license: apache-2.0
task_categories:
- text-generation
language: - en
pretty_name: GenerAlign
Dataset Card
GenerAlign is collected to help construct well-aligned LLMs in general domains, such as harmlessness, helpfulness, and honesty. It contains 31398 prompts from existed datasets, including:
Similar to UltraFeedback, we complete each prompt with responses from different LLMs, including:
All responses are annotated by ArmoRM-Llama3-8B-v0.1.
This dataset has been used to acquire Pilot-3B, which serves as the default draft model in Weak-to-Strong Decoding.
Related links:
Paper: Well Begun is Half Done: Low-resource Preference Alignment by Weak-to-Strong Decoding
Github: Weak-to-Strong-Decoding
Draft model: Pilot-3B
⚠️Caution
GenerAlign inevitably contains sensitive prompts, which can be harmless, offensive and even against human values. Furthermore, the responses contained are LLM-generated and cannot be guaranteed to be correct, factual and safe. Therefore, this dataset can be and should be used for research purposes only, instead of arbitrary abuse or distribution. Please use them at your own risk.
Citation
If you find this work useful, please consider citing:
@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}
}