SOD-1.7B/README.md

language, license, library_name, base_model, tags, pipeline_tag

language	license	library_name	base_model	tags	pipeline_tag
en	apache-2.0	transformers	Qwen/Qwen3-1.7B	agent tool-use distillation math code reasoning	text-generation

SOD-1.7B

About

SOD-1.7B is a 1.7B 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	Qwen3-1.7B
Teacher Model	SOD-GRPO_teacher-4B
Training Pipeline	Cold-Start SFT → SOD (Step-wise On-policy Distillation)
Parameters	1.7B

Related Models

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

Performance

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

1.7B Student Results

Method	AIME 2024	AIME 2025	GPQA-Diamond	LiveCodeBench-v6	Average
Vanilla	9.90	8.96	26.80	22.73	17.10
SFT	26.77	22.40	29.85	24.63	25.91
GRPO	25.63	21.67	33.55	20.70	25.39
OPD	43.86	37.04	31.73	32.45	36.27
OPSD_gt	33.85	24.69	35.02	22.73	29.07
OPSD_hint	34.42	21.43	33.46	23.12	28.11
SOD (This Model)	50.83	41.72	38.72	40.63	42.98

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

• 🏆 Recovers 69.8% of teacher performance with only 1.7B parameters (42.98 vs 61.59)
• 📈 +18.5% 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}
}