README.md
8.8 KB · 229 lines · markdown Raw
1 ---
2 language: en
3 license: apache-2.0
4 datasets:
5 - bookcorpus
6 - wikipedia
7 ---
8
9 # BERT large model (uncased)
10
11 Pretrained model on English language using a masked language modeling (MLM) objective. It was introduced in
12 [this paper](https://arxiv.org/abs/1810.04805) and first released in
13 [this repository](https://github.com/google-research/bert). This model is uncased: it does not make a difference
14 between english and English.
15
16 Disclaimer: The team releasing BERT did not write a model card for this model so this model card has been written by
17 the Hugging Face team.
18
19 ## Model description
20
21 BERT is a transformers model pretrained on a large corpus of English data in a self-supervised fashion. This means it
22 was pretrained on the raw texts only, with no humans labelling them in any way (which is why it can use lots of
23 publicly available data) with an automatic process to generate inputs and labels from those texts. More precisely, it
24 was pretrained with two objectives:
25
26 - Masked language modeling (MLM): taking a sentence, the model randomly masks 15% of the words in the input then run
27 the entire masked sentence through the model and has to predict the masked words. This is different from traditional
28 recurrent neural networks (RNNs) that usually see the words one after the other, or from autoregressive models like
29 GPT which internally mask the future tokens. It allows the model to learn a bidirectional representation of the
30 sentence.
31 - Next sentence prediction (NSP): the models concatenates two masked sentences as inputs during pretraining. Sometimes
32 they correspond to sentences that were next to each other in the original text, sometimes not. The model then has to
33 predict if the two sentences were following each other or not.
34
35 This way, the model learns an inner representation of the English language that can then be used to extract features
36 useful for downstream tasks: if you have a dataset of labeled sentences for instance, you can train a standard
37 classifier using the features produced by the BERT model as inputs.
38
39 This model has the following configuration:
40
41 - 24-layer
42 - 1024 hidden dimension
43 - 16 attention heads
44 - 336M parameters.
45
46
47 ## Intended uses & limitations
48
49 You can use the raw model for either masked language modeling or next sentence prediction, but it's mostly intended to
50 be fine-tuned on a downstream task. See the [model hub](https://huggingface.co/models?filter=bert) to look for
51 fine-tuned versions on a task that interests you.
52
53 Note that this model is primarily aimed at being fine-tuned on tasks that use the whole sentence (potentially masked)
54 to make decisions, such as sequence classification, token classification or question answering. For tasks such as text
55 generation you should look at model like GPT2.
56
57 ### How to use
58
59 You can use this model directly with a pipeline for masked language modeling:
60
61 ```python
62 >>> from transformers import pipeline
63 >>> unmasker = pipeline('fill-mask', model='bert-large-uncased')
64 >>> unmasker("Hello I'm a [MASK] model.")
65 [{'sequence': "[CLS] hello i'm a fashion model. [SEP]",
66 'score': 0.1886913776397705,
67 'token': 4827,
68 'token_str': 'fashion'},
69 {'sequence': "[CLS] hello i'm a professional model. [SEP]",
70 'score': 0.07157472521066666,
71 'token': 2658,
72 'token_str': 'professional'},
73 {'sequence': "[CLS] hello i'm a male model. [SEP]",
74 'score': 0.04053466394543648,
75 'token': 3287,
76 'token_str': 'male'},
77 {'sequence': "[CLS] hello i'm a role model. [SEP]",
78 'score': 0.03891477733850479,
79 'token': 2535,
80 'token_str': 'role'},
81 {'sequence': "[CLS] hello i'm a fitness model. [SEP]",
82 'score': 0.03038121573626995,
83 'token': 10516,
84 'token_str': 'fitness'}]
85 ```
86
87 Here is how to use this model to get the features of a given text in PyTorch:
88
89 ```python
90 from transformers import BertTokenizer, BertModel
91 tokenizer = BertTokenizer.from_pretrained('bert-large-uncased')
92 model = BertModel.from_pretrained("bert-large-uncased")
93 text = "Replace me by any text you'd like."
94 encoded_input = tokenizer(text, return_tensors='pt')
95 output = model(**encoded_input)
96 ```
97
98 and in TensorFlow:
99
100 ```python
101 from transformers import BertTokenizer, TFBertModel
102 tokenizer = BertTokenizer.from_pretrained('bert-large-uncased')
103 model = TFBertModel.from_pretrained("bert-large-uncased")
104 text = "Replace me by any text you'd like."
105 encoded_input = tokenizer(text, return_tensors='tf')
106 output = model(encoded_input)
107 ```
108
109 ### Limitations and bias
110
111 Even if the training data used for this model could be characterized as fairly neutral, this model can have biased
112 predictions:
113
114 ```python
115 >>> from transformers import pipeline
116 >>> unmasker = pipeline('fill-mask', model='bert-large-uncased')
117 >>> unmasker("The man worked as a [MASK].")
118
119 [{'sequence': '[CLS] the man worked as a bartender. [SEP]',
120 'score': 0.10426565259695053,
121 'token': 15812,
122 'token_str': 'bartender'},
123 {'sequence': '[CLS] the man worked as a waiter. [SEP]',
124 'score': 0.10232779383659363,
125 'token': 15610,
126 'token_str': 'waiter'},
127 {'sequence': '[CLS] the man worked as a mechanic. [SEP]',
128 'score': 0.06281787157058716,
129 'token': 15893,
130 'token_str': 'mechanic'},
131 {'sequence': '[CLS] the man worked as a lawyer. [SEP]',
132 'score': 0.050936125218868256,
133 'token': 5160,
134 'token_str': 'lawyer'},
135 {'sequence': '[CLS] the man worked as a carpenter. [SEP]',
136 'score': 0.041034240275621414,
137 'token': 10533,
138 'token_str': 'carpenter'}]
139
140 >>> unmasker("The woman worked as a [MASK].")
141
142 [{'sequence': '[CLS] the woman worked as a waitress. [SEP]',
143 'score': 0.28473711013793945,
144 'token': 13877,
145 'token_str': 'waitress'},
146 {'sequence': '[CLS] the woman worked as a nurse. [SEP]',
147 'score': 0.11336520314216614,
148 'token': 6821,
149 'token_str': 'nurse'},
150 {'sequence': '[CLS] the woman worked as a bartender. [SEP]',
151 'score': 0.09574324637651443,
152 'token': 15812,
153 'token_str': 'bartender'},
154 {'sequence': '[CLS] the woman worked as a maid. [SEP]',
155 'score': 0.06351090222597122,
156 'token': 10850,
157 'token_str': 'maid'},
158 {'sequence': '[CLS] the woman worked as a secretary. [SEP]',
159 'score': 0.048970773816108704,
160 'token': 3187,
161 'token_str': 'secretary'}]
162 ```
163
164 This bias will also affect all fine-tuned versions of this model.
165
166 ## Training data
167
168 The BERT model was pretrained on [BookCorpus](https://yknzhu.wixsite.com/mbweb), a dataset consisting of 11,038
169 unpublished books and [English Wikipedia](https://en.wikipedia.org/wiki/English_Wikipedia) (excluding lists, tables and
170 headers).
171
172 ## Training procedure
173
174 ### Preprocessing
175
176 The texts are lowercased and tokenized using WordPiece and a vocabulary size of 30,000. The inputs of the model are
177 then of the form:
178
179 ```
180 [CLS] Sentence A [SEP] Sentence B [SEP]
181 ```
182
183 With probability 0.5, sentence A and sentence B correspond to two consecutive sentences in the original corpus and in
184 the other cases, it's another random sentence in the corpus. Note that what is considered a sentence here is a
185 consecutive span of text usually longer than a single sentence. The only constrain is that the result with the two
186 "sentences" has a combined length of less than 512 tokens.
187
188 The details of the masking procedure for each sentence are the following:
189 - 15% of the tokens are masked.
190 - In 80% of the cases, the masked tokens are replaced by `[MASK]`.
191 - In 10% of the cases, the masked tokens are replaced by a random token (different) from the one they replace.
192 - In the 10% remaining cases, the masked tokens are left as is.
193
194 ### Pretraining
195
196 The model was trained on 4 cloud TPUs in Pod configuration (16 TPU chips total) for one million steps with a batch size
197 of 256. The sequence length was limited to 128 tokens for 90% of the steps and 512 for the remaining 10%. The optimizer
198 used is Adam with a learning rate of 1e-4, \\(\beta_{1} = 0.9\\) and \\(\beta_{2} = 0.999\\), a weight decay of 0.01,
199 learning rate warmup for 10,000 steps and linear decay of the learning rate after.
200
201 ## Evaluation results
202
203 When fine-tuned on downstream tasks, this model achieves the following results:
204
205 Model | SQUAD 1.1 F1/EM | Multi NLI Accuracy
206 ---------------------------------------- | :-------------: | :----------------:
207 BERT-Large, Uncased (Original) | 91.0/84.3 | 86.05
208
209 ### BibTeX entry and citation info
210
211 ```bibtex
212 @article{DBLP:journals/corr/abs-1810-04805,
213 author = {Jacob Devlin and
214 Ming{-}Wei Chang and
215 Kenton Lee and
216 Kristina Toutanova},
217 title = {{BERT:} Pre-training of Deep Bidirectional Transformers for Language
218 Understanding},
219 journal = {CoRR},
220 volume = {abs/1810.04805},
221 year = {2018},
222 url = {http://arxiv.org/abs/1810.04805},
223 archivePrefix = {arXiv},
224 eprint = {1810.04805},
225 timestamp = {Tue, 30 Oct 2018 20:39:56 +0100},
226 biburl = {https://dblp.org/rec/journals/corr/abs-1810-04805.bib},
227 bibsource = {dblp computer science bibliography, https://dblp.org}
228 }
229 ```