---
base_model:
- meta-llama/Llama-3.2-3B-Instruct
datasets:
- songff/GenerAlign
language:
- en
license: apache-2.0
pipeline_tag: text-generation
library_name: transformers
---

Pilot-3B is designed to be a draft model in efficient preference alignment of LLMs for its small size while high performance in general domains. It is trained from [Llama-3.2-3B-Instruct](https://huggingface.co/meta-llama/Llama-3.2-3B-Instruct) on [GenerAlign](https://huggingface.co/datasets/songff/GenerAlign).

Related links:

- Paper: [Well Begun is Half Done: Low-resource Preference Alignment by Weak-to-Strong Decoding](https://arxiv.org/abs/2506.07434)

- Github: [Weak-to-Strong-Decoding](https://github.com/F2-Song/Weak-to-Strong-Decoding)

- Dataset: [GenerAlign](https://huggingface.co/datasets/songff/GenerAlign)

# ⚠️Caution

Pilot-3B is not guaranteed always to provide safe and correct responses. Please use it at your own risk.

# Citation
If you find this work useful, please consider citing:
```bibtex
@misc{song2025well,
  title={Well Begun is Half Done: Low-resource Preference Alignment by Weak-to-Strong Decoding},
  author={Song, Feifan and Wei, Shaohang and Luo, Wen and Fan, Yuxuan and Liu, Tianyu and Wang, Guoyin and Wang, Houfeng},
  year={2025},
  eprint={2506.07434},
  archivePrefix={arXiv},
  primaryClass={cs.CL}
}
```