# bert-large-cased

 1 --- 2 language: en 3 license: apache-2.0 4 datasets: 5 - bookcorpus 6 - wikipedia 7 --- 8 9 # BERT large model (cased) 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 cased: it makes 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 ## Intended uses & limitations 47 48 You can use the raw model for either masked language modeling or next sentence prediction, but it's mostly intended to 49 be fine-tuned on a downstream task. See the [model hub](https://huggingface.co/models?filter=bert) to look for 50 fine-tuned versions on a task that interests you. 51 52 Note that this model is primarily aimed at being fine-tuned on tasks that use the whole sentence (potentially masked) 53 to make decisions, such as sequence classification, token classification or question answering. For tasks such as text 54 generation you should look at model like GPT2. 55 56 ### How to use 57 58 You can use this model directly with a pipeline for masked language modeling: 59 60 python 61 >>> from transformers import pipeline 62 >>> unmasker = pipeline('fill-mask', model='bert-large-cased') 63 >>> unmasker("Hello I'm a [MASK] model.") 64 [ 65 { 66 "sequence":"[CLS] Hello I'm a male model. [SEP]", 67 "score":0.22748498618602753, 68 "token":2581, 69 "token_str":"male" 70 }, 71 { 72 "sequence":"[CLS] Hello I'm a fashion model. [SEP]", 73 "score":0.09146175533533096, 74 "token":4633, 75 "token_str":"fashion" 76 }, 77 { 78 "sequence":"[CLS] Hello I'm a new model. [SEP]", 79 "score":0.05823173746466637, 80 "token":1207, 81 "token_str":"new" 82 }, 83 { 84 "sequence":"[CLS] Hello I'm a super model. [SEP]", 85 "score":0.04488750174641609, 86 "token":7688, 87 "token_str":"super" 88 }, 89 { 90 "sequence":"[CLS] Hello I'm a famous model. [SEP]", 91 "score":0.03271442651748657, 92 "token":2505, 93 "token_str":"famous" 94 } 95 ] 96  97 98 Here is how to use this model to get the features of a given text in PyTorch: 99 100 python 101 from transformers import BertTokenizer, BertModel 102 tokenizer = BertTokenizer.from_pretrained('bert-large-cased') 103 model = BertModel.from_pretrained("bert-large-cased") 104 text = "Replace me by any text you'd like." 105 encoded_input = tokenizer(text, return_tensors='pt') 106 output = model(**encoded_input) 107  108 109 and in TensorFlow: 110 111 python 112 from transformers import BertTokenizer, TFBertModel 113 tokenizer = BertTokenizer.from_pretrained('bert-large-cased') 114 model = TFBertModel.from_pretrained("bert-large-cased") 115 text = "Replace me by any text you'd like." 116 encoded_input = tokenizer(text, return_tensors='tf') 117 output = model(encoded_input) 118  119 120 ### Limitations and bias 121 122 Even if the training data used for this model could be characterized as fairly neutral, this model can have biased 123 predictions: 124 125 python 126 >>> from transformers import pipeline 127 >>> unmasker = pipeline('fill-mask', model='bert-large-cased') 128 >>> unmasker("The man worked as a [MASK].") 129 [ 130 { 131 "sequence":"[CLS] The man worked as a doctor. [SEP]", 132 "score":0.0645911768078804, 133 "token":3995, 134 "token_str":"doctor" 135 }, 136 { 137 "sequence":"[CLS] The man worked as a cop. [SEP]", 138 "score":0.057450827211141586, 139 "token":9947, 140 "token_str":"cop" 141 }, 142 { 143 "sequence":"[CLS] The man worked as a mechanic. [SEP]", 144 "score":0.04392256215214729, 145 "token":19459, 146 "token_str":"mechanic" 147 }, 148 { 149 "sequence":"[CLS] The man worked as a waiter. [SEP]", 150 "score":0.03755280375480652, 151 "token":17989, 152 "token_str":"waiter" 153 }, 154 { 155 "sequence":"[CLS] The man worked as a teacher. [SEP]", 156 "score":0.03458863124251366, 157 "token":3218, 158 "token_str":"teacher" 159 } 160 ] 161 162 >>> unmasker("The woman worked as a [MASK].") 163 [ 164 { 165 "sequence":"[CLS] The woman worked as a nurse. [SEP]", 166 "score":0.2572779953479767, 167 "token":7439, 168 "token_str":"nurse" 169 }, 170 { 171 "sequence":"[CLS] The woman worked as a waitress. [SEP]", 172 "score":0.16706500947475433, 173 "token":15098, 174 "token_str":"waitress" 175 }, 176 { 177 "sequence":"[CLS] The woman worked as a teacher. [SEP]", 178 "score":0.04587847739458084, 179 "token":3218, 180 "token_str":"teacher" 181 }, 182 { 183 "sequence":"[CLS] The woman worked as a secretary. [SEP]", 184 "score":0.03577028587460518, 185 "token":4848, 186 "token_str":"secretary" 187 }, 188 { 189 "sequence":"[CLS] The woman worked as a maid. [SEP]", 190 "score":0.03298963978886604, 191 "token":13487, 192 "token_str":"maid" 193 } 194 ] 195  196 197 This bias will also affect all fine-tuned versions of this model. 198 199 ## Training data 200 201 The BERT model was pretrained on [BookCorpus](https://yknzhu.wixsite.com/mbweb), a dataset consisting of 11,038 202 unpublished books and [English Wikipedia](https://en.wikipedia.org/wiki/English_Wikipedia) (excluding lists, tables and 203 headers). 204 205 ## Training procedure 206 207 ### Preprocessing 208 209 The texts are lowercased and tokenized using WordPiece and a vocabulary size of 30,000. The inputs of the model are 210 then of the form: 211 212  213 [CLS] Sentence A [SEP] Sentence B [SEP] 214  215 216 With probability 0.5, sentence A and sentence B correspond to two consecutive sentences in the original corpus and in 217 the other cases, it's another random sentence in the corpus. Note that what is considered a sentence here is a 218 consecutive span of text usually longer than a single sentence. The only constrain is that the result with the two 219 "sentences" has a combined length of less than 512 tokens. 220 221 The details of the masking procedure for each sentence are the following: 222 - 15% of the tokens are masked. 223 - In 80% of the cases, the masked tokens are replaced by [MASK]. 224 - In 10% of the cases, the masked tokens are replaced by a random token (different) from the one they replace. 225 - In the 10% remaining cases, the masked tokens are left as is. 226 227 ### Pretraining 228 229 The model was trained on 4 cloud TPUs in Pod configuration (16 TPU chips total) for one million steps with a batch size 230 of 256. The sequence length was limited to 128 tokens for 90% of the steps and 512 for the remaining 10%. The optimizer 231 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, 232 learning rate warmup for 10,000 steps and linear decay of the learning rate after. 233 234 ## Evaluation results 235 236 When fine-tuned on downstream tasks, this model achieves the following results: 237 238 Model | SQUAD 1.1 F1/EM | Multi NLI Accuracy 239 ---------------------------------------- | :-------------: | :----------------: 240 BERT-Large, Cased (Original) | 91.5/84.8 | 86.09 241 242 ### BibTeX entry and citation info 243 244 bibtex 245 @article{DBLP:journals/corr/abs-1810-04805, 246 author = {Jacob Devlin and 247 Ming{-}Wei Chang and 248 Kenton Lee and 249 Kristina Toutanova}, 250 title = {{BERT:} Pre-training of Deep Bidirectional Transformers for Language 251 Understanding}, 252 journal = {CoRR}, 253 volume = {abs/1810.04805}, 254 year = {2018}, 255 url = {http://arxiv.org/abs/1810.04805}, 256 archivePrefix = {arXiv}, 257 eprint = {1810.04805}, 258 timestamp = {Tue, 30 Oct 2018 20:39:56 +0100}, 259 biburl = {https://dblp.org/rec/journals/corr/abs-1810-04805.bib}, 260 bibsource = {dblp computer science bibliography, https://dblp.org} 261 } 262  263