Qwen3-4B-Shadow-FT-BAAI-2k

Shadow-FT

<a href="https://arxiv.org/pdf/2505.12716"><b>[📜 Paper]</b></a> • <a href="https://huggingface.co/collections/taki555/shadow-ft-683288b49e1e5e1edcf03135"><b>[🤗 HF Models]</b></a> • <a href="https://github.com/wutaiqiang/Shadow-FT"><b>[🐱 GitHub]</b></a>

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.

* for equal contributions.

Overview

<img src="framework.png" width="100%" />

Observation:

• Directly tuning the INSTRUCT (i.e., instruction tuned) models often leads to marginal improvements and even performance degeneration.

• 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).

$\Rightarrow$ 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.

Performance

This repository contains the Qwen-3-4B tuned on BAAI-2k subsets using Shadow-FT.

<img src="performance.png" width="100%" />

please refer to our paper for details.

☕️ Citation

If you find this repository helpful, please consider citing our paper:

@article{wu2025shadow,
  title={Shadow-FT: Tuning Instruct via Base},
  author={Wu, Taiqiang and Yang, Runming and Li, Jiayi and Hu, Pengfei and Wong, Ngai and Yang, Yujiu},
  journal={arXiv preprint arXiv:2505.12716},
  year={2025}
}

For any questions, please pull an issue or email at [email protected]