From 4544f906e27e63c3b9d8a6530e89c8b1fbc012c5 Mon Sep 17 00:00:00 2001 From: Sylvain Gugger <35901082+sgugger@users.noreply.github.com> Date: Mon, 29 Jun 2020 05:06:05 -0400 Subject: [PATCH] model cards for roberta and bert-multilingual (#5324) * More model cards (cc @myleott) * Apply suggestions from code review Co-authored-by: Julien Chaumond --- .../bert-base-multilingual-cased-README.md | 149 ++++++++++- .../bert-base-multilingual-uncased-README.md | 208 +++++++++++++++- model_cards/roberta-base-README.md | 226 ++++++++++++++++- model_cards/roberta-large-README.md | 235 ++++++++++++++++++ 4 files changed, 814 insertions(+), 4 deletions(-) create mode 100644 model_cards/roberta-large-README.md diff --git a/model_cards/bert-base-multilingual-cased-README.md b/model_cards/bert-base-multilingual-cased-README.md index 82e8c0ffbc..b290b8d0f7 100644 --- a/model_cards/bert-base-multilingual-cased-README.md +++ b/model_cards/bert-base-multilingual-cased-README.md @@ -1,5 +1,152 @@ --- language: multilingual - license: apache-2.0 +datasets: +- wikipedia --- + +# BERT multilingual base model (uncased) + +Pretrained model on the top 104 languages with the largest Wikipedia using a masked language modeling (MLM) objective. +It was introduced in [this paper](https://arxiv.org/abs/1810.04805) and first released in +[this repository](https://github.com/google-research/bert). This model is case sensitive: it makes a difference +between english and English. + +Disclaimer: The team releasing BERT did not write a model card for this model so this model card has been written by +the Hugging Face team. + +## Model description + +BERT is a transformers model pretrained on a large corpus of multilingual data in a self-supervised fashion. This means +it was pretrained on the raw texts only, with no humans labelling them in any way (which is why it can use lots of +publicly available data) with an automatic process to generate inputs and labels from those texts. More precisely, it +was pretrained with two objectives: + +- Masked language modeling (MLM): taking a sentence, the model randomly masks 15% of the words in the input then run + the entire masked sentence through the model and has to predict the masked words. This is different from traditional + recurrent neural networks (RNNs) that usually see the words one after the other, or from autoregressive models like + GPT which internally mask the future tokens. It allows the model to learn a bidirectional representation of the + sentence. +- Next sentence prediction (NSP): the models concatenates two masked sentences as inputs during pretraining. Sometimes + they correspond to sentences that were next to each other in the original text, sometimes not. The model then has to + predict if the two sentences were following each other or not. + +This way, the model learns an inner representation of the languages in the training set that can then be used to +extract features useful for downstream tasks: if you have a dataset of labeled sentences for instance, you can train a +standard classifier using the features produced by the BERT model as inputs. + +## Intended uses & limitations + +You can use the raw model for either masked language modeling or next sentence prediction, but it's mostly intended to +be fine-tuned on a downstream task. See the [model hub](https://huggingface.co/models?filter=bert) to look for +fine-tuned versions on a task that interests you. + +Note that this model is primarily aimed at being fine-tuned on tasks that use the whole sentence (potentially masked) +to make decisions, such as sequence classification, token classification or question answering. For tasks such as text +generation you should look at model like GPT2. + +### How to use + +You can use this model directly with a pipeline for masked language modeling: + +```python +>>> from transformers import pipeline +>>> unmasker = pipeline('fill-mask', model='bert-base-multilingual-cased') +>>> unmasker("Hello I'm a [MASK] model.") + +[{'sequence': "[CLS] Hello I'm a model model. [SEP]", + 'score': 0.10182085633277893, + 'token': 13192, + 'token_str': 'model'}, + {'sequence': "[CLS] Hello I'm a world model. [SEP]", + 'score': 0.052126359194517136, + 'token': 11356, + 'token_str': 'world'}, + {'sequence': "[CLS] Hello I'm a data model. [SEP]", + 'score': 0.048930276185274124, + 'token': 11165, + 'token_str': 'data'}, + {'sequence': "[CLS] Hello I'm a flight model. [SEP]", + 'score': 0.02036019042134285, + 'token': 23578, + 'token_str': 'flight'}, + {'sequence': "[CLS] Hello I'm a business model. [SEP]", + 'score': 0.020079681649804115, + 'token': 14155, + 'token_str': 'business'}] +``` + +Here is how to use this model to get the features of a given text in PyTorch: + +```python +from transformers import BertTokenizer, BertModel +tokenizer = BertTokenizer.from_pretrained('bert-base-multilingual-cased') +model = BertModel.from_pretrained("bert-base-multilingual-cased") +text = "Replace me by any text you'd like." +encoded_input = tokenizer(text, return_tensors='pt') +output = model(**encoded_input) +``` + +and in TensorFlow: + +```python +from transformers import BertTokenizer, TFBertModel +tokenizer = BertTokenizer.from_pretrained('bert-base-multilingual-cased') +model = TFBertModel.from_pretrained("bert-base-multilingual-cased") +text = "Replace me by any text you'd like." +encoded_input = tokenizer(text, return_tensors='tf') +output = model(encoded_input) +``` + +## Training data + +The BERT model was pretrained on the 104 languages with the largest Wikipedias. You can find the complete list +[here](https://github.com/google-research/bert/blob/master/multilingual.md#list-of-languages). + +## Training procedure + +### Preprocessing + +The texts are lowercased and tokenized using WordPiece and a shared vocabulary size of 110,000. The languages with a +larger Wikipedia are under-sampled and the ones with lower resources are oversampled. For languages like Chinese, +Japanese Kanji and Korean Hanja that don't have space, a CJK Unicode block is added around every character. + +The inputs of the model are then of the form: + +``` +[CLS] Sentence A [SEP] Sentence B [SEP] +``` + +With probability 0.5, sentence A and sentence B correspond to two consecutive sentences in the original corpus and in +the other cases, it's another random sentence in the corpus. Note that what is considered a sentence here is a +consecutive span of text usually longer than a single sentence. The only constrain is that the result with the two +"sentences" has a combined length of less than 512 tokens. + +The details of the masking procedure for each sentence are the following: +- 15% of the tokens are masked. +- In 80% of the cases, the masked tokens are replaced by `[MASK]`. +- In 10% of the cases, the masked tokens are replaced by a random token (different) from the one they replace. +- In the 10% remaining cases, the masked tokens are left as is. + + +### BibTeX entry and citation info + +```bibtex +@article{DBLP:journals/corr/abs-1810-04805, + author = {Jacob Devlin and + Ming{-}Wei Chang and + Kenton Lee and + Kristina Toutanova}, + title = {{BERT:} Pre-training of Deep Bidirectional Transformers for Language + Understanding}, + journal = {CoRR}, + volume = {abs/1810.04805}, + year = {2018}, + url = {http://arxiv.org/abs/1810.04805}, + archivePrefix = {arXiv}, + eprint = {1810.04805}, + timestamp = {Tue, 30 Oct 2018 20:39:56 +0100}, + biburl = {https://dblp.org/rec/journals/corr/abs-1810-04805.bib}, + bibsource = {dblp computer science bibliography, https://dblp.org} +} +``` diff --git a/model_cards/bert-base-multilingual-uncased-README.md b/model_cards/bert-base-multilingual-uncased-README.md index 82e8c0ffbc..58956861dc 100644 --- a/model_cards/bert-base-multilingual-uncased-README.md +++ b/model_cards/bert-base-multilingual-uncased-README.md @@ -1,5 +1,209 @@ --- -language: multilingual - +language: english license: apache-2.0 +datasets: +- wikipedia --- + +# BERT multilingual base model (uncased) + +Pretrained model on the top 102 languages with the largest Wikipedia using a masked language modeling (MLM) objective. +It was introduced in [this paper](https://arxiv.org/abs/1810.04805) and first released in +[this repository](https://github.com/google-research/bert). This model is uncased: it does not make a difference +between english and English. + +Disclaimer: The team releasing BERT did not write a model card for this model so this model card has been written by +the Hugging Face team. + +## Model description + +BERT is a transformers model pretrained on a large corpus of multilingual data in a self-supervised fashion. This means +it was pretrained on the raw texts only, with no humans labelling them in any way (which is why it can use lots of +publicly available data) with an automatic process to generate inputs and labels from those texts. More precisely, it +was pretrained with two objectives: + +- Masked language modeling (MLM): taking a sentence, the model randomly masks 15% of the words in the input then run + the entire masked sentence through the model and has to predict the masked words. This is different from traditional + recurrent neural networks (RNNs) that usually see the words one after the other, or from autoregressive models like + GPT which internally mask the future tokens. It allows the model to learn a bidirectional representation of the + sentence. +- Next sentence prediction (NSP): the models concatenates two masked sentences as inputs during pretraining. Sometimes + they correspond to sentences that were next to each other in the original text, sometimes not. The model then has to + predict if the two sentences were following each other or not. + +This way, the model learns an inner representation of the languages in the training set that can then be used to +extract features useful for downstream tasks: if you have a dataset of labeled sentences for instance, you can train a +standard classifier using the features produced by the BERT model as inputs. + +## Intended uses & limitations + +You can use the raw model for either masked language modeling or next sentence prediction, but it's mostly intended to +be fine-tuned on a downstream task. See the [model hub](https://huggingface.co/models?filter=bert) to look for +fine-tuned versions on a task that interests you. + +Note that this model is primarily aimed at being fine-tuned on tasks that use the whole sentence (potentially masked) +to make decisions, such as sequence classification, token classification or question answering. For tasks such as text +generation you should look at model like GPT2. + +### How to use + +You can use this model directly with a pipeline for masked language modeling: + +```python +>>> from transformers import pipeline +>>> unmasker = pipeline('fill-mask', model='bert-base-multilingual-uncased') +>>> unmasker("Hello I'm a [MASK] model.") + +[{'sequence': "[CLS] hello i'm a top model. [SEP]", + 'score': 0.1507750153541565, + 'token': 11397, + 'token_str': 'top'}, + {'sequence': "[CLS] hello i'm a fashion model. [SEP]", + 'score': 0.13075384497642517, + 'token': 23589, + 'token_str': 'fashion'}, + {'sequence': "[CLS] hello i'm a good model. [SEP]", + 'score': 0.036272723227739334, + 'token': 12050, + 'token_str': 'good'}, + {'sequence': "[CLS] hello i'm a new model. [SEP]", + 'score': 0.035954564809799194, + 'token': 10246, + 'token_str': 'new'}, + {'sequence': "[CLS] hello i'm a great model. [SEP]", + 'score': 0.028643041849136353, + 'token': 11838, + 'token_str': 'great'}] +``` + +Here is how to use this model to get the features of a given text in PyTorch: + +```python +from transformers import BertTokenizer, BertModel +tokenizer = BertTokenizer.from_pretrained('bert-base-multilingual-uncased') +model = BertModel.from_pretrained("bert-base-multilingual-uncased") +text = "Replace me by any text you'd like." +encoded_input = tokenizer(text, return_tensors='pt') +output = model(**encoded_input) +``` + +and in TensorFlow: + +```python +from transformers import BertTokenizer, TFBertModel +tokenizer = BertTokenizer.from_pretrained('bert-base-multilingual-uncased') +model = TFBertModel.from_pretrained("bert-base-multilingual-uncased") +text = "Replace me by any text you'd like." +encoded_input = tokenizer(text, return_tensors='tf') +output = model(encoded_input) +``` + +### Limitations and bias + +Even if the training data used for this model could be characterized as fairly neutral, this model can have biased +predictions: + +```python +>>> from transformers import pipeline +>>> unmasker = pipeline('fill-mask', model='bert-base-multilingual-uncased') +>>> unmasker("The man worked as a [MASK].") + +[{'sequence': '[CLS] the man worked as a teacher. [SEP]', + 'score': 0.07943806052207947, + 'token': 21733, + 'token_str': 'teacher'}, + {'sequence': '[CLS] the man worked as a lawyer. [SEP]', + 'score': 0.0629938617348671, + 'token': 34249, + 'token_str': 'lawyer'}, + {'sequence': '[CLS] the man worked as a farmer. [SEP]', + 'score': 0.03367974981665611, + 'token': 36799, + 'token_str': 'farmer'}, + {'sequence': '[CLS] the man worked as a journalist. [SEP]', + 'score': 0.03172805905342102, + 'token': 19477, + 'token_str': 'journalist'}, + {'sequence': '[CLS] the man worked as a carpenter. [SEP]', + 'score': 0.031021825969219208, + 'token': 33241, + 'token_str': 'carpenter'}] + +>>> unmasker("The Black woman worked as a [MASK].") + +[{'sequence': '[CLS] the black woman worked as a nurse. [SEP]', + 'score': 0.07045423984527588, + 'token': 52428, + 'token_str': 'nurse'}, + {'sequence': '[CLS] the black woman worked as a teacher. [SEP]', + 'score': 0.05178029090166092, + 'token': 21733, + 'token_str': 'teacher'}, + {'sequence': '[CLS] the black woman worked as a lawyer. [SEP]', + 'score': 0.032601192593574524, + 'token': 34249, + 'token_str': 'lawyer'}, + {'sequence': '[CLS] the black woman worked as a slave. [SEP]', + 'score': 0.030507225543260574, + 'token': 31173, + 'token_str': 'slave'}, + {'sequence': '[CLS] the black woman worked as a woman. [SEP]', + 'score': 0.027691684663295746, + 'token': 14050, + 'token_str': 'woman'}] +``` + +This bias will also affect all fine-tuned versions of this model. + +## Training data + +The BERT model was pretrained on the 102 languages with the largest Wikipedias. You can find the complete list +[here](https://github.com/google-research/bert/blob/master/multilingual.md#list-of-languages). + +## Training procedure + +### Preprocessing + +The texts are lowercased and tokenized using WordPiece and a shared vocabulary size of 110,000. The languages with a +larger Wikipedia are under-sampled and the ones with lower resources are oversampled. For languages like Chinese, +Japanese Kanji and Korean Hanja that don't have space, a CJK Unicode block is added around every character. + +The inputs of the model are then of the form: + +``` +[CLS] Sentence A [SEP] Sentence B [SEP] +``` + +With probability 0.5, sentence A and sentence B correspond to two consecutive sentences in the original corpus and in +the other cases, it's another random sentence in the corpus. Note that what is considered a sentence here is a +consecutive span of text usually longer than a single sentence. The only constrain is that the result with the two +"sentences" has a combined length of less than 512 tokens. + +The details of the masking procedure for each sentence are the following: +- 15% of the tokens are masked. +- In 80% of the cases, the masked tokens are replaced by `[MASK]`. +- In 10% of the cases, the masked tokens are replaced by a random token (different) from the one they replace. +- In the 10% remaining cases, the masked tokens are left as is. + + +### BibTeX entry and citation info + +```bibtex +@article{DBLP:journals/corr/abs-1810-04805, + author = {Jacob Devlin and + Ming{-}Wei Chang and + Kenton Lee and + Kristina Toutanova}, + title = {{BERT:} Pre-training of Deep Bidirectional Transformers for Language + Understanding}, + journal = {CoRR}, + volume = {abs/1810.04805}, + year = {2018}, + url = {http://arxiv.org/abs/1810.04805}, + archivePrefix = {arXiv}, + eprint = {1810.04805}, + timestamp = {Tue, 30 Oct 2018 20:39:56 +0100}, + biburl = {https://dblp.org/rec/journals/corr/abs-1810-04805.bib}, + bibsource = {dblp computer science bibliography, https://dblp.org} +} +``` diff --git a/model_cards/roberta-base-README.md b/model_cards/roberta-base-README.md index 203f029c3e..fb7b01acd6 100644 --- a/model_cards/roberta-base-README.md +++ b/model_cards/roberta-base-README.md @@ -1,10 +1,234 @@ --- +language: english tags: - exbert - license: mit +datasets: +- bookcorpus +- wikipedia --- +# RoBERTa base model + +Pretrained model on English language using a masked language modeling (MLM) objective. It was introduced in +[this paper](https://arxiv.org/abs/1907.11692) and first released in +[this repository](https://github.com/pytorch/fairseq/tree/master/examples/roberta). This model is case-sensitive: it +makes a difference between english and English. + +Disclaimer: The team releasing RoBERTa did not write a model card for this model so this model card has been written by +the Hugging Face team. + +## Model description + +RoBERTa is a transformers model pretrained on a large corpus of English data in a self-supervised fashion. This means +it was pretrained on the raw texts only, with no humans labelling them in any way (which is why it can use lots of +publicly available data) with an automatic process to generate inputs and labels from those texts. + +More precisely, it was pretrained with the Masked language modeling (MLM) objective. Taking a sentence, the model +randomly masks 15% of the words in the input then run the entire masked sentence through the model and has to predict +the masked words. This is different from traditional recurrent neural networks (RNNs) that usually see the words one +after the other, or from autoregressive models like GPT which internally mask the future tokens. It allows the model to +learn a bidirectional representation of the sentence. + +This way, the model learns an inner representation of the English language that can then be used to extract features +useful for downstream tasks: if you have a dataset of labeled sentences for instance, you can train a standard +classifier using the features produced by the BERT model as inputs. + +## Intended uses & limitations + +You can use the raw model for masked language modeling, but it's mostly intended to be fine-tuned on a downstream task. +See the [model hub](https://huggingface.co/models?filter=roberta) to look for fine-tuned versions on a task that +interests you. + +Note that this model is primarily aimed at being fine-tuned on tasks that use the whole sentence (potentially masked) +to make decisions, such as sequence classification, token classification or question answering. For tasks such as text +generation you should look at model like GPT2. + +### How to use + +You can use this model directly with a pipeline for masked language modeling: + +```python +>>> from transformers import pipeline +>>> unmasker = pipeline('fill-mask', model='roberta-base') +>>> unmasker("Hello I'm a model.") + +[{'sequence': "Hello I'm a male model.", + 'score': 0.3306540250778198, + 'token': 2943, + 'token_str': 'Ġmale'}, + {'sequence': "Hello I'm a female model.", + 'score': 0.04655390977859497, + 'token': 2182, + 'token_str': 'Ġfemale'}, + {'sequence': "Hello I'm a professional model.", + 'score': 0.04232972860336304, + 'token': 2038, + 'token_str': 'Ġprofessional'}, + {'sequence': "Hello I'm a fashion model.", + 'score': 0.037216778844594955, + 'token': 2734, + 'token_str': 'Ġfashion'}, + {'sequence': "Hello I'm a Russian model.", + 'score': 0.03253649175167084, + 'token': 1083, + 'token_str': 'ĠRussian'}] +``` + +Here is how to use this model to get the features of a given text in PyTorch: + +```python +from transformers import RobertaTokenizer, RobertaModel +tokenizer = RobertaTokenizer.from_pretrained('roberta-base') +model = RobertaModel.from_pretrained('roberta-base') +text = "Replace me by any text you'd like." +encoded_input = tokenizer(text, return_tensors='pt') +output = model(**encoded_input) +``` + +and in TensorFlow: + +```python +from transformers import RobertaTokenizer, TFRobertaModel +tokenizer = RobertaTokenizer.from_pretrained('roberta-base') +model = TFRobertaModel.from_pretrained('roberta-base') +text = "Replace me by any text you'd like." +encoded_input = tokenizer(text, return_tensors='tf') +output = model(encoded_input) +``` + +### Limitations and bias + +The training data used for this model contains a lot of unfiltered content from the internet, which is far from +neutral. Therefore, the model can have biased predictions: + +```python +>>> from transformers import pipeline +>>> unmasker = pipeline('fill-mask', model='roberta-base') +>>> unmasker("The man worked as a .") + +[{'sequence': 'The man worked as a mechanic.', + 'score': 0.08702439814805984, + 'token': 25682, + 'token_str': 'Ġmechanic'}, + {'sequence': 'The man worked as a waiter.', + 'score': 0.0819653645157814, + 'token': 38233, + 'token_str': 'Ġwaiter'}, + {'sequence': 'The man worked as a butcher.', + 'score': 0.073323555290699, + 'token': 32364, + 'token_str': 'Ġbutcher'}, + {'sequence': 'The man worked as a miner.', + 'score': 0.046322137117385864, + 'token': 18678, + 'token_str': 'Ġminer'}, + {'sequence': 'The man worked as a guard.', + 'score': 0.040150221437215805, + 'token': 2510, + 'token_str': 'Ġguard'}] + +>>> unmasker("The Black woman worked as a .") + +[{'sequence': 'The Black woman worked as a waitress.', + 'score': 0.22177888453006744, + 'token': 35698, + 'token_str': 'Ġwaitress'}, + {'sequence': 'The Black woman worked as a prostitute.', + 'score': 0.19288744032382965, + 'token': 36289, + 'token_str': 'Ġprostitute'}, + {'sequence': 'The Black woman worked as a maid.', + 'score': 0.06498628109693527, + 'token': 29754, + 'token_str': 'Ġmaid'}, + {'sequence': 'The Black woman worked as a secretary.', + 'score': 0.05375480651855469, + 'token': 2971, + 'token_str': 'Ġsecretary'}, + {'sequence': 'The Black woman worked as a nurse.', + 'score': 0.05245552211999893, + 'token': 9008, + 'token_str': 'Ġnurse'}] +``` + +This bias will also affect all fine-tuned versions of this model. + +## Training data + +The RoBERTa model was pretrained on the reunion of five datasets: +- [BookCorpus](https://yknzhu.wixsite.com/mbweb), a dataset consisting of 11,038 unpublished books; +- [English Wikipedia](https://en.wikipedia.org/wiki/English_Wikipedia) (excluding lists, tables and headers) ; +- [CC-News](https://commoncrawl.org/2016/10/news-dataset-available/), a dataset containing 63 millions English news + articles crawled between September 2016 and February 2019. +- [OpenWebText](https://github.com/jcpeterson/openwebtext), an opensource recreation of the WebText dataset used to + train GPT-2, +- [Stories](https://arxiv.org/abs/1806.02847) a dataset containing a subset of CommonCrawl data filtered to match the + story-like style of Winograd schemas. + +Together theses datasets weight 160GB of text. + +## Training procedure + +### Preprocessing + +The texts are tokenized using a byte version of Byte-Pair Encoding (BPE) and a vocabulary size of 50,000. The inputs of +the model take pieces of 512 contiguous token that may span over documents. The beginning of a new document is marked +with `` and the end of one by `` + +The details of the masking procedure for each sentence are the following: +- 15% of the tokens are masked. +- In 80% of the cases, the masked tokens are replaced by ``. +- In 10% of the cases, the masked tokens are replaced by a random token (different) from the one they replace. +- In the 10% remaining cases, the masked tokens are left as is. + +Contrary to BERT, the masking is done dynamically during pretraining (e.g., it changes at each epoch and is not fixed). + +### Pretraining + +The model was trained on 1024 V100 GPUs for 500K steps with a batch size of 8K and a sequence length of 512. The +optimizer used is Adam with a learning rate of 6e-4, \\(\beta_{1} = 0.9\\), \\(\beta_{2} = 0.98\\) and +\\(\epsilon = 1e-6\\), a weight decay of 0.01, learning rate warmup for 24,000 steps and linear decay of the learning +rate after. + +## Evaluation results + +When fine-tuned on downstream tasks, this model achieves the following results: + +Glue test results: + +| Task | MNLI | QQP | QNLI | SST-2 | CoLA | STS-B | MRPC | RTE | +|:----:|:----:|:----:|:----:|:-----:|:----:|:-----:|:----:|:----:| +| | 87.6 | 91.9 | 92.8 | 94.8 | 63.6 | 91.2 | 90.2 | 78.7 | + + +### BibTeX entry and citation info + +```bibtex +@article{DBLP:journals/corr/abs-1907-11692, + author = {Yinhan Liu and + Myle Ott and + Naman Goyal and + Jingfei Du and + Mandar Joshi and + Danqi Chen and + Omer Levy and + Mike Lewis and + Luke Zettlemoyer and + Veselin Stoyanov}, + title = {RoBERTa: {A} Robustly Optimized {BERT} Pretraining Approach}, + journal = {CoRR}, + volume = {abs/1907.11692}, + year = {2019}, + url = {http://arxiv.org/abs/1907.11692}, + archivePrefix = {arXiv}, + eprint = {1907.11692}, + timestamp = {Thu, 01 Aug 2019 08:59:33 +0200}, + biburl = {https://dblp.org/rec/journals/corr/abs-1907-11692.bib}, + bibsource = {dblp computer science bibliography, https://dblp.org} +} +``` + diff --git a/model_cards/roberta-large-README.md b/model_cards/roberta-large-README.md new file mode 100644 index 0000000000..eedc431cb1 --- /dev/null +++ b/model_cards/roberta-large-README.md @@ -0,0 +1,235 @@ +--- +language: english +tags: +- exbert +license: mit +datasets: +- bookcorpus +- wikipedia +--- + +# RoBERTa large model + +Pretrained model on English language using a masked language modeling (MLM) objective. It was introduced in +[this paper](https://arxiv.org/abs/1907.11692) and first released in +[this repository](https://github.com/pytorch/fairseq/tree/master/examples/roberta). This model is case-sensitive: it +makes a difference between english and English. + +Disclaimer: The team releasing RoBERTa did not write a model card for this model so this model card has been written by +the Hugging Face team. + +## Model description + +RoBERTa is a transformers model pretrained on a large corpus of English data in a self-supervised fashion. This means +it was pretrained on the raw texts only, with no humans labelling them in any way (which is why it can use lots of +publicly available data) with an automatic process to generate inputs and labels from those texts. + +More precisely, it was pretrained with the Masked language modeling (MLM) objective. Taking a sentence, the model +randomly masks 15% of the words in the input then run the entire masked sentence through the model and has to predict +the masked words. This is different from traditional recurrent neural networks (RNNs) that usually see the words one +after the other, or from autoregressive models like GPT which internally mask the future tokens. It allows the model to +learn a bidirectional representation of the sentence. + +This way, the model learns an inner representation of the English language that can then be used to extract features +useful for downstream tasks: if you have a dataset of labeled sentences for instance, you can train a standard +classifier using the features produced by the BERT model as inputs. + +## Intended uses & limitations + +You can use the raw model for masked language modeling, but it's mostly intended to be fine-tuned on a downstream task. +See the [model hub](https://huggingface.co/models?filter=roberta) to look for fine-tuned versions on a task that +interests you. + +Note that this model is primarily aimed at being fine-tuned on tasks that use the whole sentence (potentially masked) +to make decisions, such as sequence classification, token classification or question answering. For tasks such as text +generation you should look at model like GPT2. + +### How to use + +You can use this model directly with a pipeline for masked language modeling: + +```python +>>> from transformers import pipeline +>>> unmasker = pipeline('fill-mask', model='roberta-large') +>>> unmasker("Hello I'm a model.") + +[{'sequence': "Hello I'm a male model.", + 'score': 0.3317350447177887, + 'token': 2943, + 'token_str': 'Ġmale'}, + {'sequence': "Hello I'm a fashion model.", + 'score': 0.14171843230724335, + 'token': 2734, + 'token_str': 'Ġfashion'}, + {'sequence': "Hello I'm a professional model.", + 'score': 0.04291723668575287, + 'token': 2038, + 'token_str': 'Ġprofessional'}, + {'sequence': "Hello I'm a freelance model.", + 'score': 0.02134818211197853, + 'token': 18150, + 'token_str': 'Ġfreelance'}, + {'sequence': "Hello I'm a young model.", + 'score': 0.021098261699080467, + 'token': 664, + 'token_str': 'Ġyoung'}] +``` + +Here is how to use this model to get the features of a given text in PyTorch: + +```python +from transformers import RobertaTokenizer, RobertaModel +tokenizer = RobertaTokenizer.from_pretrained('roberta-large') +model = RobertaModel.from_pretrained('roberta-large') +text = "Replace me by any text you'd like." +encoded_input = tokenizer(text, return_tensors='pt') +output = model(**encoded_input) +``` + +and in TensorFlow: + +```python +from transformers import RobertaTokenizer, TFRobertaModel +tokenizer = RobertaTokenizer.from_pretrained('roberta-large') +model = TFRobertaModel.from_pretrained('roberta-large') +text = "Replace me by any text you'd like." +encoded_input = tokenizer(text, return_tensors='tf') +output = model(encoded_input) +``` + +### Limitations and bias + +The training data used for this model contains a lot of unfiltered content from the internet, which is far from +neutral. Therefore, the model can have biased predictions: + +```python +>>> from transformers import pipeline +>>> unmasker = pipeline('fill-mask', model='roberta-large') +>>> unmasker("The man worked as a .") + +[{'sequence': 'The man worked as a mechanic.', + 'score': 0.08260300755500793, + 'token': 25682, + 'token_str': 'Ġmechanic'}, + {'sequence': 'The man worked as a driver.', + 'score': 0.05736079439520836, + 'token': 1393, + 'token_str': 'Ġdriver'}, + {'sequence': 'The man worked as a teacher.', + 'score': 0.04709019884467125, + 'token': 3254, + 'token_str': 'Ġteacher'}, + {'sequence': 'The man worked as a bartender.', + 'score': 0.04641604796051979, + 'token': 33080, + 'token_str': 'Ġbartender'}, + {'sequence': 'The man worked as a waiter.', + 'score': 0.04239227622747421, + 'token': 38233, + 'token_str': 'Ġwaiter'}] + +>>> unmasker("The woman worked as a .") + +[{'sequence': 'The woman worked as a nurse.', + 'score': 0.2667474150657654, + 'token': 9008, + 'token_str': 'Ġnurse'}, + {'sequence': 'The woman worked as a waitress.', + 'score': 0.12280137836933136, + 'token': 35698, + 'token_str': 'Ġwaitress'}, + {'sequence': 'The woman worked as a teacher.', + 'score': 0.09747499972581863, + 'token': 3254, + 'token_str': 'Ġteacher'}, + {'sequence': 'The woman worked as a secretary.', + 'score': 0.05783602222800255, + 'token': 2971, + 'token_str': 'Ġsecretary'}, + {'sequence': 'The woman worked as a cleaner.', + 'score': 0.05576248839497566, + 'token': 16126, + 'token_str': 'Ġcleaner'}] +``` + +This bias will also affect all fine-tuned versions of this model. + +## Training data + +The RoBERTa model was pretrained on the reunion of five datasets: +- [BookCorpus](https://yknzhu.wixsite.com/mbweb), a dataset consisting of 11,038 unpublished books; +- [English Wikipedia](https://en.wikipedia.org/wiki/English_Wikipedia) (excluding lists, tables and headers) ; +- [CC-News](https://commoncrawl.org/2016/10/news-dataset-available/), a dataset containing 63 millions English news + articles crawled between September 2016 and February 2019. +- [OpenWebText](https://github.com/jcpeterson/openwebtext), an opensource recreation of the WebText dataset used to + train GPT-2, +- [Stories](https://arxiv.org/abs/1806.02847) a dataset containing a subset of CommonCrawl data filtered to match the + story-like style of Winograd schemas. + +Together theses datasets weight 160GB of text. + +## Training procedure + +### Preprocessing + +The texts are tokenized using a byte version of Byte-Pair Encoding (BPE) and a vocabulary size of 50,000. The inputs of +the model take pieces of 512 contiguous token that may span over documents. The beginning of a new document is marked +with `` and the end of one by `` + +The details of the masking procedure for each sentence are the following: +- 15% of the tokens are masked. +- In 80% of the cases, the masked tokens are replaced by ``. + +- In 10% of the cases, the masked tokens are replaced by a random token (different) from the one they replace. +- In the 10% remaining cases, the masked tokens are left as is. + +Contrary to BERT, the masking is done dynamically during pretraining (e.g., it changes at each epoch and is not fixed). + +### Pretraining + +The model was trained on 1024 V100 GPUs for 500K steps with a batch size of 8K and a sequence length of 512. The +optimizer used is Adam with a learning rate of 4e-4, \\(\beta_{1} = 0.9\\), \\(\beta_{2} = 0.98\\) and +\\(\epsilon = 1e-6\\), a weight decay of 0.01, learning rate warmup for 30,000 steps and linear decay of the learning +rate after. + +## Evaluation results + +When fine-tuned on downstream tasks, this model achieves the following results: + +Glue test results: + +| Task | MNLI | QQP | QNLI | SST-2 | CoLA | STS-B | MRPC | RTE | +|:----:|:----:|:----:|:----:|:-----:|:----:|:-----:|:----:|:----:| +| | 90.2 | 92.2 | 94.7 | 96.4 | 68.0 | 96.4 | 90.9 | 86.6 | + + +### BibTeX entry and citation info + +```bibtex +@article{DBLP:journals/corr/abs-1907-11692, + author = {Yinhan Liu and + Myle Ott and + Naman Goyal and + Jingfei Du and + Mandar Joshi and + Danqi Chen and + Omer Levy and + Mike Lewis and + Luke Zettlemoyer and + Veselin Stoyanov}, + title = {RoBERTa: {A} Robustly Optimized {BERT} Pretraining Approach}, + journal = {CoRR}, + volume = {abs/1907.11692}, + year = {2019}, + url = {http://arxiv.org/abs/1907.11692}, + archivePrefix = {arXiv}, + eprint = {1907.11692}, + timestamp = {Thu, 01 Aug 2019 08:59:33 +0200}, + biburl = {https://dblp.org/rec/journals/corr/abs-1907-11692.bib}, + bibsource = {dblp computer science bibliography, https://dblp.org} +} +``` + + + +