darrenxuu/TrustMed-DoctorR1-baseline-8B

TrustMed Doctor-R1 baseline (8B) Overview

This model, darrenxuu/TrustMed-DoctorR1-baseline-8B, is an 8 billion parameter language model designed for multi-turn medical consultations. It is a reimplementation of the Doctor-R1 recipe, utilizing a Multi-turn GRPO (Generative Reinforcement Learning with Policy Optimization) approach. The base model for this implementation is Qwen/Qwen3-8B, notably without an initial Supervised Fine-Tuning (SFT) stage, as the original Doctor-R1 SFT data is not publicly available.

Key Capabilities and Training

• Multi-turn Consultation: Specialized for engaging in extended doctor-patient dialogues, supporting up to 10 doctor turns.
• GRPO Training: Employs a GRPO recipe with specific parameters (n=8, train batch 448, mini-batch 448, lr 1e-6, no KL divergence, experience library enabled).
• Structured Response Generation: Generates responses in a defined action space, including <think> ... </think> for internal reasoning and <answer> Question: ... | Recommendation: ... </answer> for structured patient interaction.
• Two-tier LLM Judge Reward System: Utilizes a Qwen3-8B judge for reward, comprising an 8-dimension consultation evaluator and a terminal diagnosis judge, trained on a proprietary consultation pool.
• Context Length: Supports a context window of up to 32768 tokens, with a response budget of 3072 tokens.

Ideal Use Cases

This model is particularly suited for research and development in:

• Simulated Medical Consultations: Creating AI agents that can engage in realistic, multi-turn doctor-patient dialogues.
• Medical Dialogue Systems: Exploring reinforcement learning approaches for generating structured and contextually appropriate medical advice or questions.
• Evaluating Consultation Quality: As a baseline for developing and testing new methods in medical AI, especially those involving complex interaction patterns and structured outputs.