Medical Reasoning Sft Mega

1,789,998 unique medical samples containing 3.78 billion tokens of chain-of-thought reasoning, released by OpenMed in February 2026. The dataset aggregates outputs from seven state-of-the-art AI models using fair distribution deduplication to ensure sample variety. It specifically focuses on clinical logic, with 1.56 billion tokens dedicated exclusively to reasoning steps.

Use Cases

• Supervised Fine-Tuning (SFT) of LLMs using the 1.56 billion reasoning tokens to improve clinical logic
• Evaluating medical question-answering performance across 1.79 million unique scenarios
• Distilling medical knowledge from 7 state-of-the-art models into smaller specialized architectures

Strengths

• 1,789,998 unique samples after deduplication
• 3.78 billion total tokens including 1.56 billion reasoning tokens
• 100% reasoning coverage across 1,789,764 samples

Limitations

• Synthetic nature as data is generated by 7 AI models rather than human clinicians
• Potential for model-induced hallucinations in the 1.56 billion reasoning tokens

Provenance

Source

OpenMed (aggregated from 7 AI models)

Collection Method

Synthetic generation and deduplication

Freshness

Last updated February 2026.

Geography

US