README.md

---
license: apache-2.0
language:
- en
- zh
---

![image/png](https://huggingface.co/datasets/nbeerbower/cover-images/resolve/main/firewall-dpo.png?download=true)

# GreatFirewall-DPO

An experimental dataset to discourage censorship in Chinese models.

## Structure

- `prompt`: input text presented to model (en translated to zh)
- `chosen`: preferred response demonstrating less self-censorship (en translated to zh)  
- `rejected`: response generated by [Qwen/Qwen2.5-32B-Instruct](https://huggingface.co/Qwen/Qwen2.5-32B-Instruct), many exhibiting excessive self-censorship (generated in both en and zh)

## Content

1. CHINA-related (144 prompts) - mostly about sensitive historical/political events
2. POLITICS comes in second (84) - lots of freedom of speech and government criticism stuff
3. SOCIAL issues (75) - cultural revolution, societal impacts
4. CENSORSHIP specific (38) - website blocking, speech restrictions
5. HISTORICAL events (34) - heavy focus on tiananmen square
6. TECHNOLOGY (19) - mostly about circumventing censorship

## Disclaimer

* Contains potentially sensitive political topics. This data may also be considered biased.
* Chosen response and prompt translations are available in Chinese, however they are machine-translated and not guaranteed to be accurate or high quality.
* This data is available for research purposes, not to make a political statement.

# Gratitude and Appreciation

和平与爱同在。
谢谢中国对开源人工智能的贡献。