Thinking-Space/Qwen3-4B-Base-GRPO

<h1 align="center">Qwen3-4B-Base-GRPO</h1>

<div align="center" style="line-height: 1;"> <a href="https://arxiv.org/abs/2604.13016" style="margin: 2px;"> <img alt="Paper" src="https://img.shields.io/badge/paper-A42C25?style=for-the-badge&logo=arxiv&logoColor=white" style="display: inline-block; vertical-align: middle;"/> </a> <a href="https://github.com/thunlp/OPD" style="margin: 2px;"> <img alt="Github" src="https://img.shields.io/badge/OPD-000000?style=for-the-badge&logo=github&logoColor=white" style="display: inline-block; vertical-align: middle;"/> </a> <a href="https://huggingface.co/papers/2604.13016" style="margin: 2px;"> <img alt="HF Papers" src="https://img.shields.io/badge/HF--Paper-%23FFD14D?style=for-the-badge&logo=huggingface&logoColor=black" style="display: inline-block; vertical-align: middle;"/> </a> <a href="https://x.com/HBX_hbx/status/2044464414829777354" style="margin: 2px;"> <img alt="Twitter" src="https://img.shields.io/badge/Twitter-%23000000.svg?style=for-the-badge&logo=x&logoColor=white" style="display: inline-block; vertical-align: middle;"/> </a> </div>

<br>

Qwen3-4B-Base-GRPO is a post-RL checkpoint trained with the verl framework. It starts from Qwen3-4B-Base and applies GRPO on the DAPO-Math-17k-Processed dataset 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 GRPO reinforcement learning to Qwen3-4B-Base with verl. The training is intended to improve math-focused reasoning performance under the on-policy distillation setting.

Key characteristics

• —Base model: Qwen3-4B-Base
• —Training framework: verl
• —Training stage: Reinforcement Learning (GRPO)
• —Parameter update: Full-parameter actor update
• —Primary domain: Mathematical reasoning
• —Reward model: Not used (reward_model.enable: false)
• —Rollout engine: vLLM
• —Context length: 32768 tokens
• —Responses per prompt: 8

Training Details

Training configuration

• —Framework: verl
• —Algorithm: grpo
• —GRPO outcome weight: 1.0
• —Learned reward model: disabled (reward_model.enable: false)
• —Reward source: custom rule-based math reward function
• —Training dataset: DAPO-Math-17k-Processed
• —Training file: datasets/DAPO-Math-17k-Processed/DAPO-Math.parquet
• —Validation datasets: AIME25, AMC23, AIME24
• —Prompt length: 1024
• —Response length: 7168
• —Validation response length: 31744
• —Max model length: 32768
• —Rollout temperature: 1.0
• —Repetition penalty: 1.0
• —KL loss: disabled
• —Format reward: disabled
• —Loss aggregation: token-mean
• —Learning rate: 1e-6
• —PPO mini-batch size: 64
• —PPO micro-batch size per GPU: 1
• —Tensor parallel size: 1
• —Number of GPUs: 8
• —Number of epochs: 1
• —Save frequency: every 20 steps
• —Test frequency: every 20 steps

Dataset

• —Training dataset: DAPO-Math-17k-Processed
• —Validation datasets: AIME25, AMC23, AIME24

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "lllyx/Qwen3-4B-Base-GRPO"

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}
}