qwen2.5-1.5B-longcot-reasoning-HPD/README.md

license, library_name, pipeline_tag

license	library_name	pipeline_tag
apache-2.0	transformers	text-generation

Hybrid Policy Distillation: Qwen2.5-1.5B Student

This repository contains a Qwen2.5-1.5B student model distilled from Qwen2.5-7B-Thinking using Hybrid Policy Distillation (HPD), as presented in the paper Hybrid Policy Distillation for LLMs.

Overview

Knowledge distillation (KD) is a powerful paradigm for compressing large language models (LLMs). Hybrid Policy Distillation (HPD) is a framework designed to make policy distillation more stable and efficient for reasoning-oriented models. It integrates the complementary advantages of forward and reverse KL to balance mode coverage and mode-seeking, and combines off-policy data with lightweight, approximate on-policy sampling.

• Paper: Hybrid Policy Distillation for LLMs
• Repository: zwhong714/Hybrid-Policy-Distillation

Benchmark Performance

The following table shows the performance of the distilled student model compared to the teacher model across various reasoning benchmarks:

Model	AIME24	AIME25	AMC	MATH	OlympiadMath	GPQA
Qwen2.5-7B-Thinking (Teacher)	28.13	27.19	71.72	87.48	58.50	43.43
Qwen2.5-1.5B-Thinking (Student)	7.71	9.89	39.84	63.40	32.53	28.09

Citation

If you find this model or the HPD framework useful in your research, please cite the following work:

@article{hong2024hybrid,
  title={Hybrid Policy Distillation for LLMs},
  author={Hong, Zhang-Wei and others},
  journal={arXiv preprint arXiv:2604.20244},
  year={2024}
}