metadata

tags:
  - bert
  - oBERT
  - sparsity
  - pruning
  - compression
language: en
datasets: mnli

oBERT-12-upstream-pruned-unstructured-97-finetuned-mnli

This model is obtained with The Optimal BERT Surgeon: Scalable and Accurate Second-Order Pruning for Large Language Models.

It corresponds to the model presented in the Table 2 - oBERT - MNLI 97%.

Pruning method: oBERT upstream unstructured + sparse-transfer to downstream
Paper: https://arxiv.org/abs/2203.07259
Dataset: MNLI
Sparsity: 97%
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 97%    | m-acc | mm-acc|
| ------------ | ----- | ----- |
| seed=42      | 78.55 | 79.90 |
| seed=3407    | 78.88 | 79.78 |
| seed=54321(*)| 79.11 | 79.71 |
| ------------ | ----- | ----- |
| mean         | 78.85 | 79.80 |
| stdev        | 0.281 | 0.096 |

Code: https://github.com/neuralmagic/sparseml/tree/main/research/optimal_BERT_surgeon_oBERT

If you find the model useful, please consider citing our work.

Citation info

@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}
}