Qwen3-1.7B-Base-OPD/README.md

license, language, datasets, library_name, pipeline_tag, tags, base_model, base_model_relation

license	language	datasets	library_name	pipeline_tag	tags	base_model	base_model_relation
other	en zh	BytedTsinghua-SIA/DAPO-Math-17k	transformers	text-generation	qwen qwen3 math dapo opd on-policy-distillation knowledge-distillation reasoning safetensors arxiv:2604.13016	Qwen/Qwen3-1.7B-Base	finetune

Qwen3-1.7B-Base-OPD

Qwen3-1.7B-Base-OPD is an on-policy distillation (OPD) checkpoint initialized from Qwen3-1.7B-Base. It is distilled from the teacher model Qwen3-4B-Base-GRPO using the DAPO-Math-17k dataset, and is intended for mathematical reasoning and problem-solving.

This model is associated with the paper:
Rethinking On-Policy Distillation of Large Language Models: Phenomenology, Mechanism, and Recipe
Paper link: https://arxiv.org/abs/2604.13016

Model Description

This model is obtained by applying on-policy distillation (OPD) to Qwen3-1.7B-Base, with Qwen3-4B-Base-GRPO serving as the teacher model. The OPD training uses DAPO math prompts/data and is designed to transfer the teacher's math-focused reasoning behavior into a smaller 1.7B-parameter student model.

Key characteristics

• Student/base model: Qwen3-1.7B-Base
• Teacher model: lllyx/Qwen3-4B-Base-GRPO
• Training data: DAPO-Math-17k
• Training stage: On-Policy Distillation (OPD)
• Training framework: verl
• Rollout engine: vLLM
• Primary domain: Mathematical reasoning
• Model architecture: Qwen3ForCausalLM
• Precision: bfloat16
• Context length: 32768 tokens

Training Details

Training configuration

• Base checkpoint: Qwen/Qwen3-1.7B-Base
• Teacher checkpoint: lllyx/Qwen3-4B-Base-GRPO
• Training framework: verl
• Training method: on-policy distillation with GRPO-style rollouts
• Distillation loss mode: k1
• Policy-gradient term: enabled
• Training dataset: DAPO-Math-17k/DAPO-Math.parquet
• Primary task domain: math reasoning
• Chat template thinking mode: disabled (enable_thinking=False)
• Model type: qwen3

Rollout and optimization

• Rollout engine: vLLM
• Responses per prompt: 4
• Prompt length: 1024
• Response length: 7168
• Max rollout model length: 8193
• Train batch size: 64
• PPO mini-batch size: 16
• PPO micro-batch size per GPU: 1
• Max PPO token length per GPU: 8192
• Actor learning rate: 1e-6
• Total epochs: 1
• Save frequency: every 20 steps

Runtime setup

• Distributed backend: Ray
• Number of nodes: 1
• GPUs per node: 4
• Teacher world size: 4
• Rollout tensor parallel size: 1
• Teacher tensor parallel size: 1
• Actor training: FSDP with parameter and optimizer offload
• Gradient checkpointing: enabled
• Padding removal: enabled
• Torch compile for actor: enabled
• Reward function: rule-based math reward from verl/recipe/r1_ascend/deepscaler.py::compute_score

Dataset

• Training data: BytedTsinghua-SIA/DAPO-Math-17k
• Teacher rollout/source model: lllyx/Qwen3-4B-Base-GRPO
• Student initialization: Qwen/Qwen3-1.7B-Base

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "lllyx/Qwen3-1.7B-Base-OPD"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype="auto",
    device_map="auto",
)

Citation

If you use this model, please consider citing the related paper:

@article{li2026rethinking,
  title={Rethinking On-Policy Distillation of Large Language Models: Phenomenology, Mechanism, and Recipe},
  author={Li, Yaxuan and Zuo, Yuxin and He, Bingxiang and Zhang, Jinqian and Xiao, Chaojun and Qian, Cheng and Yu, Tianyu and Gao, Huan-ang and Yang, Wenkai and Liu, Zhiyuan and Ding, Ning},
  journal={arXiv preprint arXiv:2604.13016},
  year={2026}
}