Fact-Qwen3-1.7B-Qwen3-8B-GRPO-Binary-RAR

Paper · GitHub · Project · Collection

This repository contains an on-policy distillation (OPD) student checkpoint released with Towards Understanding On-Policy Distillation through the Lens of Test-Time Scaling. Its base model, teacher, and training data are listed below.

The figure reports paper results across factual-QA settings; it is not a scorecard for this checkpoint alone.

Checkpoint details

• Base model: Qwen/Qwen3-1.7B
• Teacher model: chentong00/Qwen3-8B-GRPO-Binary-RAR
• Training data: A closed-book question mixture from Natural Questions (NQ) and HotpotQA.

This is a closed-book setup: this student has no retrieval or other external tool. Fact evaluation covers HotpotQA, TriviaQA, PopQA, and 2WikiMultiHopQA; its scorer expects one unique <answer>...</answer> response.

More information

See the paper and GitHub repository for training settings, evaluation protocols, and analysis. Upstream model and data terms remain applicable; see third-party notices.

Citation

@misc{ge2026understandingonpolicydistillationlens,
  title={Towards Understanding On-Policy Distillation through the Lens of Test-Time Scaling},
  author={Xinmu Ge and Zizhuo Zhang and Yu Huang and Jianing Zhu and Lin Yuan and Wanli Gu and Weichang Wu and Weiran Huang and Bo Han and Xiaolu Zhang and Jiangchao Yao},
  year={2026},
  eprint={2608.11829},
  archivePrefix={arXiv},
  primaryClass={cs.LG},
  doi={10.48550/arXiv.2608.11829},
  url={https://arxiv.org/abs/2608.11829}
}