finnianx/LMSYS-GPT-5-RESPONSE

🤖 LMSYS-Chat-GPT-5-Chat-Response

• The dataset used in Black-Box On-Policy Distillation of Large Language Models paper. Homepage at here.
• This dataset is an extension of the LMSYS-Chat-1M-Clean corpus, specifically curated by collecting high-quality, non-refusal responses from the GPT-5-Chat API.
• The LMSYS-Chat-1M dataset collects real-world user queries from the Chatbot Arena.
• There is no tool calls or reasoning in the GPT-5-Chat response.

💾 Dataset Structure

The dataset contains the following splits and columns:

Split Name	Number of Examples	Description
train	Around 200,000	Train set
test	Around 500	Test set

Column Name	Data Type	Description
content	string	The original user prompt/question from the LMSYS-Chat dataset
teacher_response	string	The response generated by the GPT-5-Chat API

📊 Diversity of Categories

The underlying LMSYS-Chat dataset contains a wide and realistic range of user intentions.

The categories present in the data include:

Type of Task/Query
Code	coding	debugging	translation
Logic/Reasoning	logical reasoning	spatial reasoning	pattern recognition	debating
Instruction Following	instruction following	specific format writing	information extraction	summarization
Creative/Writing	creative writing	copywriting	roleplaying	text completion
Analysis	sentiment analysis	text comparison	text classification	explanation
General	question answering	free-form chat	trivia	brainstorming
Math & Planning	math	planning and scheduling
Editing/Correction	proofreading	paraphrasing	text manipulation
Ethics	ethical reasoning
Other	tutorial	question generation

📄 Citation

If you find this work useful, please cite our paper:

@article{ye2025blackboxonpolicydistillationlarge,
  title={Black-Box On-Policy Distillation of Large Language Models},
  author={Tianzhu Ye and Li Dong and Zewen Chi and Xun Wu and Shaohan Huang and Furu Wei},
  journal={arXiv preprint arXiv:2511.10643},
  year={2025},
  url={https://arxiv.org/abs/2511.10643}
}