00 62.03 21.51 65.69 64.79 72.92 69.12 77.89 26.71 39.02 26.93
11 ⋆ Baseline (pre-train/fine-tune) 83.28 53.84 92.68 92.07 88.92 88.02 87.94 88.67 91.56 84.24 84.57 90.48 76.28 41.33 19.24 38.77 80.88 88.81 71.36 76.62 91.22 91.96 66.20 66.13 25.78 69.05 68.16 75.34 68.04 78.56 26.98 39.82 27.65
11 Equal 76.13 39.47 90.94 82.90 75.74 78.83 78.44 86.45 89.71 82.08 82.92 90.13 59.93 40.95 19.02 38.39 76.51 85.61 63.37 73.06 82.37 83.93 65.00 60.89 17.52 60.51 59.70 61.01 60.03 65.38 23.89 34.31 26.78
11 Examples-proportional, K = 216 80.45 42.07 91.97 90.97 87.50 85.41 85.04 86.89 90.10 83.01 83.66 90.74 72.56 41.16 19.04 38.59 77.25 85.72 69.95 76.67 86.38 89.29 70.00 65.93 27.91 62.78 61.95 76.90 65.83 73.08 24.35 34.99 27.10
11 Examples-proportional, K = 217 81.56 47.35 91.40 91.55 88.24 86.15 85.93 86.94 90.06 82.76 84.12 90.79 75.09 41.06 19.12 38.47 77.00 85.87 67.91 77.89 77.54 85.71 57.00 67.78 27.07 61.51 60.54 79.06 65.20 74.04 24.36 35.00 27.25
11 Examples-proportional, K = 218 81.67 46.85 91.63 91.99 88.73 87.68 87.20 86.93 90.35 83.30 84.01 91.47 73.29 40.96 19.07 38.43 78.17 86.74 67.94 76.57 78.88 87.50 62.00 67.70 30.85 63.43 62.54 76.53 65.67 67.31 24.57 35.19 27.39
11 Examples-proportional, K = 219 81.42 45.94 91.63 92.20 89.22 88.44 88.32 86.84 90.10 83.73 84.29 91.84 70.40 41.26 19.24 38.71 79.78 88.15 67.30 75.66 75.59 87.50 59.00 68.22 30.64 65.32 64.29 73.65 65.05 69.23 25.21 36.30 27.76
11 Examples-proportional, K = 220 80.80 42.55 92.78 91.27 87.99 88.36 88.10 86.10 89.62 84.15 84.26 92.20 68.95 41.05 19.24 38.46 80.36 88.27 67.38 73.21 76.18 83.93 62.00 67.57 26.86 66.12 65.22 76.90 64.73 69.23 25.66 36.93 27.68
11 Examples-proportional, K = 221 79.83 44.45 91.28 89.00 84.31 87.54 87.40 84.93 88.53 82.54 84.16 90.85 67.87 40.51 18.79 37.92 79.50 87.48 65.10 71.16 68.88 85.71 57.00 62.75 23.40 64.50 63.65 72.92 64.11 71.15 25.82 37.22 27.13
11 Temperature-scaled, T = 2 81.90 54.00 91.74 90.56 86.76 85.11 84.60 86.40 89.74 83.47 84.15 91.51 72.56 41.09 19.28 38.54 79.42 87.77 69.92 76.73 92.37 92.86 57.00 69.80 31.90 66.65 65.74 72.92 67.08 75.96 25.42 36.72 27.20
11 Temperature-scaled, T = 4 80.56 45.38 91.97 89.68 85.78 83.13 82.76 86.39 90.00 82.78 84.19 91.16 73.65 41.09 19.22 38.51 77.99 86.81 69.54 76.76 97.36 96.43 59.00 68.10 31.48 64.26 63.27 74.73 64.26 71.15 25.04 35.82 27.45
11 Temperature-scaled, T = 8 77.21 40.07 91.06 88.11 83.33 79.20 79.06 86.60 89.90 83.05 83.56 90.21 59.93 41.01 19.10 38.40 77.14 85.99 66.07 73.94 93.70 94.64 60.00 66.36 26.86 63.46 62.60 62.09 63.32 65.38 24.55 35.35 27.17
12 ⋆ Unsupervised pre-training + fine-tuning 83.28 53.84 92.68 92.07 88.92 88.02 87.94 88.67 91.56 84.24 84.57 90.48 76.28 41.33 19.24 38.77 80.88 88.81 71.36 76.62 91.22 91.96 66.20 66.13 25.78 69.05 68.16 75.34 68.04 78.56 26.98 39.82 27.65
12 Multi-task training 81.42 45.94 91.63 92.20 89.22 88.44 88.32 86.84 90.10 83.73 84.29 91.84 70.40 41.26 19.24 38.71 79.78 88.15 67.30 75.66 75.59 87.50 59.00 68.22 30.64 65.32 64.29 73.65 65.05 69.23 25.21 36.30 27.76
12 Multi-task pre-training + fine-tuning 83.11 51.42 92.66 91.73 88.73 88.06 87.70 88.61 91.61 84.09 84.31 91.85 76.53 41.15 19.12 38.59 80.26 88.50 71.03 79.54 81.69 87.50 65.00 70.72 31.48 65.94 65.03 81.23 68.18 73.08 27.08 39.80 28.07
12 Leave-one-out multi-task training 81.98 48.00 93.23 91.72 88.24 87.76 87.32 88.61 91.44 84.00 84.11 90.79 72.20 41.34 19.05 38.77 79.97 88.10 71.68 78.35 86.76 89.29 66.00 68.09 29.49 66.23 65.27 79.06 68.65 78.85 26.93 39.79 27.87
12 Supervised multi-task pre-training 79.93 36.60 92.43 91.58 88.24 87.03 86.78 88.15 91.20 82.87 83.16 90.13 70.76 41.12 18.96 38.49 77.38 85.65 65.36 75.66 68.87 83.93 58.00 64.81 21.93 55.37 54.61 71.12 67.40 75.96 26.81 40.13 28.04
13 ⋆ Baseline 83.28 53.84 92.68 92.07 88.92 88.02 87.94 88.67 91.56 84.24 84.57 90.48 76.28 41.33 19.24 38.77 80.88 88.81 71.36 76.62 91.22 91.96 66.20 66.13 25.78 69.05 68.16 75.34 68.04 78.56 26.98 39.82 27.65
13 1× size, 4× training steps 85.33 60.29 93.81 94.06 91.67 89.42 89.25 89.15 91.87 86.01 85.70 91.63 78.34 41.52 19.33 38.96 82.45 90.19 74.72 79.17 94.75 92.86 71.00 67.34 29.70 72.63 71.59 78.34 72.10 82.69 27.08 40.66 27.93
13 1× size, 4× batch size 84.60 56.08 93.12 92.31 89.22 88.85 88.84 89.35 92.07 85.98 86.13 91.07 80.14 41.70 19.42 39.08 82.52 90.21 74.64 78.78 93.69 94.64 72.00 68.09 30.95 74.73 73.90 76.53 70.06 81.73 27.07 40.60 27.84
13 2× size, 2× training steps 86.18 62.04 93.69 93.36 90.69 89.18 89.23 89.35 92.05 87.23 87.05 92.68 81.95 41.74 19.66 39.14 84.18 91.29 77.18 80.98 97.36 96.43 74.00 71.34 35.68 77.11 76.34 80.51 69.28 85.58 27.52 41.03 28.19
13 4× size, 1× training steps 85.91 57.58 94.38 92.67 89.95 89.60 89.60 89.44 92.14 87.05 87.12 93.12 83.39 41.60 19.73 39.08 83.86 91.32 78.04 81.38 89.09 94.64 73.00 73.74 40.40 78.25 77.40 81.59 70.22 91.35 27.47 40.71 28.10
13 4× ensembled 84.77 56.14 93.46 93.31 90.67 89.71 89.60 89.62 92.24 86.22 86.53 91.60 77.98 42.10 20.10 39.56 83.09 90.40 71.74 77.58 89.85 91.07 66.00 69.32 29.49 72.67 71.94 76.90 69.12 72.12 28.05 40.53 28.09
13 4× ensembled, fine-tune only 84.05 54.78 92.78 93.15 90.44 88.34 88.12 89.27 91.97 85.33 85.88 90.98 77.62 41.66 19.57 39.12 82.36 89.86 71.56 77.43 90.07 92.86 69.00 67.31 26.34 70.47 69.64 75.45 68.18 74.04 27.55 40.22 28.09
Table 16: Score achieved on every task we consider for all of the experiments in this paper. In the first column, we list the table where the condensed results were presented for a given experiment. As in the main text, a row marked with ⋆ denotes our baseline model (described in Section 3.1).
Raffel, Shazeer, Roberts, Lee, Narang, Matena, Zhou, Li and Liu
References
Rami Al-Rfou, Dokook Choe, Noah Constant, Mandy Guo, and Llion Jones. Character-level
language modeling with deeper self-attention. In Proceedings of the AAAI Conference on
Artificial Intelligence, 2019.
Rohan Anil, Vineet Gupta, Tomer Koren, and Yoram Singer. Memory-efficient adaptive
optimization for large-scale learning. arXiv preprint arXiv:1901.11150, 2019.
Naveen Arivazhagan, Ankur Bapna, Orhan Firat, Dmitry Lepikhin, Melvin Johnson, Maxim
Krikun, Mia Xu Chen, Yuan Cao, George Foster, Colin Cherry, et al. Massively multi-
lingual neural machine translation in the wild: Findings and challenges. arXiv preprint
arXiv:1907.05019, 2019.
Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E. Hinton. Layer normalization. arXiv
preprint arXiv:1607.06450, 2016.
Alexei Baevski, Sergey Edunov, Yinhan Liu, Luke Zettlemoyer, and Michael Auli. Cloze-
driven pretraining of self-attention networks. arXiv preprint arXiv:1903.07785, 2019.
Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. Neural machine translation by
jointly learning to align and translate. In Third International Conference on Learning
Representations, 2015.
Ankur Bapna, Naveen Arivazhagan, and Orhan Firat. Simple, scalable adaptation for neural
machine translation. arXiv preprint arXiv:1909.08478, 2019.
Iz Beltagy, Kyle Lo, and Arman Cohan. SciBERT: A pretrained language model for scientific
text. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language
Processing and the 9th International Joint Conference on Natural Language Processing
(EMNLP-IJCNLP), 2019.
Ondřej Bojar, Christian Buck, Christian Federmann, Barry Haddow, Philipp Koehn, Jo-
hannes Leveling, Christof Monz, Pavel Pecina, Matt Post, Herve Saint-Amand, et al.
Findings of the 2014 workshop on statistical machine translation. In Proceedings of the
Ninth Workshop on Statistical Machine Translation, 2014.
Ondřej Bojar, Rajen Chatterjee, Christian Federmann, Barry Haddow, Matthias Huck,
Chris Hokamp, Philipp Koehn, Varvara Logacheva, Christof Monz, Matteo Negri, et al.
Findings of the 2015 workshop on statistical machine translation. In Proceedings of the
Tenth Workshop on Statistical Machine Translation, 2015.
Ondřej Bojar, Rajen Chatterjee, Christian Federmann, Yvette Graham, Barry Haddow,
Matthias Huck, Antonio Jimeno Yepes, Philipp Koehn, Varvara Logacheva, Christof Monz,
et al. Findings of the 2016 conference on machine translation. In Proceedings of the First
Conference on Machine Translation, 2016.
Samuel R. Bowman, Luke Vilnis, Oriol Vinyals, Andrew M. Dai, Rafal Jozefowicz, and Samy
Bengio. Generating sentences from a continuous space. arXiv preprint arXiv:1511.06349,
2015.
58
Exploring the Limits of Transfer Learning
Christian Buck, Kenneth Heafield, and Bas Van Ooyen. N-gram counts and language models
from the common crawl. In LREC, 2014.
Rich Caruana. Multitask learning. Machine learning, 28(1), 1997.
Daniel Cer, Mona Diab, Eneko Agirre, Inigo Lopez-Gazpio, and Lucia Specia. Semeval-2017
task 1: Semantic textual similarity-multilingual and cross-lingual focused evaluation.
arXiv preprint arXiv:1708.00055, 2017.
Jianpeng Cheng, Li Dong, and Mirella Lapata. Long short-term memory-networks for
machine reading. arXiv preprint arXiv:1601.06733, 2016.
Christopher Clark, Kenton Lee, Ming-Wei Chang, Tom Kwiatkowski, Michael Collins, and
Kristina Toutanova. BoolQ: Exploring the surprising difficulty of natural yes/no questions.
arXiv preprint arXiv:1905.10044, 2019.
Kevin Clark, Minh-Thang Luong, Quoc V Le, and Christopher D Manning. Electra:
Pre-training text encoders as discriminators rather than generators. arXiv preprint
arXiv:2003.10555, 2020.
Alexis Conneau and Douwe Kiela. SentEval: An evaluation toolkit for universal sentence
representations. arXiv preprint arXiv:1803.05449, 2018.
Alexis Conneau, Douwe Kiela, Holger Schwenk, Loic Barrault, and Antoine Bordes. Super-
vised learning of universal sentence representations from natural language inference data.
arXiv preprint arXiv:1705.02364, 2017.
Ido Dagan, Oren Glickman, and Bernardo Magnini. The PASCAL recognising textual
entailment challenge. In Machine Learning Challenges Workshop, 2005.
Andrew M. Dai and Quoc V. Le. Semi-supervised sequence learning. In Advances in neural
information processing systems, 2015.
Marie-Catherine De Marneff, Mandy Simons, and Judith Tonhauser. The CommitmentBank:
Investigating projection in naturally occurring discourse. In Sinn und Bedeutung 23, 2019.
Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: A
large-scale hierarchical image database. In 2009 IEEE conference on computer vision and
pattern recognition, 2009.
Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-
training of deep bidirectional transformers for language understanding. arXiv preprint
arXiv:1810.04805, 2018.
William B. Dolan and Chris Brockett. Automatically constructing a corpus of sentential para-
phrases. In Proceedings of the Third International Workshop on Paraphrasing (IWP2005),
2005.
Li Dong, Nan Yang, Wenhui Wang, Furu Wei, Xiaodong Liu, Yu Wang, Jianfeng Gao, Ming
Zhou, and Hsiao-Wuen Hon. Unified language model pre-training for natural language
understanding and generation. arXiv preprint arXiv:1905.03197, 2019.
59
Raffel, Shazeer, Roberts, Lee, Narang, Matena, Zhou, Li and Liu
Sergey Edunov, Myle Ott, Michael Auli, and David Grangier. Understanding back-translation
at scale. arXiv preprint arXiv:1808.09381, 2018.
Edouard Grave, Piotr Bojanowski, Prakhar Gupta, Armand Joulin, and Tomas Mikolov.
Learning word vectors for 157 languages. arXiv preprint arXiv:1802.06893, 2018.
Alex Graves. Generating sequences with recurrent neural networks. arXiv preprint
arXiv:1308.0850, 2013.
Ivan Habernal, Omnia Zayed, and Iryna Gurevych. C4Corpus: Multilingual web-size corpus
with free license. In Proceedings of the Tenth International Conference on Language
Resources and Evaluation (LREC’16), pages 914–922, 2016.
Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for
image recognition. In Proceedings of the IEEE conference on computer vision and pattern
recognition, 2016.
Kaiming He, Ross Girshick, and Piotr Dollár. Rethinking ImageNet pre-training. arXiv
preprint arXiv:1811.08883, 2018.
Pengcheng He, Xiaodong Liu, Weizhu Chen, and Jianfeng Gao. A hybrid neural network
model for commonsense reasoning. arXiv preprint arXiv:1907.11983, 2019.
Karl Moritz Hermann, Tomas Kocisky, Edward Grefenstette, Lasse Espeholt, Will Kay,
Mustafa Suleyman, and Phil Blunsom. Teaching machines to read and comprehend. In
Advances in neural information processing systems, 2015.
Joel Hestness, Sharan Narang, Newsha Ardalani, Gregory Diamos, Heewoo Jun, Hassan
Kianinejad, Md. Mostofa Ali Patwary, Yang Yang, and Yanqi Zhou. Deep learning scaling
is predictable, empirically. arXiv preprint arXiv:1712.00409, 2017.
Felix Hill, Kyunghyun Cho, and Anna Korhonen. Learning distributed representations of
sentences from unlabelled data. arXiv preprint arXiv:1602.03483, 2016.
Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network.
arXiv preprint arXiv:1503.02531, 2015.
Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe,
Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer
learning for NLP. arXiv preprint arXiv:1902.00751, 2019.
Jeremy Howard and Sebastian Ruder. Universal language model fine-tuning for text classifi-
cation. arXiv preprint arXiv:1801.06146, 2018.
Cheng-Zhi Anna Huang, Ashish Vaswani, Jakob Uszkoreit, Ian Simon, Curtis Hawthorne,
Noam Shazeer, Andrew M. Dai, Matthew D. Hoffman, Monica Dinculescu, and Dou-
glas Eck. Music transformer: Generating music with long-term structure. In Seventh
International Conference on Learning Representations, 2018a.
60
Exploring the Limits of Transfer Learning
Yanping Huang, Yonglong Cheng, Dehao Chen, HyoukJoong Lee, Jiquan Ngiam, Quoc V
Le, and Zhifeng Chen. GPipe: Efficient training of giant neural networks using pipeline
parallelism. arXiv preprint arXiv:1811.06965, 2018b.
Minyoung Huh, Pulkit Agrawal, and Alexei A. Efros. What makes ImageNet good for
transfer learning? arXiv preprint arXiv:1608.08614, 2016.
Shankar Iyer, Nikhil Dandekar, and Kornel Csernai. First Quora dataset release: Question
pairs. https://data.quora.com/First-Quora-Dataset-Release-Question-Pairs,
2017.
Yangqing Jia, Evan Shelhamer, Jeff Donahue, Sergey Karayev, Jonathan Long, Ross Girshick,
Sergio Guadarrama, and Trevor Darrell. Caffe: Convolutional architecture for fast feature
embedding. In Proceedings of the 22nd ACM international conference on Multimedia,
2014.
Xiaoqi Jiao, Yichun Yin, Lifeng Shang, Xin Jiang, Xiao Chen, Linlin Li, Fang Wang, and
Qun Liu. TinyBERT: Distilling BERT for natural language understanding. arXiv preprint
arXiv:1909.10351, 2019.
Mandar Joshi, Eunsol Choi, Daniel S. Weld, and Luke Zettlemoyer. TriviaQA: A large
scale distantly supervised challenge dataset for reading comprehension. arXiv preprint
arXiv:1705.03551, 2017.
Mandar Joshi, Danqi Chen, Yinhan Liu, Daniel S. Weld, Luke Zettlemoyer, and Omer Levy.
SpanBERT: Improving pre-training by representing and predicting spans. arXiv preprint
arXiv:1907.10529, 2019.
Rafal Jozefowicz, Oriol Vinyals, Mike Schuster, Noam Shazeer, and Yonghui Wu. Exploring
the limits of language modeling. arXiv preprint arXiv:1602.02410, 2016.
Nal Kalchbrenner, Edward Grefenstette, and Phil Blunsom. A convolutional neural network
for modelling sentences. In Proceedings of the 52nd Annual Meeting of the Association for
Computational Linguistics, 2014.
Nitish Shirish Keskar, Bryan McCann, Lav R. Varshney, Caiming Xiong, and Richard
Socher. CTRL: A conditional transformer language model for controllable generation.
arXiv preprint arXiv:1909.05858, 2019a.
Nitish Shirish Keskar, Bryan McCann, Caiming Xiong, and Richard Socher. Unifying question
answering and text classification via span extraction. arXiv preprint arXiv:1904.09286,
2019b.
Daniel Khashabi, Snigdha Chaturvedi, Michael Roth, Shyam Upadhyay, and Dan Roth.
Looking beyond the surface: A challenge set for reading comprehension over multiple
sentences. In Proceedings of North American Chapter of the Association for Computational
Linguistics (NAACL), 2018.
61
Raffel, Shazeer, Roberts, Lee, Narang, Matena, Zhou, Li and Liu
Ryan Kiros, Yukun Zhu, Ruslan R. Salakhutdinov, Richard Zemel, Raquel Urtasun, Antonio
Torralba, and Sanja Fidler. Skip-thought vectors. In Advances in neural information
processing systems, 2015.
Vid Kocijan, Ana-Maria Cretu, Oana-Maria Camburu, Yordan Yordanov, and Thomas
Lukasiewicz. A surprisingly robust trick for Winograd schema challenge. arXiv preprint
arXiv:1905.06290, 2019.
Jakub Konečnỳ, Brendan McMahan, and Daniel Ramage. Federated optimization: Dis-
tributed optimization beyond the datacenter. arXiv preprint arXiv:1511.03575, 2015.
Jakub Konečnỳ, H. Brendan McMahan, Felix X. Yu, Peter Richtárik, Ananda Theertha
Suresh, and Dave Bacon. Federated learning: Strategies for improving communication
efficiency. arXiv preprint arXiv:1610.05492, 2016.
Simon Kornblith, Jonathon Shlens, and Quoc V. Le. Do better ImageNet models transfer
better? arXiv preprint arXiv:1805.08974, 2018.
Alex Krizhevsky. One weird trick for parallelizing convolutional neural networks. arXiv
preprint arXiv:1404.5997, 2014.
Taku Kudo. Subword regularization: Improving neural network translation models with
multiple subword candidates. arXiv preprint arXiv:1804.10959, 2018.
Taku Kudo and John Richardson. SentencePiece: A simple and language independent sub-
word tokenizer and detokenizer for neural text processing. arXiv preprint arXiv:1808.06226,
2018.
Guillaume Lample and Alexis Conneau. Cross-lingual language model pretraining. arXiv
preprint arXiv:1901.07291, 2019.
Zhenzhong Lan, Mingda Chen, Sebastian Goodman, Kevin Gimpel, Piyush Sharma, and
Radu Soricut. ALBERT: A lite BERT for self-supervised learning of language representa-
tions. arXiv preprint arXiv:1909.11942, 2019.
Hector Levesque, Ernest Davis, and Leora Morgenstern. The Winograd schema challenge.
In Thirteenth International Conference on the Principles of Knowledge Representation
and Reasoning, 2012.
Qi Li. Literature survey: domain adaptation algorithms for natural language processing.
2012.
Chin-Yew Lin. ROUGE: A package for automatic evaluation of summaries. In Text
summarization branches out, 2004.
Peter J. Liu, Mohammad Saleh, Etienne Pot, Ben Goodrich, Ryan Sepassi, Lukasz Kaiser,
and Noam Shazeer. Generating Wikipedia by summarizing long sequences. arXiv preprint
arXiv:1801.10198, 2018.
Peter J. Liu, Yu-An Chung, and Jie Ren. SummAE: Zero-shot abstractive text summarization
using length-agnostic auto-encoders. arXiv preprint arXiv:1910.00998, 2019a.
62
Exploring the Limits of Transfer Learning
Xiaodong Liu, Jianfeng Gao, Xiaodong He, Li Deng, Kevin Duh, and Ye-Yi Wang. Rep-
resentation learning using multi-task deep neural networks for semantic classification
and information retrieval. In Proceedings of the 2015 Conference of the North American
Chapter of the Association for Computational Linguistics: Human Language Technologies,
2015.
Xiaodong Liu, Pengcheng He, Weizhu Chen, and Jianfeng Gao. Multi-task deep neural
networks for natural language understanding. arXiv preprint arXiv:1901.11504, 2019b.
Yang Liu. Fine-tune BERT for extractive summarization. arXiv preprint arXiv:1903.10318,
2019.
Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy,
Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. RoBERTa: A robustly optimized
BERT pretraining approach. arXiv preprint arXiv:1907.11692, 2019c.
Lajanugen Logeswaran and Honglak Lee. An efficient framework for learning sentence
representations. arXiv preprint arXiv:1803.02893, 2018.
Dhruv Mahajan, Ross Girshick, Vignesh Ramanathan, Kaiming He, Manohar Paluri, Yixuan
Li, Ashwin Bharambe, and Laurens van der Maaten. Exploring the limits of weakly
supervised pretraining. In Proceedings of the European Conference on Computer Vision
(ECCV), 2018.
Bryan McCann, Nitish Shirish Keskar, Caiming Xiong, and Richard Socher. The nat-
ural language decathlon: Multitask learning as question answering. arXiv preprint
arXiv:1806.08730, 2018.
Tomas Mikolov, Kai Chen, Greg Corrado, and Jeffrey Dean. Efficient estimation of word
representations in vector space. arXiv preprint arXiv:1301.3781, 2013a.
Tomas Mikolov, Ilya Sutskever, Kai Chen, Greg S. Corrado, and Jeff Dean. Distributed
representations of words and phrases and their compositionality. In Advances in neural
information processing systems, 2013b.
Ramesh Nallapati, Bowen Zhou, Cicero Nogueira dos santos, Caglar Gulcehre, and Bing
Xiang. Abstractive text summarization using sequence-to-sequence RNNs and beyond.
arXiv preprint arXiv:1602.06023, 2016.
Maxime Oquab, Leon Bottou, Ivan Laptev, and Josef Sivic. Learning and transferring
mid-level image representations using convolutional neural networks. In Proceedings of
the IEEE conference on computer vision and pattern recognition, 2014.
Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. BLEU: a method for
automatic evaluation of machine translation. In Proceedings of the 40th annual meeting on
association for computational linguistics. Association for Computational Linguistics, 2002.
Romain Paulus, Caiming Xiong, and Richard Socher. A deep reinforced model for abstractive
summarization. arXiv preprint arXiv:1705.04304, 2017.
63
Raffel, Shazeer, Roberts, Lee, Narang, Matena, Zhou, Li and Liu
Jeffrey Pennington, Richard Socher, and Christopher Manning. GloVe: Global vectors
for word representation. In Proceedings of the 2014 conference on empirical methods in
natural language processing (EMNLP), 2014.
Matthew Peters, Sebastian Ruder, and Noah A. Smith. To tune or not to tune? adapting
pretrained representations to diverse tasks. arXiv preprint arXiv:1903.05987, 2019.
Matthew E. Peters, Mark Neumann, Mohit Iyyer, Matt Gardner, Christopher Clark, Kenton
Lee, and Luke Zettlemoyer. Deep contextualized word representations. arXiv preprint
arXiv:1802.05365, 2018.
Jason Phang, Thibault Févry, and Samuel R. Bowman. Sentence encoders on STILTs: Sup-
plementary training on intermediate labeled-data tasks. arXiv preprint arXiv:1811.01088,
2018.
Mohammad Taher Pilehvar and Jose Camacho-Collados. WIC: 10,000 example pairs for
evaluating context-sensitive representations. arXiv preprint arXiv:1808.09121, 2018.
Matt Post. A call for clarity in reporting BLEU scores. arXiv preprint arXiv:1804.08771,
2018.
Alec Radford, Karthik Narasimhan, Tim Salimans, and Ilya Sutskever. Improving language
understanding by generative pre-training, 2018.
Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever.
Language models are unsupervised multitask learners, 2019.
Altaf Rahman and Vincent Ng. Resolving complex cases of definite pronouns: the Winograd
schema challenge. In Proceedings of the 2012 Joint Conference on Empirical Methods in
Natural Language Processing and Computational Natural Language Learning. Association
for Computational Linguistics, 2012.
Pranav Rajpurkar, Jian Zhang, Konstantin Lopyrev, and Percy Liang. Squad: 100,000+
questions for machine comprehension of text. arXiv preprint arXiv:1606.05250, 2016.
Prajit Ramachandran, Peter J. Liu, and Quoc V. Le. Unsupervised pretraining for sequence
to sequence learning. arXiv preprint arXiv:1611.02683, 2016.
Alex Ratner, Braden Hancock, Jared Dunnmon, Roger Goldman, and Christopher Ré.
Snorkel MeTaL: Weak supervision for multi-task learning. In Proceedings of the Second
Workshop on Data Management for End-To-End Machine Learning, 2018.
Melissa Roemmele, Cosmin Adrian Bejan, and Andrew S Gordon. Choice of plausible
alternatives: An evaluation of commonsense causal reasoning. In 2011 AAAI Spring
Symposium Series, 2011.
Sebastian Ruder. An overview of multi-task learning in deep neural networks. arXiv preprint
arXiv:1706.05098, 2017.
Sebastian Ruder. Neural transfer learning for natural language processing. PhD thesis, NUI
Galway, 2019.
64
Exploring the Limits of Transfer Learning
Sebastian Ruder, Matthew E. Peters, Swabha Swayamdipta, and Thomas Wolf. Transfer
learning in natural language processing. In Proceedings of the 2019 Conference of the
North American Chapter of the Association for Computational Linguistics: Tutorials,
pages 15–18, 2019.
Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng
Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. ImageNet large scale
visual recognition challenge. International journal of computer vision, 2015.
Victor Sanh, Lysandre Debut, Julien Chaumond, and Thomas Wolf. DistilBERT, a distilled
version of BERT: smaller, faster, cheaper and lighter. arXiv preprint arXiv:1910.01108,
2019.
Abigail See, Peter J. Liu, and Christopher D. Manning. Get to the point: Summarization
with pointer-generator networks. arXiv preprint arXiv:1704.04368, 2017.
Rico Sennrich, Barry Haddow, and Alexandra Birch. Neural machine translation of rare
words with subword units. arXiv preprint arXiv:1508.07909, 2015.
Christopher J Shallue, Jaehoon Lee, Joe Antognini, Jascha Sohl-Dickstein, Roy Frostig, and
George E. Dahl. Measuring the effects of data parallelism on neural network training.
arXiv preprint arXiv:1811.03600, 2018.
Peter Shaw, Jakob Uszkoreit, and Ashish Vaswani. Self-attention with relative position
representations. arXiv preprint arXiv:1803.02155, 2018.
Noam Shazeer and Mitchell Stern. Adafactor: Adaptive learning rates with sublinear memory
cost. arXiv preprint arXiv:1804.04235, 2018.
Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc Le, Geoffrey Hinton,
and Jeff Dean. Outrageously large neural networks: The sparsely-gated mixture-of-experts
layer. arXiv preprint arXiv:1701.06538, 2017.
Noam Shazeer, Youlong Cheng, Niki Parmar, Dustin Tran, Ashish Vaswani, Penporn
Koanantakool, Peter Hawkins, HyoukJoong Lee, Mingsheng Hong, Cliff Young, Ryan
Sepassi, and Blake Hechtman. Mesh-tensorflow: Deep learning for supercomputers. In
Advances in Neural Information Processing Systems, 2018.
Jason R. Smith, Herve Saint-Amand, Magdalena Plamada, Philipp Koehn, Chris Callison-
Burch, and Adam Lopez. Dirt cheap web-scale parallel text from the common crawl. In
Proceedings of the 51st Annual Meeting of the Association for Computational Linguistics,
2013.
Richard Socher, Alex Perelygin, Jean Wu, Jason Chuang, Christopher D. Manning, Andrew
Ng, and Christopher Potts. Recursive deep models for semantic compositionality over a
sentiment treebank. In Proceedings of the 2013 conference on empirical methods in natural
language processing, 2013.
Kaitao Song, Xu Tan, Tao Qin, Jianfeng Lu, and Tie-Yan Liu. MASS: Masked sequence to
sequence pre-training for language generation. arXiv preprint arXiv:1905.02450, 2019.
65
Raffel, Shazeer, Roberts, Lee, Narang, Matena, Zhou, Li and Liu
Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdi-
nov. Dropout: a simple way to prevent neural networks from overfitting. The Journal of
Machine Learning Research, 2014.
Sandeep Subramanian, Adam Trischler, Yoshua Bengio, and Christopher J. Pal. Learning
general purpose distributed sentence representations via large scale multi-task learning.
arXiv preprint arXiv:1804.00079, 2018.
Ilya Sutskever, Oriol Vinyals, and Quoc V. Le. Sequence to sequence learning with neural
networks. In Advances in neural information processing systems, 2014.
Richard S. Sutton. The bitter lesson. http://www.incompleteideas.net/IncIdeas/
BitterLesson.html, 2019.
Wilson L. Taylor. “Cloze procedure”: A new tool for measuring readability. Journalism
Bulletin, 1953.
Trieu H. Trinh and Quoc V. Le. A simple method for commonsense reasoning. arXiv preprint
arXiv:1806.02847, 2018.
Adam Trischler, Tong Wang, Xingdi Yuan, Justin Harris, Alessandro Sordoni, Philip
Bachman, and Kaheer Suleman. NewsQA: A machine comprehension dataset. arXiv
preprint arXiv:1611.09830, 2016.
Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez,
Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in neural
information processing systems, 2017.
Elena Voita, Rico Sennrich, and Ivan Titov. The bottom-up evolution of representations
in the transformer: A study with machine translation and language modeling objectives.
arXiv preprint arXiv:1909.01380, 2019.
Alex Wang, Amapreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel R. Bowman.
GLUE: A multi-task benchmark and analysis platform for natural language understanding.
arXiv preprint arXiv:1804.07461, 2018.
Alex Wang, Jan Hula, Patrick Xia, Raghavendra Pappagari, R. Thomas McCoy, Roma
Patel, Najoung Kim, Ian Tenney, Yinghui Huang, Katherin Yu, et al. Can you tell me
how to get past Sesame Street? Sentence-level pretraining beyond language modeling. In
Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics,
2019a.
Alex Wang, Yada Pruksachatkun, Nikita Nangia, Amanpreet Singh, Julian Michael, Felix
Hill, Omer Levy, and Samuel R. Bowman. SuperGLUE: A stickier benchmark for general-
purpose language understanding systems. arXiv preprint arXiv:1905.00537, 2019b.
Wei Wang, Bin Bi, Ming Yan, Chen Wu, Zuyi Bao, Liwei Peng, and Luo Si. StructBERT:
Incorporating language structures into pre-training for deep language understanding.
arXiv preprint arXiv:1908.04577, 2019c.
66
Exploring the Limits of Transfer Learning
Alex Warstadt, Amanpreet Singh, and Samuel R. Bowman. Neural network acceptability
judgments. arXiv preprint arXiv:1805.12471, 2018.
Adina Williams, Nikita Nangia, and Samuel R. Bowman. A broad-coverage challenge corpus
for sentence understanding through inference. arXiv preprint arXiv:1704.05426, 2017.
Ronald J. Williams and David Zipser. A learning algorithm for continually running fully
recurrent neural networks. Neural computation, 1989.
Yonghui Wu, Mike Schuster, Zhifeng Chen, Quoc V. Le, Mohammad Norouzi, Wolfgang
Macherey, Maxim Krikun, Yuan Cao, Qin Gao, Klaus Macherey, et al. Google’s neural
machine translation system: Bridging the gap between human and machine translation.
arXiv preprint arXiv:1609.08144, 2016.
Zhilin Yang, Zihang Dai, Yiming Yang, Jaime Carbonell, Ruslan Salakhutdinov, and Quoc V.
Le. XLNet: Generalized autoregressive pretraining for language understanding. arXiv
preprint arXiv:1906.08237, 2019.
Jason Yosinski, Jeff Clune, Yoshua Bengio, and Hod Lipson. How transferable are features
in deep neural networks? In Advances in neural information processing systems, 2014.
Adams Wei Yu, David Dohan, Minh-Thang Luong, Rui Zhao, Kai Chen, Mohammad
Norouzi, and Quoc V. Le. QAnet: Combining local convolution with global self-attention
for reading comprehension. arXiv preprint arXiv:1804.09541, 2018.
Rowan Zellers, Ari Holtzman, Hannah Rashkin, Yonatan Bisk, Ali Farhadi, Franziska Roes-
ner, and Yejin Choi. Defending against neural fake news. arXiv preprint arXiv:1905.12616,
2019.
Sheng Zhang, Xiaodong Liu, Jingjing Liu, Jianfeng Gao, Kevin Duh, and Benjamin
Van Durme. ReCoRD: Bridging the gap between human and machine commonsense
reading comprehension. arXiv preprint arXiv:1810.12885, 2018.
Chen Zhu, Yu Cheng, Zhe Gan, Siqi Sun, Thomas Goldstein, and Jingjing Liu. Freelb: En-
hanced adversarial training for language understanding. arXiv preprint arXiv:1909.11764,
2019.
Yukun Zhu, Ryan Kiros, Rich Zemel, Ruslan Salakhutdinov, Raquel Urtasun, Antonio
Torralba, and Sanja Fidler. Aligning books and movies: Towards story-like visual expla-
nations by watching movies and reading books. In Proceedings of the IEEE international
conference on computer vision, 2015.
67
来源材料