Add metadata
#1
by
nielsr
HF Staff
- opened
README.md
CHANGED
@@ -1,59 +1,61 @@
|
|
1 |
-
---
|
2 |
-
license: llama3.1
|
3 |
-
datasets:
|
4 |
-
- BAAI/Infinity-Instruct
|
5 |
-
base_model:
|
6 |
-
- meta-llama/Llama-3.1-8B-Instruct
|
7 |
-
tags:
|
8 |
-
- Instruct_Tuning
|
9 |
-
|
10 |
-
|
11 |
-
|
12 |
-
|
13 |
-
|
14 |
-
|
15 |
-
<a href="https://
|
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 |
-
|
|
|
|
|
|
1 |
+
---
|
2 |
+
license: llama3.1
|
3 |
+
datasets:
|
4 |
+
- BAAI/Infinity-Instruct
|
5 |
+
base_model:
|
6 |
+
- meta-llama/Llama-3.1-8B-Instruct
|
7 |
+
tags:
|
8 |
+
- Instruct_Tuning
|
9 |
+
library_name: transformers
|
10 |
+
pipeline_tag: text-generation
|
11 |
+
---
|
12 |
+
|
13 |
+
# Shadow-FT
|
14 |
+
|
15 |
+
<a href="https://arxiv.org/pdf/2505.12716"><b>[π Paper]</b></a> β’
|
16 |
+
<a href="https://huggingface.co/collections/taki555/shadow-ft-683288b49e1e5e1edcf03135"><b>[π€ HF Models]</b></a> β’
|
17 |
+
<a href="https://github.com/wutaiqiang/Shadow-FT"><b>[π± GitHub]</b></a>
|
18 |
+
|
19 |
+
This repo contains the weights from our paper: <a href="https://arxiv.org/pdf/2505.12716" target="_blank">Shadow-FT: Tuning Instruct via Base</a> by <a href="https://wutaiqiang.github.io" target="_blank">Taiqiang Wu*</a> <a href="https://rummyyang.github.io/" target="_blank">Runming Yang*</a>, Jiayi Li, Pengfei Hu, Ngai Wong and Yujiu Yang.
|
20 |
+
|
21 |
+
\* for equal contributions.
|
22 |
+
|
23 |
+
|
24 |
+
|
25 |
+
## Overview
|
26 |
+
|
27 |
+
<img src="framework.png" width="100%" />
|
28 |
+
|
29 |
+
Observation:
|
30 |
+
|
31 |
+
- Directly tuning the INSTRUCT (i.e., instruction tuned) models often leads to marginal improvements and even performance degeneration.
|
32 |
+
|
33 |
+
- Paired BASE models, the foundation for these INSTRUCT variants, contain highly similar weight values (i.e., less than 2% on average for Llama 3.1 8B).
|
34 |
+
|
35 |
+
$\Rightarrow$ We propose the Shadow-FT framework to tune the INSTRUCT models by leveraging the corresponding BASE models. The key insight is to fine-tune the BASE model, and then _directly_ graft the learned weight updates to the INSTRUCT model.
|
36 |
+
|
37 |
+
|
38 |
+
## Performance
|
39 |
+
|
40 |
+
This repository contains the Llama-3.1-8B tuned on BAAI-2k subsets using Shadow-FT.
|
41 |
+
|
42 |
+
<img src="performance.png" width="100%" />
|
43 |
+
|
44 |
+
please refer to [our paper](https://arxiv.org/pdf/2505.12716) for details.
|
45 |
+
|
46 |
+
|
47 |
+
|
48 |
+
## βοΈ Citation
|
49 |
+
|
50 |
+
If you find this repository helpful, please consider citing our paper:
|
51 |
+
|
52 |
+
```
|
53 |
+
@article{wu2025shadow,
|
54 |
+
title={Shadow-FT: Tuning Instruct via Base},
|
55 |
+
author={Wu, Taiqiang and Yang, Runming and Li, Jiayi and Hu, Pengfei and Wong, Ngai and Yang, Yujiu},
|
56 |
+
journal={arXiv preprint arXiv:2505.12716},
|
57 |
+
year={2025}
|
58 |
+
}
|
59 |
+
```
|
60 |
+
|
61 |
+
For any questions, please pull an issue or email at `[email protected]`
|