The first open "medical o1" — complex reasoning for medicine via a two-stage recipe: construct verifiable medical problems from exam questions, learn complex reasoning trajectories through a medical verifier, then RL against verifier rewards. With only 40K verifiable problems it outperformed general and prior medical baselines, and the verifiable-medical-problems + verifier-RL recipe became the heavily cited template for domain reasoning models.

From Benyou Wang's FreedomIntelligence group (CUHK-Shenzhen). Models are RL-tuned open bases (Llama-3.1 8B/70B and Qwen2.5 7B/72B), not from-scratch pretrains. The line continues in HuatuoGPT-3.

Model Details

Base model llama-3.1

Variants

Name Parameters Notes
HuatuoGPT-o1-8B RL-tuned from Llama-3.1-8B
HuatuoGPT-o1-70B RL-tuned from Llama-3.1-70B
HuatuoGPT-o1-7B RL-tuned from Qwen2.5-7B
HuatuoGPT-o1-72B RL-tuned from Qwen2.5-72B

Paper

medicalreasoningpost-trainingopen-weight

Related