File size: 1,299 Bytes
b59d9b9
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
# oBERT-12-upstream-pruned-unstructured-90-finetuned-squadv1

This model is obtained with [The Optimal BERT Surgeon: Scalable and Accurate Second-Order Pruning for Large Language Models](https://arxiv.org/abs/2203.07259).


It corresponds to the model presented in the `Table 2 - oBERT - SQuADv1 90%`.

```
Pruning method: oBERT upstream unstructured + sparse-transfer to downstream
Paper: https://arxiv.org/abs/2203.07259
Dataset: SQuADv1
Sparsity: 90%
Number of layers: 12
```

The dev-set performance reported in the paper is averaged over three seeds, and we release the best model (marked with `(*)`):

```
| oBERT 90%    | F1    | EM    |
| ------------ | ----- | ----- |
| seed=42   (*)| 88.47 | 81.43 |
| seed=3407    | 88.32 | 81.13 |
| seed=54321   | 88.47 | 81.38 |
| ------------ | ----- | ----- |
| mean         | 88.42 | 81.31 |
| stdev        | 0.086 | 0.160 |
```

Code: _coming soon_

## BibTeX entry and citation info
```bibtex
@article{kurtic2022optimal,
  title={The Optimal BERT Surgeon: Scalable and Accurate Second-Order Pruning for Large Language Models},
  author={Kurtic, Eldar and Campos, Daniel and Nguyen, Tuan and Frantar, Elias and Kurtz, Mark and Fineran, Benjamin and Goin, Michael and Alistarh, Dan},
  journal={arXiv preprint arXiv:2203.07259},
  year={2022}
}
```