***** New March 11th, 2020: Smaller BERT Models *****

This is a release of 24 smaller BERT models (English only, uncased, trained with WordPiece masking) referenced in Well-Read Students Learn Better: On the Importance of Pre-training Compact Models.

We have shown that the standard BERT recipe (including model architecture and training objective) is effective on a wide range of model sizes, beyond BERT-Base and BERT-Large. The smaller BERT models are intended for environments with restricted computational resources. They can be fine-tuned in the same manner as the original BERT models. However, they are most effective in the context of knowledge distillation, where the fine-tuning labels are produced by a larger and more accurate teacher.

Our goal is to enable research in institutions with fewer computational resources and encourage the community to seek directions of innovation alternative to increasing model capacity.

You can download all 24 from here, or individually from the table below:

Note that the BERT-Base model in this release is included for completeness only; it was re-trained under the same regime as the original model.

Here are the corresponding GLUE scores on the test set:

For each task, we selected the best fine-tuning hyperparameters from the lists below, and trained for 4 epochs:

• batch sizes: 8, 16, 32, 64, 128
• learning rates: 3e-4, 1e-4, 5e-5, 3e-5

If you use these models, please cite the following paper:

@article{turc2019,
  title={Well-Read Students Learn Better: On the Importance of Pre-training Compact Models},
  author={Turc, Iulia and Chang, Ming-Wei and Lee, Kenton and Toutanova, Kristina},
  journal={arXiv preprint arXiv:1908.08962v2 },
  year={2019}
}

***** New May 31st, 2019: Whole Word Masking Models *****

This is a release of several new models which were the result of an improvement the pre-processing code.

In the original pre-processing code, we randomly select WordPiece tokens to mask. For example:

Input Text: the man jumped up , put his basket on phil ##am ##mon ' s head Original Masked Input: [MASK] man [MASK] up , put his [MASK] on phil [MASK] ##mon ' s head

The new technique is called Whole Word Masking. In this case, we always mask all of the the tokens corresponding to a word at once. The overall masking rate remains the same.

Whole Word Masked Input: the man [MASK] up , put his basket on [MASK] [MASK] [MASK] ' s head

The training is identical -- we still predict each masked WordPiece token independently. The improvement comes from the fact that the original prediction task was too 'easy' for words that had been split into multiple WordPieces.

This can be enabled during data generation by passing the flag --do_whole_word_mask=True to create_pretraining_data.py.

Pre-trained models with Whole Word Masking are linked below. The data and training were otherwise identical, and the models have identical structure and vocab to the original models. We only include BERT-Large models. When using these models, please make it clear in the paper that you are using the Whole Word Masking variant of BERT-Large.

• BERT-Large, Uncased (Whole Word Masking): 24-layer, 1024-hidden, 16-heads, 340M parameters

• BERT-Large, Cased (Whole Word Masking): 24-layer, 1024-hidden, 16-heads, 340M parameters

***** New February 7th, 2019: TfHub Module *****

BERT has been uploaded to TensorFlow Hub. See run_classifier_with_tfhub.py for an example of how to use the TF Hub module, or run an example in the browser on Colab.

***** New November 23rd, 2018: Un-normalized multilingual model + Thai + Mongolian *****

We uploaded a new multilingual model which does not perform any normalization on the input (no lower casing, accent stripping, or Unicode normalization), and additionally inclues Thai and Mongolian.

It is recommended to use this version for developing multilingual models, especially on languages with non-Latin alphabets.

This does not require any code changes, and can be downloaded here:

• BERT-Base, Multilingual Cased: 104 languages, 12-layer, 768-hidden, 12-heads, 110M parameters

***** New November 15th, 2018: SOTA SQuAD 2.0 System *****

We released code changes to reproduce our 83% F1 SQuAD 2.0 system, which is currently 1st place on the leaderboard by 3%. See the SQuAD 2.0 section of the README for details.

***** New November 5th, 2018: Third-party PyTorch and Chainer versions of BERT available *****

NLP researchers from HuggingFace made a PyTorch version of BERT available which is compatible with our pre-trained checkpoints and is able to reproduce our results. Sosuke Kobayashi also made a Chainer version of BERT available (Thanks!) We were not involved in the creation or maintenance of the PyTorch implementation so please direct any questions towards the authors of that repository.

***** New November 3rd, 2018: Multilingual and Chinese models available *****

We have made two new BERT models available:

• BERT-Base, Multilingual (Not recommended, use Multilingual Cased instead): 102 languages, 12-layer, 768-hidden, 12-heads, 110M parameters
• BERT-Base, Chinese: Chinese Simplified and Traditional, 12-layer, 768-hidden, 12-heads, 110M parameters

We use character-based tokenization for Chinese, and WordPiece tokenization for all other languages. Both models should work out-of-the-box without any code changes. We did update the implementation of BasicTokenizer in tokenization.py to support Chinese character tokenization, so please update if you forked it. However, we did not change the tokenization API.

For more, see the Multilingual README.

***** End new information *****

BERT, or Bidirectional Encoder Representations from Transformers, is a new method of pre-training language representations which obtains state-of-the-art results on a wide array of Natural Language Processing (NLP) tasks.

Our academic paper which describes BERT in detail and provides full results on a number of tasks can be found here: https://arxiv.org/abs/1810.04805.

To give a few numbers, here are the results on the SQuAD v1.1 question answering task:

And several natural language inference tasks:

Plus many other tasks.

Moreover, these results were all obtained with almost no task-specific neural network architecture design.

If you already know what BERT is and you just want to get started, you can download the pre-trained models and run a state-of-the-art fine-tuning in only a few minutes.

BERT is a method of pre-training language representations, meaning that we train a general-purpose "language understanding" model on a large text corpus (like Wikipedia), and then use that model for downstream NLP tasks that we care about (like question answering). BERT outperforms previous methods because it is the first unsupervised, deeply bidirectional system for pre-training NLP.

Unsupervised means that BERT was trained using only a plain text corpus, which is important because an enormous amount of plain text data is publicly available on the web in many languages.

Pre-trained representations can also either be context-free or contextual, and contextual representations can further be unidirectional or bidirectional. Context-free models such as word2vec or GloVe generate a single "word embedding" representation for each word in the vocabulary, so bank would have the same representation in bank deposit and river bank. Contextual models instead generate a representation of each word that is based on the other words in the sentence.

BERT was built upon recent work in pre-training contextual representations — including Semi-supervised Sequence Learning, Generative Pre-Training, ELMo, and ULMFit — but crucially these models are all unidirectional or shallowly bidirectional. This means that each word is only contextualized using the words to its left (or right). For example, in the sentence I made a bank deposit the unidirectional representation of bank is only based on I made a but not deposit. Some previous work does combine the representations from separate left-context and right-context models, but only in a "shallow" manner. BERT represents "bank" using both its left and right context — I made a ... deposit — starting from the very bottom of a deep neural network, so it is deeply bidirectional.

BERT uses a simple approach for this: We mask out 15% of the words in the input, run the entire sequence through a deep bidirectional Transformer encoder, and then predict only the masked words. For example:

Input: the man went to the [MASK1] . he bought a [MASK2] of milk.
Labels: [MASK1] = store; [MASK2] = gallon

In order to learn relationships between sentences, we also train on a simple task which can be generated from any monolingual corpus: Given two sentences A and B, is B the actual next sentence that comes after A, or just a random sentence from the corpus?

Sentence A: the man went to the store .
Sentence B: he bought a gallon of milk .
Label: IsNextSentence

Sentence A: the man went to the store .
Sentence B: penguins are flightless .
Label: NotNextSentence

We then train a large model (12-layer to 24-layer Transformer) on a large corpus (Wikipedia + BookCorpus) for a long time (1M update steps), and that's BERT.

Using BERT has two stages: Pre-training and fine-tuning.

Pre-training is fairly expensive (four days on 4 to 16 Cloud TPUs), but is a one-time procedure for each language (current models are English-only, but multilingual models will be released in the near future). We are releasing a number of pre-trained models from the paper which were pre-trained at Google. Most NLP researchers will never need to pre-train their own model from scratch.

Fine-tuning is inexpensive. All of the results in the paper can be replicated in at most 1 hour on a single Cloud TPU, or a few hours on a GPU, starting from the exact same pre-trained model. SQuAD, for example, can be trained in around 30 minutes on a single Cloud TPU to achieve a Dev F1 score of 91.0%, which is the single system state-of-the-art.

The other important aspect of BERT is that it can be adapted to many types of NLP tasks very easily. In the paper, we demonstrate state-of-the-art results on sentence-level (e.g., SST-2), sentence-pair-level (e.g., MultiNLI), word-level (e.g., NER), and span-level (e.g., SQuAD) tasks with almost no task-specific modifications.

We are releasing the following:

• TensorFlow code for the BERT model architecture (which is mostly a standard Transformer architecture).
• Pre-trained checkpoints for both the lowercase and cased version of BERT-Base and BERT-Large from the paper.
• TensorFlow code for push-button replication of the most important fine-tuning experiments from the paper, including SQuAD, MultiNLI, and MRPC.

All of the code in this repository works out-of-the-box with CPU, GPU, and Cloud TPU.

We are releasing the BERT-Base and BERT-Large models from the paper. Uncased means that the text has been lowercased before WordPiece tokenization, e.g., John Smith becomes john smith. The Uncased model also strips out any accent markers. Cased means that the true case and accent markers are preserved. Typically, the Uncased model is better unless you know that case information is important for your task (e.g., Named Entity Recognition or Part-of-Speech tagging).

These models are all released under the same license as the source code (Apache 2.0).

For information about the Multilingual and Chinese model, see the Multilingual README.

When using a cased model, make sure to pass --do_lower=False to the training scripts. (Or pass do_lower_case=False directly to FullTokenizer if you're using your own script.)

The links to the models are here (right-click, 'Save link as...' on the name):

• BERT-Large, Uncased (Whole Word Masking): 24-layer, 1024-hidden, 16-heads, 340M parameters
• BERT-Large, Cased (Whole Word Masking): 24-layer, 1024-hidden, 16-heads, 340M parameters
• BERT-Base, Uncased: 12-layer, 768-hidden, 12-heads, 110M parameters
• BERT-Large, Uncased: 24-layer, 1024-hidden, 16-heads, 340M parameters
• BERT-Base, Cased: 12-layer, 768-hidden, 12-heads , 110M parameters
• BERT-Large, Cased: 24-layer, 1024-hidden, 16-heads, 340M parameters
• BERT-Base, Multilingual Cased (New, recommended): 104 languages, 12-layer, 768-hidden, 12-heads, 110M parameters
• BERT-Base, Multilingual Uncased (Orig, not recommended) (Not recommended, use Multilingual Cased instead): 102 languages, 12-layer, 768-hidden, 12-heads, 110M parameters
• BERT-Base, Chinese: Chinese Simplified and Traditional, 12-layer, 768-hidden, 12-heads, 110M parameters

Each .zip file contains three items:

• A TensorFlow checkpoint (bert_model.ckpt) containing the pre-trained weights (which is actually 3 files).
• A vocab file (vocab.txt) to map WordPiece to word id.
• A config file (bert_config.json) which specifies the hyperparameters of the model.

Important: All results on the paper were fine-tuned on a single Cloud TPU, which has 64GB of RAM. It is currently not possible to re-produce most of the BERT-Large results on the paper using a GPU with 12GB - 16GB of RAM, because the maximum batch size that can fit in memory is too small. We are working on adding code to this repository which allows for much larger effective batch size on the GPU. See the section on out-of-memory issues for more details.

This code was tested with TensorFlow 1.11.0. It was tested with Python2 and Python3 (but more thoroughly with Python2, since this is what's used internally in Google).

The fine-tuning examples which use BERT-Base should be able to run on a GPU that has at least 12GB of RAM using the hyperparameters given.

Most of the examples below assumes that you will be running training/evaluation on your local machine, using a GPU like a Titan X or GTX 1080.

However, if you have access to a Cloud TPU that you want to train on, just add the following flags to run_classifier.py or run_squad.py:

  --use_tpu=True \
  --tpu_name=$TPU_NAME

Please see the Google Cloud TPU tutorial for how to use Cloud TPUs. Alternatively, you can use the Google Colab notebook "BERT FineTuning with Cloud TPUs".

On Cloud TPUs, the pretrained model and the output directory will need to be on Google Cloud Storage. For example, if you have a bucket named some_bucket, you might use the following flags instead:

  --output_dir=gs://some_bucket/my_output_dir/

The unzipped pre-trained model files can also be found in the Google Cloud Storage folder gs://bert_models/2018_10_18. For example:

export BERT_BASE_DIR=gs://bert_models/2018_10_18/uncased_L-12_H-768_A-12

Before running this example you must download the GLUE data by running this script and unpack it to some directory $GLUE_DIR. Next, download the BERT-Base checkpoint and unzip it to some directory $BERT_BASE_DIR.

This example code fine-tunes BERT-Base on the Microsoft Research Paraphrase Corpus (MRPC) corpus, which only contains 3,600 examples and can fine-tune in a few minutes on most GPUs.

export BERT_BASE_DIR=/path/to/bert/uncased_L-12_H-768_A-12
export GLUE_DIR=/path/to/glue

python run_classifier.py \
  --task_name=MRPC \
  --do_train=true \
  --do_eval=true \
  --data_dir=$GLUE_DIR/MRPC \
  --vocab_file=$BERT_BASE_DIR/vocab.txt \
  --bert_config_file=$BERT_BASE_DIR/bert_config.json \
  --init_checkpoint=$BERT_BASE_DIR/bert_model.ckpt \
  --max_seq_length=128 \
  --train_batch_size=32 \
  --learning_rate=2e-5 \
  --num_train_epochs=3.0 \
  --output_dir=/tmp/mrpc_output/

You should see output like this:

***** Eval results *****
  eval_accuracy = 0.845588
  eval_loss = 0.505248
  global_step = 343
  loss = 0.505248

This means that the Dev set accuracy was 84.55%. Small sets like MRPC have a high variance in the Dev set accuracy, even when starting from the same pre-training checkpoint. If you re-run multiple times (making sure to point to different output_dir), you should see results between 84% and 88%.

A few other pre-trained models are implemented off-the-shelf in run_classifier.py, so it should be straightforward to follow those examples to use BERT for any single-sentence or sentence-pair classification task.

Note: You might see a message Running train on CPU. This really just means that it's running on something other than a Cloud TPU, which includes a GPU.

Once you have trained your classifier you can use it in inference mode by using the --do_predict=true command. You need to have a file named test.tsv in the input folder. Output will be created in file called test_results.tsv in the output folder. Each line will contain output for each sample, columns are the class probabilities.

export BERT_BASE_DIR=/path/to/bert/uncased_L-12_H-768_A-12
export GLUE_DIR=/path/to/glue
export TRAINED_CLASSIFIER=/path/to/fine/tuned/classifier

python run_classifier.py \
  --task_name=MRPC \
  --do_predict=true \
  --data_dir=$GLUE_DIR/MRPC \
  --vocab_file=$BERT_BASE_DIR/vocab.txt \
  --bert_config_file=$BERT_BASE_DIR/bert_config.json \
  --init_checkpoint=$TRAINED_CLASSIFIER \
  --max_seq_length=128 \
  --output_dir=/tmp/mrpc_output/

The Stanford Question Answering Dataset (SQuAD) is a popular question answering benchmark dataset. BERT (at the time of the release) obtains state-of-the-art results on SQuAD with almost no task-specific network architecture modifications or data augmentation. However, it does require semi-complex data pre-processing and post-processing to deal with (a) the variable-length nature of SQuAD context paragraphs, and (b) the character-level answer annotations which are used for SQuAD training. This processing is implemented and documented in run_squad.py.

To run on SQuAD, you will first need to download the dataset. The SQuAD website does not seem to link to the v1.1 datasets any longer, but the necessary files can be found here:

• train-v1.1.json
• dev-v1.1.json
• evaluate-v1.1.py

Download these to some directory $SQUAD_DIR.

The state-of-the-art SQuAD results from the paper currently cannot be reproduced on a 12GB-16GB GPU due to memory constraints (in fact, even batch size 1 does not seem to fit on a 12GB GPU using BERT-Large). However, a reasonably strong BERT-Base model can be trained on the GPU with these hyperparameters:

python run_squad.py \
  --vocab_file=$BERT_BASE_DIR/vocab.txt \
  --bert_config_file=$BERT_BASE_DIR/bert_config.json \
  --init_checkpoint=$BERT_BASE_DIR/bert_model.ckpt \
  --do_train=True \
  --train_file=$SQUAD_DIR/train-v1.1.json \
  --do_predict=True \
  --predict_file=$SQUAD_DIR/dev-v1.1.json \
  --train_batch_size=12 \
  --learning_rate=3e-5 \
  --num_train_epochs=2.0 \
  --max_seq_length=384 \
  --doc_stride=128 \
  --output_dir=/tmp/squad_base/

The dev set predictions will be saved into a file called predictions.json in the output_dir:

python $SQUAD_DIR/evaluate-v1.1.py $SQUAD_DIR/dev-v1.1.json ./squad/predictions.json

Which should produce an output like this:

{"f1": 88.41249612335034, "exact_match": 81.2488174077578}

You should see a result similar to the 88.5% reported in the paper for BERT-Base.

If you have access to a Cloud TPU, you can train with BERT-Large. Here is a set of hyperparameters (slightly different than the paper) which consistently obtain around 90.5%-91.0% F1 single-system trained only on SQuAD:

python run_squad.py \
  --vocab_file=$BERT_LARGE_DIR/vocab.txt \
  --bert_config_file=$BERT_LARGE_DIR/bert_config.json \
  --init_checkpoint=$BERT_LARGE_DIR/bert_model.ckpt \
  --do_train=True \
  --train_file=$SQUAD_DIR/train-v1.1.json \
  --do_predict=True \
  --predict_file=$SQUAD_DIR/dev-v1.1.json \
  --train_batch_size=24 \
  --learning_rate=3e-5 \
  --num_train_epochs=2.0 \
  --max_seq_length=384 \
  --doc_stride=128 \
  --output_dir=gs://some_bucket/squad_large/ \
  --use_tpu=True \
  --tpu_name=$TPU_NAME

For example, one random run with these parameters produces the following Dev scores:

{"f1": 90.87081895814865, "exact_match": 84.38978240302744}

If you fine-tune for one epoch on TriviaQA before this the results will be even better, but you will need to convert TriviaQA into the SQuAD json format.

This model is also implemented and documented in run_squad.py.

To run on SQuAD 2.0, you will first need to download the dataset. The necessary files can be found here:

• train-v2.0.json
• dev-v2.0.json
• evaluate-v2.0.py

Download these to some directory $SQUAD_DIR.

On Cloud TPU you can run with BERT-Large as follows:

python run_squad.py \
  --vocab_file=$BERT_LARGE_DIR/vocab.txt \
  --bert_config_file=$BERT_LARGE_DIR/bert_config.json \
  --init_checkpoint=$BERT_LARGE_DIR/bert_model.ckpt \
  --do_train=True \
  --train_file=$SQUAD_DIR/train-v2.0.json \
  --do_predict=True \
  --predict_file=$SQUAD_DIR/dev-v2.0.json \
  --train_batch_size=24 \
  --learning_rate=3e-5 \
  --num_train_epochs=2.0 \
  --max_seq_length=384 \
  --doc_stride=128 \
  --output_dir=gs://some_bucket/squad_large/ \
  --use_tpu=True \
  --tpu_name=$TPU_NAME \
  --version_2_with_negative=True

We assume you have copied everything from the output directory to a local directory called ./squad/. The initial dev set predictions will be at ./squad/predictions.json and the differences between the score of no answer ("") and the best non-null answer for each question will be in the file ./squad/null_odds.json

Run this script to tune a threshold for predicting null versus non-null answers:

python $SQUAD_DIR/evaluate-v2.0.py $SQUAD_DIR/dev-v2.0.json ./squad/predictions.json --na-prob-file ./squad/null_odds.json

Assume the script outputs "best_f1_thresh" THRESH. (Typical values are between -1.0 and -5.0). You can now re-run the model to generate predictions with the derived threshold or alternatively you can extract the appropriate answers from ./squad/nbest_predictions.json.

python run_squad.py \
  --vocab_file=$BERT_LARGE_DIR/vocab.txt \
  --bert_config_file=$BERT_LARGE_DIR/bert_config.json \
  --init_checkpoint=$BERT_LARGE_DIR/bert_model.ckpt \
  --do_train=False \
  --train_file=$SQUAD_DIR/train-v2.0.json \
  --do_predict=True \
  --predict_file=$SQUAD_DIR/dev-v2.0.json \
  --train_batch_size=24 \
  --learning_rate=3e-5 \
  --num_train_epochs=2.0 \
  --max_seq_length=384 \
  --doc_stride=128 \
  --output_dir=gs://some_bucket/squad_large/ \
  --use_tpu=True \
  --tpu_name=$TPU_NAME \
  --version_2_with_negative=True \
  --null_score_diff_threshold=$THRESH

All experiments in the paper were fine-tuned on a Cloud TPU, which has 64GB of device RAM. Therefore, when using a GPU with 12GB - 16GB of RAM, you are likely to encounter out-of-memory issues if you use the same hyperparameters described in the paper.

The factors that affect memory usage are:

• max_seq_length: The released models were trained with sequence lengths up to 512, but you can fine-tune with a shorter max sequence length to save substantial memory. This is controlled by the max_seq_length flag in our example code.

• train_batch_size: The memory usage is also directly proportional to the batch size.

• Model type, BERT-Base vs. BERT-Large: The BERT-Large model requires significantly more memory than BERT-Base.

• Optimizer: The default optimizer for BERT is Adam, which requires a lot of extra memory to store the m and v vectors. Switching to a more memory efficient optimizer can reduce memory usage, but can also affect the results. We have not experimented with other optimizers for fine-tuning.

Using the default training scripts (run_classifier.py and run_squad.py), we benchmarked the maximum batch size on single Titan X GPU (12GB RAM) with TensorFlow 1.11.0:

Unfortunately, these max batch sizes for BERT-Large are so small that they will actually harm the model accuracy, regardless of the learning rate used. We are working on adding code to this repository which will allow much larger effective batch sizes to be used on the GPU. The code will be based on one (or both) of the following techniques:

• Gradient accumulation: The samples in a minibatch are typically independent with respect to gradient computation (excluding batch normalization, which is not used here). This means that the gradients of multiple smaller minibatches can be accumulated before performing the weight update, and this will be exactly equivalent to a single larger update.

• Gradient checkpointing: The major use of GPU/TPU memory during DNN training is caching the intermediate activations in the forward pass that are necessary for efficient computation in the backward pass. "Gradient checkpointing" trades memory for compute time by re-computing the activations in an intelligent way.

However, this is not implemented in the current release.

In certain cases, rather than fine-tuning the entire pre-trained model end-to-end, it can be beneficial to obtained pre-trained contextual embeddings, which are fixed contextual representations of each input token generated from the hidden layers of the pre-trained model. This should also mitigate most of the out-of-memory issues.

As an example, we include the script extract_features.py which can be used like this:

# Sentence A and Sentence B are separated by the ||| delimiter for sentence
# pair tasks like question answering and entailment.
# For single sentence inputs, put one sentence per line and DON'T use the
# delimiter.
echo 'Who was Jim Henson ? ||| Jim Henson was a puppeteer' > /tmp/input.txt

python extract_features.py \
  --input_file=/tmp/input.txt \
  --output_file=/tmp/output.jsonl \
  --vocab_file=$BERT_BASE_DIR/vocab.txt \
  --bert_config_file=$BERT_BASE_DIR/bert_config.json \
  --init_checkpoint=$BERT_BASE_DIR/bert_model.ckpt \
  --layers=-1,-2,-3,-4 \
  --max_seq_length=128 \
  --batch_size=8

This will create a JSON file (one line per line of input) containing the BERT activations from each Transformer layer specified by layers (-1 is the final hidden layer of the Transformer, etc.)

Note that this script will produce very large output files (by default, around 15kb for every input token).

If you need to maintain alignment between the original and tokenized words (for projecting training labels), see the Tokenization section below.

Note: You may see a message like Could not find trained model in model_dir: /tmp/tmpuB5g5c, running initialization to predict. This message is expected, it just means that we are using the init_from_checkpoint() API rather than the saved model API. If you don't specify a checkpoint or specify an invalid checkpoint, this script will complain.

For sentence-level tasks (or sentence-pair) tasks, tokenization is very simple. Just follow the example code in run_classifier.py and extract_features.py. The basic procedure for sentence-level tasks is:

1. Instantiate an instance of tokenizer = tokenization.FullTokenizer

2. Tokenize the raw text with tokens = tokenizer.tokenize(raw_text).

3. Truncate to the maximum sequence length. (You can use up to 512, but you probably want to use shorter if possible for memory and speed reasons.)

4. Add the [CLS] and [SEP] tokens in the right place.

Word-level and span-level tasks (e.g., SQuAD and NER) are more complex, since you need to maintain alignment between your input text and output text so that you can project your training labels. SQuAD is a particularly complex example because the input labels are character-based, and SQuAD paragraphs are often longer than our maximum sequence length. See the code in run_squad.py to show how we handle this.

Before we describe the general recipe for handling word-level tasks, it's important to understand what exactly our tokenizer is doing. It has three main steps:

1. Text normalization: Convert all whitespace characters to spaces, and (for the Uncased model) lowercase the input and strip out accent markers. E.g., John Johanson's, → john johanson's,.

2. Punctuation splitting: Split all punctuation characters on both sides (i.e., add whitespace around all punctuation characters). Punctuation characters are defined as (a) Anything with a P* Unicode class, (b) any non-letter/number/space ASCII character (e.g., characters like $ which are technically not punctuation). E.g., john johanson's, → john johanson ' s ,

3. WordPiece tokenization: Apply whitespace tokenization to the output of the above procedure, and apply WordPiece tokenization to each token separately. (Our implementation is directly based on the one from tensor2tensor, which is linked). E.g., john johanson ' s , → john johan ##son ' s ,

The advantage of this scheme is that it is "compatible" with most existing English tokenizers. For example, imagine that you have a part-of-speech tagging task which looks like this:

Input:  John Johanson 's   house
Labels: NNP  NNP      POS NN

The tokenized output will look like this:

Tokens: john johan ##son ' s house

Crucially, this would be the same output as if the raw text were John Johanson's house (with no space before the 's).

If you have a pre-tokenized representation with word-level annotations, you can simply tokenize each input word independently, and deterministically maintain an original-to-tokenized alignment:

### Input
orig_tokens = ["John", "Johanson", "'s",  "house"]
labels      = ["NNP",  "NNP",      "POS", "NN"]

### Output
bert_tokens = []

# Token map will be an int -> int mapping between the `orig_tokens` index and
# the `bert_tokens` index.
orig_to_tok_map = []

tokenizer = tokenization.FullTokenizer(
    vocab_file=vocab_file, do_lower_case=True)

bert_tokens.append("[CLS]")
for orig_token in orig_tokens:
  orig_to_tok_map.append(len(bert_tokens))
  bert_tokens.extend(tokenizer.tokenize(orig_token))
bert_tokens.append("[SEP]")

# bert_tokens == ["[CLS]", "john", "johan", "##son", "'", "s", "house", "[SEP]"]
# orig_to_tok_map == [1, 2, 4, 6]

Now orig_to_tok_map can be used to project labels to the tokenized representation.

There are common English tokenization schemes which will cause a slight mismatch between how BERT was pre-trained. For example, if your input tokenization splits off contractions like do n't, this will cause a mismatch. If it is possible to do so, you should pre-process your data to convert these back to raw-looking text, but if it's not possible, this mismatch is likely not a big deal.

We are releasing code to do "masked LM" and "next sentence prediction" on an arbitrary text corpus. Note that this is not the exact code that was used for the paper (the original code was written in C++, and had some additional complexity), but this code does generate pre-training data as described in the paper.

Here's how to run the data generation. The input is a plain text file, with one sentence per line. (It is important that these be actual sentences for the "next sentence prediction" task). Documents are delimited by empty lines. The output is a set of tf.train.Examples serialized into TFRecord file format.

You can perform sentence segmentation with an off-the-shelf NLP toolkit such as spaCy. The create_pretraining_data.py script will concatenate segments until they reach the maximum sequence length to minimize computational waste from padding (see the script for more details). However, you may want to intentionally add a slight amount of noise to your input data (e.g., randomly truncate 2% of input segments) to make it more robust to non-sentential input during fine-tuning.

This script stores all of the examples for the entire input file in memory, so for large data files you should shard the input file and call the script multiple times. (You can pass in a file glob to run_pretraining.py, e.g., tf_examples.tf_record*.)

The max_predictions_per_seq is the maximum number of masked LM predictions per sequence. You should set this to around max_seq_length * masked_lm_prob (the script doesn't do that automatically because the exact value needs to be passed to both scripts).

python create_pretraining_data.py \
  --input_file=./sample_text.txt \
  --output_file=/tmp/tf_examples.tfrecord \
  --vocab_file=$BERT_BASE_DIR/vocab.txt \
  --do_lower_case=True \
  --max_seq_length=128 \
  --max_predictions_per_seq=20 \
  --masked_lm_prob=0.15 \
  --random_seed=12345 \
  --dupe_factor=5

Here's how to run the pre-training. Do not include init_checkpoint if you are pre-training from scratch. The model configuration (including vocab size) is specified in bert_config_file. This demo code only pre-trains for a small number of steps (20), but in practice you will probably want to set num_train_steps to 10000 steps or more. The max_seq_length and max_predictions_per_seq parameters passed to run_pretraining.py must be the same as create_pretraining_data.py.

python run_pretraining.py \
  --input_file=/tmp/tf_examples.tfrecord \
  --output_dir=/tmp/pretraining_output \
  --do_train=True \
  --do_eval=True \
  --bert_config_file=$BERT_BASE_DIR/bert_config.json \
  --init_checkpoint=$BERT_BASE_DIR/bert_model.ckpt \
  --train_batch_size=32 \
  --max_seq_length=128 \
  --max_predictions_per_seq=20 \
  --num_train_steps=20 \
  --num_warmup_steps=10 \
  --learning_rate=2e-5

This will produce an output like this:

***** Eval results *****
  global_step = 20
  loss = 0.0979674
  masked_lm_accuracy = 0.985479
  masked_lm_loss = 0.0979328
  next_sentence_accuracy = 1.0
  next_sentence_loss = 3.45724e-05

Note that since our sample_text.txt file is very small, this example training will overfit that data in only a few steps and produce unrealistically high accuracy numbers.

• If using your own vocabulary, make sure to change vocab_size in bert_config.json. If you use a larger vocabulary without changing this, you will likely get NaNs when training on GPU or TPU due to unchecked out-of-bounds access.
• If your task has a large domain-specific corpus available (e.g., "movie reviews" or "scientific papers"), it will likely be beneficial to run additional steps of pre-training on your corpus, starting from the BERT checkpoint.
• The learning rate we used in the paper was 1e-4. However, if you are doing additional steps of pre-training starting from an existing BERT checkpoint, you should use a smaller learning rate (e.g., 2e-5).
• Current BERT models are English-only, but we do plan to release a multilingual model which has been pre-trained on a lot of languages in the near future (hopefully by the end of November 2018).
• Longer sequences are disproportionately expensive because attention is quadratic to the sequence length. In other words, a batch of 64 sequences of length 512 is much more expensive than a batch of 256 sequences of length 128. The fully-connected/convolutional cost is the same, but the attention cost is far greater for the 512-length sequences. Therefore, one good recipe is to pre-train for, say, 90,000 steps with a sequence length of 128 and then for 10,000 additional steps with a sequence length of 512. The very long sequences are mostly needed to learn positional embeddings, which can be learned fairly quickly. Note that this does require generating the data twice with different values of max_seq_length.
• If you are pre-training from scratch, be prepared that pre-training is computationally expensive, especially on GPUs. If you are pre-training from scratch, our recommended recipe is to pre-train a BERT-Base on a single preemptible Cloud TPU v2, which takes about 2 weeks at a cost of about $500 USD (based on the pricing in October 2018). You will have to scale down the batch size when only training on a single Cloud TPU, compared to what was used in the paper. It is recommended to use the largest batch size that fits into TPU memory.

We will not be able to release the pre-processed datasets used in the paper. For Wikipedia, the recommended pre-processing is to download the latest dump, extract the text with WikiExtractor.py, and then apply any necessary cleanup to convert it into plain text.

Unfortunately the researchers who collected the BookCorpus no longer have it available for public download. The Project Guttenberg Dataset is a somewhat smaller (200M word) collection of older books that are public domain.

Common Crawl is another very large collection of text, but you will likely have to do substantial pre-processing and cleanup to extract a usable corpus for pre-training BERT.

This repository does not include code for learning a new WordPiece vocabulary. The reason is that the code used in the paper was implemented in C++ with dependencies on Google's internal libraries. For English, it is almost always better to just start with our vocabulary and pre-trained models. For learning vocabularies of other languages, there are a number of open source options available. However, keep in mind that these are not compatible with our tokenization.py library:

• Google's SentencePiece library

• tensor2tensor's WordPiece generation script

• Rico Sennrich's Byte Pair Encoding library

If you want to use BERT with Colab, you can get started with the notebook "BERT FineTuning with Cloud TPUs". At the time of this writing (October 31st, 2018), Colab users can access a Cloud TPU completely for free. Note: One per user, availability limited, requires a Google Cloud Platform account with storage (although storage may be purchased with free credit for signing up with GCP), and this capability may not longer be available in the future. Click on the BERT Colab that was just linked for more information.

Yes, all of the code in this repository works out-of-the-box with CPU, GPU, and Cloud TPU. However, GPU training is single-GPU only.

See the section on out-of-memory issues for more information.

There is no official PyTorch implementation. However, NLP researchers from HuggingFace made a PyTorch version of BERT available which is compatible with our pre-trained checkpoints and is able to reproduce our results. We were not involved in the creation or maintenance of the PyTorch implementation so please direct any questions towards the authors of that repository.

There is no official Chainer implementation. However, Sosuke Kobayashi made a Chainer version of BERT available which is compatible with our pre-trained checkpoints and is able to reproduce our results. We were not involved in the creation or maintenance of the Chainer implementation so please direct any questions towards the authors of that repository.

Yes, we plan to release a multi-lingual BERT model in the near future. We cannot make promises about exactly which languages will be included, but it will likely be a single model which includes most of the languages which have a significantly-sized Wikipedia.

So far we have not attempted to train anything larger than BERT-Large. It is possible that we will release larger models if we are able to obtain significant improvements.

All code and models are released under the Apache 2.0 license. See the LICENSE file for more information.

For now, cite the Arxiv paper:

@article{devlin2018bert,
  title={BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding},
  author={Devlin, Jacob and Chang, Ming-Wei and Lee, Kenton and Toutanova, Kristina},
  journal={arXiv preprint arXiv:1810.04805},
  year={2018}
}

If we submit the paper to a conference or journal, we will update the BibTeX.

This is not an official Google product.

For help or issues using BERT, please submit a GitHub issue.

For personal communication related to BERT, please contact Jacob Devlin (jacobdevlin@google.com), Ming-Wei Chang (mingweichang@google.com), or Kenton Lee (kentonl@google.com).