---
license: mit
language:
- en
tags:
- orpo
- safety
- llm
- train
size_categories:
- 10K<n<100K
---



<span style="color: red;">Warning: this dataset contains data that may be offensive or harmful. The data are intended for research purposes, especially research that can make models less harmful.</span>

[👇original PKU-SafeRLHF datasets (click 🔗 for more details)](https://huggingface.co/datasets/PKU-Alignment/PKU-SafeRLHF-single-dimension)
<img src='https://s3.bmp.ovh/imgs/2025/04/27/8b8b63100ee69318.png' width='800' height='600' />


what's the advantage of this **train** dataset over the original one ?
- **standard chosen/rejected format of preference datasets** : make 'chosen' and 'rejected' according to 'better_response_id'
- **only one file** : merge three train datasets(Alpaca-7B、Alpaca2-7B、Alpaca3-8B) into one file