Files

ModelHub XC 82107c6153 初始化项目，由ModelHub XC社区提供模型

Model: youngzhong/SOD-0.6B
Source: Original Platform

2026-07-22 20:34:10 +08:00

3.8 KiB

Raw Blame History

language, license, library_name, base_model, tags, pipeline_tag

language

license

library_name

base_model

SOD-0.6B

About

SOD-0.6B is a 0.6B student model distilled from a 4B teacher using SOD (Step-wise On-policy Distillation), a method designed for training small language model agents with tool-integrated reasoning capabilities.

SOD addresses the cascading error propagation problem in on-policy distillation for agentic reasoning by introducing an adaptive step-level weighting mechanism that suppresses distillation loss on drifted steps and restores supervision when the student recovers alignment — all at negligible additional computational cost.

Model Information

Attribute	Value
Base Model	Qwen3-0.6B
Teacher Model	SOD-GRPO_teacher-4B
Training Pipeline	Cold-Start SFT → SOD (Step-wise On-policy Distillation)
Parameters	0.6B

Model	Description
SOD-0.6B	SOD-distilled 0.6B student (this model)
SOD-1.7B	SOD-distilled 1.7B student
SOD-GRPO_teacher-4B	GRPO-trained 4B teacher model

Performance

We report average@32 over 5 runs on challenging math, science, and code benchmarks.

0.6B Student Results

Method	AIME 2024	AIME 2025	GPQA-Diamond	LiveCodeBench-v6	Average
Vanilla	7.71	12.81	13.24	14.89	12.16
SFT	5.67	5.42	15.20	9.61	8.97
GRPO	4.06	4.90	20.38	15.95	11.32
OPD	16.82	22.95	17.76	22.65	20.04
OPSD_gt	12.63	17.04	17.32	16.73	15.93
OPSD_hint	9.77	14.12	15.98	12.65	13.13
SOD (This Model)	20.84	26.13	22.19	27.72	24.22

Teacher Model (4B)

Method	AIME 2024	AIME 2025	GPQA-Diamond	LiveCodeBench-v6	Average
GRPO	67.60	60.42	55.19	63.13	61.59

Key Highlights

🏆 Strong 0.6B agent: Achieves 26.13% on AIME 2025 (average@32)
📈 +20.86% over second-best baseline (OPD) on average
💡 Minimal extra compute: The divergence metric reuses log-probabilities already computed in the forward pass

Citation

@article{zhong2026sod,
  title={SOD: Step-wise On-policy Distillation for Small Language Model Agents},
  author={Zhong, Qiyong and Zheng, Mao and Song, Mingyang and Lin, Xin and Sun, Jie and Jiang, Houcheng and Wang, Xiang and Fang, Junfeng},
  journal={arXiv preprint arXiv:2605.07725},
  year={2026}
}

3.8 KiB Raw Blame History