You need to agree to share your Hugging Face information to access this dataset

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

YAML Metadata Warning:empty or missing yaml metadata in repo card

Check out the documentation for more information.

rare_sync: Qwen3-8B trained on the synthetic rare-disease package

Results, analysis and caveats: see SUMMARY.md. Run log: AGENT_LOG.md. Raw eval results: results.jsonl, eval//.

models/sft-qwen3-8b/ LoRA SFT on cot_sft (adapter/ on Qwen/Qwen3-8B, merged/ bf16, done.json = args + log)
models/grpo-qwen3-8b/ GRPO from Qwen/Qwen3-8B (adapter/ on Qwen/Qwen3-8B, merged/, done.json)
models/sftgrpo-qwen3-8b/ GRPO from the SFT model (adapter/ on sft-qwen3-8b/merged, NOT on Qwen3-8B; merged/, done.json)
eval// generations.jsonl + summary.json; -mt2048 = eval max_tokens 2048 (main), plain = 1024
flow2/ run pack (PBS, Flow 2.0); see flow2/RUN.md. bash submit.sh e12 / bash submit.sh e2 reproduce both GRPO runs.

vLLM / transformers: load models//merged directly.
sftgrpo adapter: merge sft-qwen3-8b/adapter into Qwen3-8B first (or use sft-qwen3-8b/merged), then apply sftgrpo-qwen3-8b/adapter.
Prompt format: Qwen3 chat template, non-thinking mode, empty removed and "\n" prefilled (flow2/scripts/scaffold.py build_prompt). Answers are scored inside ....