|
--- |
|
tags: |
|
- bert |
|
- oBERT |
|
- sparsity |
|
- pruning |
|
- compression |
|
language: en |
|
datasets: |
|
- bookcorpus |
|
- wikipedia |
|
--- |
|
# oBERT-3-upstream-pretrained-dense |
|
|
|
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 3 layers from `neuralmagic/oBERT-12-upstream-pretrained-dense`, pretrained with knowledge distillation. This model is used as a starting point for downstream finetuning and pruning runs presented in the `Table 3 - 3 Layers`. |
|
The model can also be used for finetuning on any downstream task, as a starting point instead of the three times larger `bert-base-uncased` model. |
|
|
|
Finetuned and pruned versions of this model on the SQuADv1 downstream task, as described in the paper: |
|
- 0%: `neuralmagic/oBERT-3-downstream-dense-squadv1` |
|
- 80% unstructured: `neuralmagic/oBERT-3-downstream-pruned-unstructured-80-squadv1` |
|
- 80% block-4: `neuralmagic/oBERT-3-downstream-pruned-block4-80-squadv1` |
|
- 90% unstructured: `neuralmagic/oBERT-3-downstream-pruned-unstructured-90-squadv1` |
|
- 90% block-4: `neuralmagic/oBERT-3-downstream-pruned-block4-90-squadv1` |
|
|
|
``` |
|
Training objective: masked language modeling (MLM) + knowledge distillation |
|
Paper: https://arxiv.org/abs/2203.07259 |
|
Dataset: BookCorpus and English Wikipedia |
|
Sparsity: 0% |
|
Number of layers: 3 |
|
``` |
|
|
|
Code: [https://github.com/neuralmagic/sparseml/tree/main/research/optimal_BERT_surgeon_oBERT](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 |
|
```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} |
|
} |
|
``` |
