hhyhuang/deeprefine repository overview

⁠News

• [2026/6/2] ✨DeepRefine-Skill⁠ has released! An agent skill to evolve the quality of LLM-Wiki (Graphify) at test time with the harness of DeepRefine.
• [2026/5/10] Static quants of DeepRefine-v1-8B, 🤗 mradermacher/DeepRefine-v1-8B-GGUF⁠ has been released. Thanks to the community!

⁠🪜 Environment

We provide a Dockerfile⁠ for a reproducible environment.

⁠Local setup

Create atlastune from scratch:

cd /path/to/DeepRefine
conda env create -f docker/atlastune_environment.yml
conda activate atlastune
pip install -e .

Create the environment with pip only:

conda create -n atlastune python=3.10 -y
conda activate atlastune
cd /path/to/DeepRefine

pip install torch==2.8.0 torchvision==0.23.0 torchaudio==2.8.0 \
  --index-url https://download.pytorch.org/whl/cu126
pip install -r docker/requirements-atlastune.txt
pip install flash-attn==2.8.3 --no-build-isolation
pip install -e .

pip install -e . registers the bundled verl package. For local atlas-rag development, use pip install -e ./AutoSchemaKG instead of the PyPI pin when needed.

⁠graphify + Cursor skill (DeepRefine-Skill)

The /deeprefine Cursor skill and deeprefine CLI live in a separate repo (DeepRefine-Skill⁠), sibling to this one. After atlastune is ready:

pip install -e /path/to/DeepRefine-Skill
cd /path/to/your-kb-project && deeprefine cursor install

See DeepRefine-Skill/README.md⁠ for the full graphify workflow.

⁠🚀 Quick Start Demo

We have a demo pipline test.ipynb⁠, in which you can have a quick overview about what DeepRefine is doing.

⁠📊 Training Data Preprocessing

We collect the raw training data of HotpotQA from https://hotpotqa.github.io/⁠ and then construct the data samples for RL training through the following script:

bash scripts/autograph-r1/data_prepare/hotpotqa_cons.sh

Or you can also access the training data under the folder data/.

⁠🔥 GRPO Training

Update your config in verl/third_party/autograph_r1/config.ini.

Train $\texttt{DeepRefine-4B}$ based on $\texttt{Qwen3-4B-Instruct-2507en}$ with GRPO and GBD reward:

bash scripts/train/run_qwen3-4b_graph_refiner.sh

Train $\texttt{DeepRefine-8B}$ based on $\texttt{Qwen3-8B}$ with GRPO and GBD reward:

bash scripts/train/run_qwen3-8b_graph_refiner.sh

We have also provided our model in HuggingFace⁠.

⁠🔍 Evaluation

There are six evaluation mode:

• Graph Retriever, no refinement.

bash scripts/eval/gr_refine_bench_no_refine.sh

• Graph Retriever, naive refinement (without training).

bash scripts/eval/gr_refine_bench_wo_rl.sh

• Graph Retriever, deeprefine.

bash scripts/eval/gr_refine_bench_rl.sh

• Text Retriever, no refinement.

bash scripts/eval/tr_refine_bench_no_refine.sh

• Text Retriever, naive refinement (without training).

bash scripts/eval/tr_refine_bench_wo_rl.sh

• Text Retriever, deeprefine.

bash scripts/eval/tr_refine_bench_rl.sh

⁠📖 Citation

@article{huang2026deeprefine,
  title={DeepRefine: Agent-Compiled Knowledge Refinement via Reinforcement Learning},
  author={Huang, Haoyu and Bai, Jiaxin and Liu, Shujie and Wei, Yang and Tsang, Hong Ting and Gao, Yisen and Xie, Zhongwei and Li, Yufei and Song, Yangqiu},
  journal={arXiv preprint arXiv:2605.10488},
  year={2026}
}