Critique-Coder

| 🚀Project Page | 📖Paper | 🤗Data | 🤗Model |

Overview

Get Started

Installation

git clone https://github.com/TIGER-AI-Lab/Critique-Coder.git
cd Critique-Coder
conda create --name critique-coder python=3.10
conda activate critique-coder
pip install -e verl
pip install -e ".[vllm,acecoder]"
pip install "flash-attn<2.8.0" --no-build-isolation

Dataset Preparation

We did extensive filtering for rStar-Coder to build rStar-Coder-Critique data:

hf download TIGER-Lab/rStar-Critique-Data --local-dir ./data/critique-coder-dataset --repo dataset

Training

bash examples/train/train_qwen3_4b_16k.sh

Evaluation

We use EvalPlus, BigCodeBench, Aider-Polyglot, LiveCodeBench for the evaluation of EvalPlus, BigCodeBench, Aider-Polyglot, LiveCodeBench, respectively.

Citations

If you find this work helpful, please cite us with:

@article{ruan2025critiquecoder,
    title={Critique-Coder: Enhancing Coder Models by Critique Reinforcement Learning},
    author={Ruan, Chi and Jiang, Dongfu and Wang, Yubo and Chen, Wenhu},
    journal={ArXiv},
    year={2025},
    volume={2509.22824}
}