HuatuoGPT Medical Question-Answer Pairs for Instruction Tuning

Over 276,000 conversational instruction-response pairs for medical question answering, primarily in Chinese. The dataset was created by author shibing624 by converting and merging two HuatuoGPT fine-tuning datasets into a ShareGPT format. It was last updated on the Hugging Face platform in January 2024.

Use Cases

• Fine-tuning LLMs to generate medical advice by learning from the 'instruction' and 'response' fields in the ShareGPT format.
• Training models for multi-turn medical dialogue using the conversational structure implied by the ShareGPT format.
• Benchmarking model performance on Chinese medical knowledge recall and reasoning tasks based on the QA pairs.
• Creating synthetic medical QA data by using the high-quality GPT-4 generated subset of 50,000 instruction pairs as a seed.

Strengths

• Contains 276,042 total instruction-response pairs from two merged sources.
• Includes a subset of 50,000 pairs generated with GPT-4 for higher quality instructions.
• Data is structured in a standardized ShareGPT JSONL format for compatibility with common training pipelines.

Limitations

• Specific column names and internal data schema are not documented in the provided input.
• The geographic and temporal scope of the medical knowledge within the QA pairs is unspecified.
• Potential class imbalance or topic bias within the medical subjects covered is unknown.

Provenance

Source

Derived from the FreedomIntelligence/HuatuoGPT-sft-data-v1 and HuatuoGPT2_sft_instruct_GPT4_50K datasets on Hugging Face.

Collection Method

Converted and merged from original sources into a ShareGPT (JSONL) format.

Freshness

Last updated on the platform in January 2024; underlying source data creation date is unspecified.