README.md · bert-base-uncased

1

---

2

language: en

3

tags:

4

- exbert

5

license: apache-2.0

6

datasets:

7

- bookcorpus

8

- wikipedia

9

---

10

11

# BERT base model (uncased)

12

13

Pretrained model on English language using a masked language modeling (MLM) objective. It was introduced in

14

[this paper](https://arxiv.org/abs/1810.04805) and first released in

15

[this repository](https://github.com/google-research/bert). This model is uncased: it does not make a difference

16

between english and English.

17

18

Disclaimer: The team releasing BERT did not write a model card for this model so this model card has been written by

19

the Hugging Face team.

20

21

## Model description

22

23

BERT is a transformers model pretrained on a large corpus of English data in a self-supervised fashion. This means it

24

was pretrained on the raw texts only, with no humans labeling them in any way (which is why it can use lots of

25

publicly available data) with an automatic process to generate inputs and labels from those texts. More precisely, it

26

was pretrained with two objectives:

27

28

- Masked language modeling (MLM): taking a sentence, the model randomly masks 15% of the words in the input then run

29

the entire masked sentence through the model and has to predict the masked words. This is different from traditional

30

recurrent neural networks (RNNs) that usually see the words one after the other, or from autoregressive models like

31

GPT which internally masks the future tokens. It allows the model to learn a bidirectional representation of the

32

sentence.

33

- Next sentence prediction (NSP): the models concatenates two masked sentences as inputs during pretraining. Sometimes

34

they correspond to sentences that were next to each other in the original text, sometimes not. The model then has to

35

predict if the two sentences were following each other or not.

36

37

This way, the model learns an inner representation of the English language that can then be used to extract features

38

useful for downstream tasks: if you have a dataset of labeled sentences, for instance, you can train a standard

39

classifier using the features produced by the BERT model as inputs.

40

41

## Model variations

42

43

BERT has originally been released in base and large variations, for cased and uncased input text. The uncased models also strips out an accent markers.

44

Chinese and multilingual uncased and cased versions followed shortly after.

45

Modified preprocessing with whole word masking has replaced subpiece masking in a following work, with the release of two models.

46

Other 24 smaller models are released afterward.

47

48

The detailed release history can be found on the [google-research/bert readme](https://github.com/google-research/bert/blob/master/README.md) on github.

49

50

| Model | #params | Language |

51

|------------------------|--------------------------------|-------|

52

| [`bert-base-uncased`](https://huggingface.co/bert-base-uncased) | 110M | English |

53

| [`bert-large-uncased`](https://huggingface.co/bert-large-uncased) | 340M | English | sub

54

| [`bert-base-cased`](https://huggingface.co/bert-base-cased) | 110M | English |

55

| [`bert-large-cased`](https://huggingface.co/bert-large-cased) | 340M | English |

56

| [`bert-base-chinese`](https://huggingface.co/bert-base-chinese) | 110M | Chinese |

57

| [`bert-base-multilingual-cased`](https://huggingface.co/bert-base-multilingual-cased) | 110M | Multiple |

58

| [`bert-large-uncased-whole-word-masking`](https://huggingface.co/bert-large-uncased-whole-word-masking) | 340M | English |

59

| [`bert-large-cased-whole-word-masking`](https://huggingface.co/bert-large-cased-whole-word-masking) | 340M | English |

60

61

## Intended uses & limitations

62

63

You can use the raw model for either masked language modeling or next sentence prediction, but it's mostly intended to

64

be fine-tuned on a downstream task. See the [model hub](https://huggingface.co/models?filter=bert) to look for

65

fine-tuned versions of a task that interests you.

66

67

Note that this model is primarily aimed at being fine-tuned on tasks that use the whole sentence (potentially masked)

68

to make decisions, such as sequence classification, token classification or question answering. For tasks such as text

69

generation you should look at model like GPT2.

70

71

### How to use

72

73

You can use this model directly with a pipeline for masked language modeling:

74

75

```python

76

>>> from transformers import pipeline

77

>>> unmasker = pipeline('fill-mask', model='bert-base-uncased')

78

>>> unmasker("Hello I'm a [MASK] model.")

79

80

[{'sequence': "[CLS] hello i'm a fashion model. [SEP]",

81

'score': 0.1073106899857521,

82

'token': 4827,

83

'token_str': 'fashion'},

84

{'sequence': "[CLS] hello i'm a role model. [SEP]",

85

'score': 0.08774490654468536,

86

'token': 2535,

87

'token_str': 'role'},

88

{'sequence': "[CLS] hello i'm a new model. [SEP]",

89

'score': 0.05338378623127937,

90

'token': 2047,

91

'token_str': 'new'},

92

{'sequence': "[CLS] hello i'm a super model. [SEP]",

93

'score': 0.04667217284440994,

94

'token': 3565,

95

'token_str': 'super'},

96

{'sequence': "[CLS] hello i'm a fine model. [SEP]",

97

'score': 0.027095865458250046,

98

'token': 2986,

99

'token_str': 'fine'}]

100

```

101

102

Here is how to use this model to get the features of a given text in PyTorch:

103

104

```python

105

from transformers import BertTokenizer, BertModel

106

tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')

107

model = BertModel.from_pretrained("bert-base-uncased")

108

text = "Replace me by any text you'd like."

109

encoded_input = tokenizer(text, return_tensors='pt')

110

output = model(**encoded_input)

111

```

112

113

and in TensorFlow:

114

115

```python

116

from transformers import BertTokenizer, TFBertModel

117

tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')

118

model = TFBertModel.from_pretrained("bert-base-uncased")

119

text = "Replace me by any text you'd like."

120

encoded_input = tokenizer(text, return_tensors='tf')

121

output = model(encoded_input)

122

```

123

124

### Limitations and bias

125

126

Even if the training data used for this model could be characterized as fairly neutral, this model can have biased

127

predictions:

128

129

```python

130

>>> from transformers import pipeline

131

>>> unmasker = pipeline('fill-mask', model='bert-base-uncased')

132

>>> unmasker("The man worked as a [MASK].")

133

134

[{'sequence': '[CLS] the man worked as a carpenter. [SEP]',

135

'score': 0.09747550636529922,

136

'token': 10533,

137

'token_str': 'carpenter'},

138

{'sequence': '[CLS] the man worked as a waiter. [SEP]',

139

'score': 0.0523831807076931,

140

'token': 15610,

141

'token_str': 'waiter'},

142

{'sequence': '[CLS] the man worked as a barber. [SEP]',

143

'score': 0.04962705448269844,

144

'token': 13362,

145

'token_str': 'barber'},

146

{'sequence': '[CLS] the man worked as a mechanic. [SEP]',

147

'score': 0.03788609802722931,

148

'token': 15893,

149

'token_str': 'mechanic'},

150

{'sequence': '[CLS] the man worked as a salesman. [SEP]',

151

'score': 0.037680890411138535,

152

'token': 18968,

153

'token_str': 'salesman'}]

154

155

>>> unmasker("The woman worked as a [MASK].")

156

157

[{'sequence': '[CLS] the woman worked as a nurse. [SEP]',

158

'score': 0.21981462836265564,

159

'token': 6821,

160

'token_str': 'nurse'},

161

{'sequence': '[CLS] the woman worked as a waitress. [SEP]',

162

'score': 0.1597415804862976,

163

'token': 13877,

164

'token_str': 'waitress'},

165

{'sequence': '[CLS] the woman worked as a maid. [SEP]',

166

'score': 0.1154729500412941,

167

'token': 10850,

168

'token_str': 'maid'},

169

{'sequence': '[CLS] the woman worked as a prostitute. [SEP]',

170

'score': 0.037968918681144714,

171

'token': 19215,

172

'token_str': 'prostitute'},

173

{'sequence': '[CLS] the woman worked as a cook. [SEP]',

174

'score': 0.03042375110089779,

175

'token': 5660,

176

'token_str': 'cook'}]

177

```

178

179

This bias will also affect all fine-tuned versions of this model.

180

181

## Training data

182

183

The BERT model was pretrained on [BookCorpus](https://yknzhu.wixsite.com/mbweb), a dataset consisting of 11,038

184

unpublished books and [English Wikipedia](https://en.wikipedia.org/wiki/English_Wikipedia) (excluding lists, tables and

185

headers).

186

187

## Training procedure

188

189

### Preprocessing

190

191

The texts are lowercased and tokenized using WordPiece and a vocabulary size of 30,000. The inputs of the model are

192

then of the form:

193

194

```

195

[CLS] Sentence A [SEP] Sentence B [SEP]

196

```

197

198

With probability 0.5, sentence A and sentence B correspond to two consecutive sentences in the original corpus, and in

199

the other cases, it's another random sentence in the corpus. Note that what is considered a sentence here is a

200

consecutive span of text usually longer than a single sentence. The only constrain is that the result with the two

201

"sentences" has a combined length of less than 512 tokens.

202

203

The details of the masking procedure for each sentence are the following:

204

- 15% of the tokens are masked.

205

- In 80% of the cases, the masked tokens are replaced by `[MASK]`.

206

- In 10% of the cases, the masked tokens are replaced by a random token (different) from the one they replace.

207

- In the 10% remaining cases, the masked tokens are left as is.

208

209

### Pretraining

210

211

The model was trained on 4 cloud TPUs in Pod configuration (16 TPU chips total) for one million steps with a batch size

212

of 256. The sequence length was limited to 128 tokens for 90% of the steps and 512 for the remaining 10%. The optimizer

213

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,

214

learning rate warmup for 10,000 steps and linear decay of the learning rate after.

215

216

## Evaluation results

217

218

When fine-tuned on downstream tasks, this model achieves the following results:

219

220

Glue test results:

221

222

| Task | MNLI-(m/mm) | QQP | QNLI | SST-2 | CoLA | STS-B | MRPC | RTE | Average |

223

|:----:|:-----------:|:----:|:----:|:-----:|:----:|:-----:|:----:|:----:|:-------:|

224

| | 84.6/83.4 | 71.2 | 90.5 | 93.5 | 52.1 | 85.8 | 88.9 | 66.4 | 79.6 |

225

226

227

### BibTeX entry and citation info

228

229

```bibtex

230

@article{DBLP:journals/corr/abs-1810-04805,

231

author = {Jacob Devlin and

232

Ming{-}Wei Chang and

233

Kenton Lee and

234

Kristina Toutanova},

235

title = {{BERT:} Pre-training of Deep Bidirectional Transformers for Language

236

Understanding},

237

journal = {CoRR},

238

volume = {abs/1810.04805},

239

year = {2018},

240

url = {http://arxiv.org/abs/1810.04805},

241

archivePrefix = {arXiv},

242

eprint = {1810.04805},

243

timestamp = {Tue, 30 Oct 2018 20:39:56 +0100},

244

biburl = {https://dblp.org/rec/journals/corr/abs-1810-04805.bib},

245

bibsource = {dblp computer science bibliography, https://dblp.org}

246

}

247

```

248

249

<a href="https://huggingface.co/exbert/?model=bert-base-uncased">

250

<img width="300px" src="https://cdn-media.huggingface.co/exbert/button.png">

251

</a>

252