You need to agree to share your contact information to access this model

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

TrustMed Doctor-R1 baseline (8B) — step 21

Multi-turn GRPO (verl 0.5 + sglang) reimplementation of the Doctor-R1 recipe (ICLR 2026, arXiv 2510.04284) trained on our own consultation pool.

base model: Qwen/Qwen3-8B (no SFT stage; Doctor-R1's SFT data/weights are unreleased)
recipe: GRPO n=8, train batch 448, mini-batch 448, lr 1e-6, no KL, up to 10 doctor turns, response budget 3072 tokens, experience library on
reward: two-tier LLM judge (8-dimension consultation evaluator + terminal diagnosis judge), Qwen3-8B judge
selection: best held-out validation reward at training step 21 (val reward 2.3252)
action space: <think> ... </think> <answer> Question: ... | Recommendation: ... </answer>

Checkpoint exported with verl.model_merger from FSDP shards.

Safetensors

Model size

8B params

Tensor type

BF16

Inference Providers NEW

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Base model

Finetuned

Finetuned

this model

Quantizations