File size: 2,714 Bytes
9bea1d3 5760137 e591e7b 506f950 e591e7b b4f554d 5760137 e591e7b ca6cded e591e7b 1ec8e05 e591e7b | 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 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 | ---
license: apache-2.0
---
---
**Paper**: [https://arxiv.org/pdf/2310.06694.pdf](https://arxiv.org/pdf/2310.06694.pdf)
**Code**: https://github.com/princeton-nlp/LLM-Shearing
**Models**: [Sheared-LLaMA-1.3B](https://huggingface.co/princeton-nlp/Sheared-LLaMA-1.3B), [Sheared-LLaMA-2.7B](https://huggingface.co/princeton-nlp/Sheared-LLaMA-2.7B)
**License**: Must comply with license of Llama2 since it's a model derived from Llama2.
---
Sheared-LLaMA-2.7B is a model pruned and further pre-trained from [meta-llama/Llama-2-7b-hf](https://huggingface.co/meta-llama/Llama-2-7b-hf). We dynamically load data from different domains in the [RedPajama dataset](https://github.com/togethercomputeub.com/togethercomputer/RedPajama-Data). We use 0.4B tokens for pruning and 50B tokens for continued pre-training the pruned model. This model can be loaded into huggingface via
```
model = AutoModelForCausalLM.from_pretrained("princeton-nlp/Sheared-LLaMA-2.7B")
```
- Smaller-scale
- Same vocabulary as LLaMA1 and LLaMA2
- Derived with a budget of 50B tokens by utilizing existing strong LLMs
## Downstream Tasks
We evaluate on an extensive set of downstream tasks including reasoning, reading comprehension, language modeling and knowledge intensive tasks. Our Sheared-LLaMA models outperform existing large language models.
| Model | # Pre-training Tokens | Average Performance |
| ------------------- | --------------------- | ------------------- |
| LLaMA2-7B | 2T | 64.6 |
**1.3B**
| Model | # Pre-training Tokens | Average Performance |
| ------------------- | --------------------- | ------------------- |
| OPT-1.3B | 300B | 48.2 |
| Pythia-1.4B | 300B | 48.9 |
| Sheared-LLaMA-1.3B | 50B | 51.0 |
**3B**
| Model | # Pre-training Tokens | Average Performance |
| ------------------- | --------------------- | ------------------- |
| OPT-2.7B | 300B | 51.4 |
| Pythia-2.8B | 300B | 52.5 |
| INCITE-Base-3B | 800B | 54.7 |
| Open-LLaMA-3B-v1 | 1T | 55.1 |
| Open-LLaMA-3B-v2 | 1T | 55.7 |
| **Sheared-LLaMA-2.7B** | **50B** | **56.7** |
## Bibtex
```
@article{xia2023sheared,
title={Sheared LLaMA: Accelerating Language Model Pre-training via Structured Pruning},
author={Xia, Mengzhou and Gao, Tianyu, and Zeng, Zhiyuan and Chen, Danqi},
year={2023}
}
``` |