CHIMERA-4B-RL

CHIMERA-4B-RL

CHIMERA-4B-RL is Qwen3-4B-Thinking-2507 post-trained on the CHIMERA dataset with supervised fine-tuning (SFT) followed by reinforcement learning (RL).

This model was introduced in the paper CHIMERA: Compact Synthetic Data for Generalizable LLM Reasoning.

Authors: Xinyu Zhu, Yihao Feng, Yanchao Sun, Xianzhi Du, Pingzhi Li, Olli Saarikivi, Yun Zhu, Yu Meng.

• Paper: arXiv:2603.00889
• Dataset: TianHongZXY/CHIMERA
• Models: CHIMERA-4B-SFT, CHIMERA-4B-RL, CHIMERA-Qwen3.5-4B-SFT

About CHIMERA

CHIMERA is a compact synthetic reasoning dataset of 9,225 problems spanning 8 scientific disciplines and 1,179 fine-grained topics, each paired with a long Chain-of-Thought (CoT) reasoning trajectory. All problems are generated, validated and solved by LLMs without human annotation. Despite its modest size, post-training Qwen3-4B-Thinking-2507 on CHIMERA approaches or matches the reasoning performance of substantially larger models such as DeepSeek-R1 and Qwen3-235B.

Results

All results use temperature = 0.6, top-p = 0.95, top-k = 20 and a maximum of 102,400 tokens. We sample 32 solutions per problem for AIME and HMMT, 10 for GPQA-Diamond, and 8 for Humanity's Last Exam.

Citation

@article{zhu2026chimera,
  title={CHIMERA: Compact Synthetic Data for Generalizable LLM Reasoning},
  author={Zhu, Xinyu and Feng, Yihao and Sun, Yanchao and Du, Xianzhi and Li, Pingzhi and Saarikivi, Olli and Zhu, Yun and Meng, Yu},
  journal={arXiv preprint arXiv:2603.00889},
  year={2026}
}