Timber

Official code for paper Timber: Training-free Instruct Model Refining with Base via Effective Rank

[🤗 HF Models] • [📜 Paper] • [🐱 GitHub]

This repo contains the code for our paper: Timber: Training-free Instruct Model Refining with Base via Effective Rank by Taiqiang Wu, Runming Yang, Tao Liu, Jiahao Wang, Zenan Xu, and Ngai Wong.

Overview

-> Ranks from paired Base and Instruct models are almost the same.

--> Reinforce the hypothesis that post-training is superficial.

---> Post-training improves the exploitation capabilities at the cost of limiting its exploration

----> Timber, a simple yet effective training-free method by refining the weight deltas.

Quick Start

Environment

Please follow the official guidance of Opencompass to set up a python environment.

We use the lmdeploy backend, please remember to set

pip install "opencompass[lmdeploy]"

Remember to fix the code base following this issue for thinking.

Weight Refine via Timber

Download the official weights from huggingface:

• Llama
• Qwen3

We recommend to download via the huggingface-cli, such as

hf download Qwen/Qwen3-30B-A3B --token $your_hf_token --local-dir weights/Qwen3-30B-A3B/Qwen3-30B-A3B

Then, run the timber.py:

python3 -u timber.py $path/to/base $path/to/instruct $path/to/save --gamma  0.0 --svd_cache_path $your_path/cache.pt

where gamma is the scale factor and svd_cache_path is the cache file for eRank.

Evaluate

We employ the opencompass for evaluation.

You need to modify the config files first.

For example, in Evaluation/llama_1B.py, replace the paths with your folder, modify the tp and num_gpus to fit your machine.

Then all you need is to run opencompass Evaluation/llama_1B.py and wait the final results.

License

We use the Apache‑2.0 license. Please also comply with the licenses of any upstream models and datasets.

☕️ Citation

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

@article{wu2025timber,
  title={Timber: Training-free Instruct Model Refining with Base via Effective Rank},
  author={Wu, Taiqiang and Yang, Runming and Liu, Tao and Wang, Jiahao and Xu, Zenan and Wong, Ngai.},
  journal={arXiv preprint arXiv:2509.23595},
  year={2025}
}

For any questions, please pull an issue or email at takiwu@connect.hku.hk