Skip to content

Commit eedf571

Browse files
jacobdevlin-googlejacobdevlin-google
authored
Merge pull request google-research#1027 from iuliaturc-google/master
Add links to 24 smaller BERT models.
2 parents cc7051d + 8028c04 commit eedf571

File tree

1 file changed

+71
-0
lines changed

1 file changed

+71
-0
lines changed

‎README.md

Lines changed: 71 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -1,5 +1,76 @@
11
# BERT
22

3+
**\*\*\*\*\* New March 11th, 2020: Smaller BERT Models \*\*\*\*\***
4+
5+
This is a release of 24 smaller BERT models (English only, uncased, trained with WordPiece masking) referenced in [Well-Read Students Learn Better: On the Importance of Pre-training Compact Models](https://arxiv.org/abs/1908.08962).
6+
7+
We have shown that the standard BERT recipe (including model architecture and training objective) is effective on a wide range of model sizes, beyond BERT-Base and BERT-Large. The smaller BERT models are intended for environments with restricted computational resources. They can be fine-tuned in the same manner as the original BERT models. However, they are most effective in the context of knowledge distillation, where the fine-tuning labels are produced by a larger and more accurate teacher.
8+
9+
Our goal is to enable research in institutions with fewer computational resources and encourage the community to seek directions of innovation alternative to increasing model capacity.
10+
11+
You can download all 24 from [here][all], or individually from the table below:
12+
13+
| |H=128|H=256|H=512|H=768|
14+
|---|:---:|:---:|:---:|:---:|
15+
| **L=2** |[**2/128 (BERT-Tiny)**][2_128]|[2/256][2_256]|[2/512][2_512]|[2/768][2_768]|
16+
| **L=4** |[4/128][4_128]|[**4/256 (BERT-Mini)**][4_256]|[**4/512 (BERT-Small)**][4_512]|[4/768][4_768]|
17+
| **L=6** |[6/128][6_128]|[6/256][6_256]|[6/512][6_512]|[6/768][6_768]|
18+
| **L=8** |[8/128][8_128]|[8/256][8_256]|[**8/512 (BERT-Medium)**][8_512]|[8/768][8_768]|
19+
| **L=10** |[10/128][10_128]|[10/256][10_256]|[10/512][10_512]|[10/768][10_768]|
20+
| **L=12** |[12/128][12_128]|[12/256][12_256]|[12/512][12_512]|[**12/768 (BERT-Base)**][12_768]|
21+
22+
Note that the BERT-Base model in this release is included for completeness only; it was re-trained under the same regime as the original model.
23+
24+
Here are the corresponding GLUE scores on the test set:
25+
26+
|Model|Score|CoLA|SST-2|MRPC|STS-B|QQP|MNLI-m|MNLI-mm|QNLI(v2)|RTE|WNLI|AX|
27+
|---|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|
28+
|BERT-Tiny|64.2|0.0|83.2|81.1/71.1|74.3/73.6|62.2/83.4|70.2|70.3|81.5|57.2|62.3|21.0|
29+
|BERT-Mini|65.8|0.0|85.9|81.1/71.8|75.4/73.3|66.4/86.2|74.8|74.3|84.1|57.9|62.3|26.1|
30+
|BERT-Small|71.2|27.8|89.7|83.4/76.2|78.8/77.0|68.1/87.0|77.6|77.0|86.4|61.8|62.3|28.6|
31+
|BERT-Medium|73.5|38.0|89.6|86.6/81.6|80.4/78.4|69.6/87.9|80.0|79.1|87.7|62.2|62.3|30.5|
32+
33+
For each task, we selected the best fine-tuning hyperparameters from the lists below, and trained for 4 epochs:
34+
- batch sizes: 8, 16, 32, 64, 128
35+
- learning rates: 3e-4, 1e-4, 5e-5, 3e-5
36+
37+
If you use these models, please cite the following paper:
38+
39+
```
40+
@article{turc2019,
41+
title={Well-Read Students Learn Better: On the Importance of Pre-training Compact Models},
42+
author={Turc, Iulia and Chang, Ming-Wei and Lee, Kenton and Toutanova, Kristina},
43+
journal={arXiv preprint arXiv:1908.08962v2 },
44+
year={2019}
45+
}
46+
```
47+
48+
[2_128]: https://storage.googleapis.com/bert_models/2020_02_20/uncased_L-2_H-128_A-2.zip
49+
[2_256]: https://storage.googleapis.com/bert_models/2020_02_20/uncased_L-2_H-256_A-4.zip
50+
[2_512]: https://storage.googleapis.com/bert_models/2020_02_20/uncased_L-2_H-512_A-8.zip
51+
[2_768]: https://storage.googleapis.com/bert_models/2020_02_20/uncased_L-2_H-768_A-12.zip
52+
[4_128]: https://storage.googleapis.com/bert_models/2020_02_20/uncased_L-4_H-128_A-2.zip
53+
[4_256]: https://storage.googleapis.com/bert_models/2020_02_20/uncased_L-4_H-256_A-4.zip
54+
[4_512]: https://storage.googleapis.com/bert_models/2020_02_20/uncased_L-4_H-512_A-8.zip
55+
[4_768]: https://storage.googleapis.com/bert_models/2020_02_20/uncased_L-4_H-768_A-12.zip
56+
[6_128]: https://storage.googleapis.com/bert_models/2020_02_20/uncased_L-6_H-128_A-2.zip
57+
[6_256]: https://storage.googleapis.com/bert_models/2020_02_20/uncased_L-6_H-256_A-4.zip
58+
[6_512]: https://storage.googleapis.com/bert_models/2020_02_20/uncased_L-6_H-512_A-8.zip
59+
[6_768]: https://storage.googleapis.com/bert_models/2020_02_20/uncased_L-6_H-768_A-12.zip
60+
[8_128]: https://storage.googleapis.com/bert_models/2020_02_20/uncased_L-8_H-128_A-2.zip
61+
[8_256]: https://storage.googleapis.com/bert_models/2020_02_20/uncased_L-8_H-256_A-4.zip
62+
[8_512]: https://storage.googleapis.com/bert_models/2020_02_20/uncased_L-8_H-512_A-8.zip
63+
[8_768]: https://storage.googleapis.com/bert_models/2020_02_20/uncased_L-8_H-768_A-12.zip
64+
[10_128]: https://storage.googleapis.com/bert_models/2020_02_20/uncased_L-10_H-128_A-2.zip
65+
[10_256]: https://storage.googleapis.com/bert_models/2020_02_20/uncased_L-10_H-256_A-4.zip
66+
[10_512]: https://storage.googleapis.com/bert_models/2020_02_20/uncased_L-10_H-512_A-8.zip
67+
[10_768]: https://storage.googleapis.com/bert_models/2020_02_20/uncased_L-10_H-768_A-12.zip
68+
[12_128]: https://storage.googleapis.com/bert_models/2020_02_20/uncased_L-12_H-128_A-2.zip
69+
[12_256]: https://storage.googleapis.com/bert_models/2020_02_20/uncased_L-12_H-256_A-4.zip
70+
[12_512]: https://storage.googleapis.com/bert_models/2020_02_20/uncased_L-12_H-512_A-8.zip
71+
[12_768]: https://storage.googleapis.com/bert_models/2020_02_20/uncased_L-12_H-768_A-12.zip
72+
[all]: https://storage.googleapis.com/bert_models/2020_02_20/all_bert_models.zip
73+
374
**\*\*\*\*\* New May 31st, 2019: Whole Word Masking Models \*\*\*\*\***
475

576
This is a release of several new models which were the result of an improvement

0 commit comments

Comments
 (0)