princeton-nlp
commited on
Commit
•
e591e7b
1
Parent(s):
b4f554d
Update README.md
Browse files
README.md
CHANGED
@@ -6,8 +6,54 @@ license: apache-2.0
|
|
6 |
license: apache-2.0
|
7 |
---
|
8 |
|
|
|
|
|
|
|
|
|
|
|
|
|
9 |
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
|
10 |
|
11 |
```
|
12 |
model = AutoModelForCausalLM.from_pretrained("princeton-nlp/Sheared-LLaMA-2.7B")
|
13 |
```
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
6 |
license: apache-2.0
|
7 |
---
|
8 |
|
9 |
+
**Paper**: [https://arxiv.org/pdf/2310.06694.pdf](https://arxiv.org/pdf/2310.06694.pdf)
|
10 |
+
**Code**: https://github.com/princeton-nlp/LLM-Shearing
|
11 |
+
**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)
|
12 |
+
|
13 |
+
---
|
14 |
+
|
15 |
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
|
16 |
|
17 |
```
|
18 |
model = AutoModelForCausalLM.from_pretrained("princeton-nlp/Sheared-LLaMA-2.7B")
|
19 |
```
|
20 |
+
|
21 |
+
- Smaller-scale
|
22 |
+
- Same vocabulary as LLaMA1 and LLaMA2
|
23 |
+
- Derived with 50B tokens by utilizing existing strong LLMs
|
24 |
+
|
25 |
+
## Downstream Tasks
|
26 |
+
|
27 |
+
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.
|
28 |
+
|
29 |
+
| Model | # Pre-training Tokens | Average Performance |
|
30 |
+
| ------------------- | --------------------- | ------------------- |
|
31 |
+
| LLaMA2-7B | 2T | 64.6 |
|
32 |
+
|
33 |
+
**1.3B**
|
34 |
+
|
35 |
+
| Model | # Pre-training Tokens | Average Performance |
|
36 |
+
| ------------------- | --------------------- | ------------------- |
|
37 |
+
| OPT-1.3B | 300B | 48.2 |
|
38 |
+
| Pythia-1.4B | 300B | 48.9 |
|
39 |
+
| Sheared-LLaMA-1.3B | 50B | 51.0 |
|
40 |
+
|
41 |
+
**3B**
|
42 |
+
|
43 |
+
| Model | # Pre-training Tokens | Average Performance |
|
44 |
+
| ------------------- | --------------------- | ------------------- |
|
45 |
+
| OPT-2.7B | 300B | 51.4 |
|
46 |
+
| Pythia-2.8B | 300B | 52.5 |
|
47 |
+
| INCITE-Base-3B | 800B | 54.7 |
|
48 |
+
| Open-LLaMA-3B-v1 | 1T | 55.1 |
|
49 |
+
| Open-LLaMA-3B-v2 | 1T | 55.7 |
|
50 |
+
| **Sheared-LLaMA-2.7B** | **50B** | **56.7** |
|
51 |
+
|
52 |
+
## Bibtex
|
53 |
+
```
|
54 |
+
@article{xia2023sheared,
|
55 |
+
title={Sheared LLaMA: Accelerating Language Model Pre-training via Structured Pruning},
|
56 |
+
author={Xia, Mengzhou and Gao, Tianyu, and Zeng Zhiyuan, and Chen Danqi},
|
57 |
+
year={2023}
|
58 |
+
}
|
59 |
+
```
|