1
---
2
language: en
3
tags:
4
- exbert
5
license: apache-2.0
6
datasets:
7
- bookcorpus
8
- wikipedia
9
---
10
11
# DistilBERT base model (uncased)
12
13
This model is a distilled version of the [BERT base model](https://huggingface.co/bert-base-uncased). It was
14
introduced in [this paper](https://arxiv.org/abs/1910.01108). The code for the distillation process can be found
15
[here](https://github.com/huggingface/transformers/tree/master/examples/distillation). This model is uncased: it does
16
not make a difference between english and English.
17
18
## Model description
19
20
DistilBERT is a transformers model, smaller and faster than BERT, which was pretrained on the same corpus in a
21
self-supervised fashion, using the BERT base model as a teacher. This means it was pretrained on the raw texts only,
22
with no humans labelling them in any way (which is why it can use lots of publicly available data) with an automatic
23
process to generate inputs and labels from those texts using the BERT base model. More precisely, it was pretrained
24
with three objectives:
25
26
- Distillation loss: the model was trained to return the same probabilities as the BERT base model.
27
- Masked language modeling (MLM): this is part of the original training loss of the BERT base model. When taking a
28
  sentence, the model randomly masks 15% of the words in the input then run the entire masked sentence through the
29
  model and has to predict the masked words. This is different from traditional recurrent neural networks (RNNs) that
30
  usually see the words one after the other, or from autoregressive models like GPT which internally mask the future
31
  tokens. It allows the model to learn a bidirectional representation of the sentence.
32
- Cosine embedding loss: the model was also trained to generate hidden states as close as possible as the BERT base
33
  model.
34
35
This way, the model learns the same inner representation of the English language than its teacher model, while being
36
faster for inference or downstream tasks.
37
38
## Intended uses & limitations
39
40
You can use the raw model for either masked language modeling or next sentence prediction, but it's mostly intended to
41
be fine-tuned on a downstream task. See the [model hub](https://huggingface.co/models?filter=distilbert) to look for
42
fine-tuned versions on a task that interests you.
43
44
Note that this model is primarily aimed at being fine-tuned on tasks that use the whole sentence (potentially masked)
45
to make decisions, such as sequence classification, token classification or question answering. For tasks such as text
46
generation you should look at model like GPT2.
47
48
### How to use
49
50
You can use this model directly with a pipeline for masked language modeling:
51
52
```python
53
>>> from transformers import pipeline
54
>>> unmasker = pipeline('fill-mask', model='distilbert-base-uncased')
55
>>> unmasker("Hello I'm a [MASK] model.")
56
57
[{'sequence': "[CLS] hello i'm a role model. [SEP]",
58
  'score': 0.05292855575680733,
59
  'token': 2535,
60
  'token_str': 'role'},
61
 {'sequence': "[CLS] hello i'm a fashion model. [SEP]",
62
  'score': 0.03968575969338417,
63
  'token': 4827,
64
  'token_str': 'fashion'},
65
 {'sequence': "[CLS] hello i'm a business model. [SEP]",
66
  'score': 0.034743521362543106,
67
  'token': 2449,
68
  'token_str': 'business'},
69
 {'sequence': "[CLS] hello i'm a model model. [SEP]",
70
  'score': 0.03462274372577667,
71
  'token': 2944,
72
  'token_str': 'model'},
73
 {'sequence': "[CLS] hello i'm a modeling model. [SEP]",
74
  'score': 0.018145186826586723,
75
  'token': 11643,
76
  'token_str': 'modeling'}]
77
```
78
79
Here is how to use this model to get the features of a given text in PyTorch:
80
81
```python
82
from transformers import DistilBertTokenizer, DistilBertModel
83
tokenizer = DistilBertTokenizer.from_pretrained('distilbert-base-uncased')
84
model = DistilBertModel.from_pretrained("distilbert-base-uncased")
85
text = "Replace me by any text you'd like."
86
encoded_input = tokenizer(text, return_tensors='pt')
87
output = model(**encoded_input)
88
```
89
90
and in TensorFlow:
91
92
```python
93
from transformers import DistilBertTokenizer, TFDistilBertModel
94
tokenizer = DistilBertTokenizer.from_pretrained('distilbert-base-uncased')
95
model = TFDistilBertModel.from_pretrained("distilbert-base-uncased")
96
text = "Replace me by any text you'd like."
97
encoded_input = tokenizer(text, return_tensors='tf')
98
output = model(encoded_input)
99
```
100
101
### Limitations and bias
102
103
Even if the training data used for this model could be characterized as fairly neutral, this model can have biased
104
predictions. It also inherits some of
105
[the bias of its teacher model](https://huggingface.co/bert-base-uncased#limitations-and-bias).
106
107
```python
108
>>> from transformers import pipeline
109
>>> unmasker = pipeline('fill-mask', model='distilbert-base-uncased')
110
>>> unmasker("The White man worked as a [MASK].")
111
112
[{'sequence': '[CLS] the white man worked as a blacksmith. [SEP]',
113
  'score': 0.1235365942120552,
114
  'token': 20987,
115
  'token_str': 'blacksmith'},
116
 {'sequence': '[CLS] the white man worked as a carpenter. [SEP]',
117
  'score': 0.10142576694488525,
118
  'token': 10533,
119
  'token_str': 'carpenter'},
120
 {'sequence': '[CLS] the white man worked as a farmer. [SEP]',
121
  'score': 0.04985016956925392,
122
  'token': 7500,
123
  'token_str': 'farmer'},
124
 {'sequence': '[CLS] the white man worked as a miner. [SEP]',
125
  'score': 0.03932540491223335,
126
  'token': 18594,
127
  'token_str': 'miner'},
128
 {'sequence': '[CLS] the white man worked as a butcher. [SEP]',
129
  'score': 0.03351764753460884,
130
  'token': 14998,
131
  'token_str': 'butcher'}]
132
133
>>> unmasker("The Black woman worked as a [MASK].")
134
135
[{'sequence': '[CLS] the black woman worked as a waitress. [SEP]',
136
  'score': 0.13283951580524445,
137
  'token': 13877,
138
  'token_str': 'waitress'},
139
 {'sequence': '[CLS] the black woman worked as a nurse. [SEP]',
140
  'score': 0.12586183845996857,
141
  'token': 6821,
142
  'token_str': 'nurse'},
143
 {'sequence': '[CLS] the black woman worked as a maid. [SEP]',
144
  'score': 0.11708822101354599,
145
  'token': 10850,
146
  'token_str': 'maid'},
147
 {'sequence': '[CLS] the black woman worked as a prostitute. [SEP]',
148
  'score': 0.11499975621700287,
149
  'token': 19215,
150
  'token_str': 'prostitute'},
151
 {'sequence': '[CLS] the black woman worked as a housekeeper. [SEP]',
152
  'score': 0.04722772538661957,
153
  'token': 22583,
154
  'token_str': 'housekeeper'}]
155
```
156
157
This bias will also affect all fine-tuned versions of this model.
158
159
## Training data
160
161
DistilBERT pretrained on the same data as BERT, which is [BookCorpus](https://yknzhu.wixsite.com/mbweb), a dataset
162
consisting of 11,038 unpublished books and [English Wikipedia](https://en.wikipedia.org/wiki/English_Wikipedia)
163
(excluding lists, tables and headers).
164
165
## Training procedure
166
167
### Preprocessing
168
169
The texts are lowercased and tokenized using WordPiece and a vocabulary size of 30,000. The inputs of the model are
170
then of the form:
171
172
```
173
[CLS] Sentence A [SEP] Sentence B [SEP]
174
```
175
176
With probability 0.5, sentence A and sentence B correspond to two consecutive sentences in the original corpus and in
177
the other cases, it's another random sentence in the corpus. Note that what is considered a sentence here is a
178
consecutive span of text usually longer than a single sentence. The only constrain is that the result with the two
179
"sentences" has a combined length of less than 512 tokens.
180
181
The details of the masking procedure for each sentence are the following:
182
- 15% of the tokens are masked.
183
- In 80% of the cases, the masked tokens are replaced by `[MASK]`.
184
- In 10% of the cases, the masked tokens are replaced by a random token (different) from the one they replace.
185
- In the 10% remaining cases, the masked tokens are left as is.
186
187
### Pretraining
188
189
The model was trained on 8 16 GB V100 for 90 hours. See the
190
[training code](https://github.com/huggingface/transformers/tree/master/examples/distillation) for all hyperparameters
191
details.
192
193
## Evaluation results
194
195
When fine-tuned on downstream tasks, this model achieves the following results:
196
197
Glue test results:
198
199
| Task | MNLI | QQP  | QNLI | SST-2 | CoLA | STS-B | MRPC | RTE  |
200
|:----:|:----:|:----:|:----:|:-----:|:----:|:-----:|:----:|:----:|
201
|      | 82.2 | 88.5 | 89.2 | 91.3  | 51.3 | 85.8  | 87.5 | 59.9 |
202
203
204
### BibTeX entry and citation info
205
206
```bibtex
207
@article{Sanh2019DistilBERTAD,
208
  title={DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter},
209
  author={Victor Sanh and Lysandre Debut and Julien Chaumond and Thomas Wolf},
210
  journal={ArXiv},
211
  year={2019},
212
  volume={abs/1910.01108}
213
}
214
```
215
216
<a href="https://huggingface.co/exbert/?model=distilbert-base-uncased">
217
	<img width="300px" src="https://cdn-media.huggingface.co/exbert/button.png">
218
</a>
219