Static-R0 Matched GRPO on RaR-Medicine — step 33

This is the policy after 33 global optimizer updates of the matched static-rubric GRPO run (planned total: 48). It is intentionally separate from the OnlineRubrics/dynamic-rubric checkpoints.

Experiment identity

• Method: static_r0_matched
• Reward source: rar_static_r0_only
• Domain: Medicine
• Training data: RaR-Medicine, 1,500 prompts
• Seed: 11
• Policy: Qwen/Qwen3-4B-Instruct-2507
• Base revision: cdbee75f17c01a7cc42f958dc650907174af0554
• Thinking: disabled
• GRPO global prompt batch: 96
• Rollouts per prompt: 16
• Learning rate: 5e-06

The root files are a BF16 Transformers export for inference. The original_checkpoint/ directory contains the exact original veRL/FSDP policy parameter checkpoint and its tokenizer/configuration files. Optimizer, trainer, and data-loader state are intentionally not published; the complete resume checkpoint remains on Daisy.

from transformers import AutoModelForCausalLM, AutoTokenizer

repo_id = "HYU-NLP-EVAL/qwen3-4b-rar-medicine-static-r0-matched-seed11-step-033"
tokenizer = AutoTokenizer.from_pretrained(repo_id)
model = AutoModelForCausalLM.from_pretrained(
    repo_id, torch_dtype="bfloat16", device_map="auto"
)

This is an intermediate research checkpoint, not a clinical model. No medical capability or safety claim is made.

Original actor parameter SHA256: 5a78fc291520c5f4651a9e4dbac6753940e3ab6bde9d7a4b29008546b0dd7f9f