Qwen3-1.7B Static-R0 Policy Checkpoint

This public model repository contains one policy checkpoint from the static-rubric discriminability-horizon experiment. The domain and optimizer step are encoded in the repository name.

Experiment

• Base policy: Qwen/Qwen3-1.7B
• Base revision: 70d244cc86ccca08cf5af4e1e306ecf908b1ad5e
• RL algorithm: GRPO
• Training reward: frozen prompt-specific initial rubric (R0)
• Domain: RaR Medicine or RaR Science, as named in the repository
• Training seed: 11
• Export format: Hugging Face Transformers, BF16 safetensors
• Checkpoint contents: model weights, configuration, tokenizer, and chat template
• Excluded: optimizer, scheduler, trainer state, rollouts, rubrics, and evaluation data

Medicine checkpoint repositories cover the ten planned audit points through step 48. Science training was only saved through step 3 at the time of this upload, so the public Science repositories contain steps 0 and 3.

Loading

from transformers import AutoModelForCausalLM, AutoTokenizer

repo_id = "HYU-NLP-EVAL/<repository-name>"
tokenizer = AutoTokenizer.from_pretrained(repo_id)
model = AutoModelForCausalLM.from_pretrained(
    repo_id,
    dtype="bfloat16",
    device_map="auto",
)

Intended use

These checkpoints are research artifacts for studying reward saturation and static-rubric staleness during policy optimization. Medicine checkpoints are not medical devices and must not be used as a substitute for professional medical advice.