Graph of Skills (GoS) Data

This repository contains the dataset artifacts for the paper Graph of Skills: Dependency-Aware Structural Retrieval for Massive Agent Skills.

Graph of Skills (GoS) is an inference-time structural retrieval layer for large skill libraries. It constructs an executable skill graph offline and retrieves a bounded, dependency-aware skill bundle at inference time. This repository hosts the skill libraries and prebuilt workspaces used in the paper's benchmarks (SkillsBench and ALFWorld).

Links

• Paper: https://huggingface.co/papers/2604.05333
• GitHub Repository: https://github.com/davidliuk/graph-of-skills

Dataset Structure

The data includes:

• Skill Libraries: Collections of SKILL.md files (sizes: 200, 500, 1000, and 2000 skills).
• Prebuilt Workspaces: Indexed graph workspaces containing vector embeddings and graph storage for immediate use with the GoS engine.

Sample Usage

To download the skill libraries and workspaces using the utility script provided in the GitHub repository:

# Download the skill sets (skills_200, skills_500, etc.)
./scripts/download_data.sh --skillsets

# Download the prebuilt graph workspaces
./scripts/download_data.sh --workspace

For more information on how to index custom skills or run the evaluation suite, please refer to the GitHub README.

Citation

@misc{li2026graphskillsdependencyawarestructural,
      title={Graph of Skills: Dependency-Aware Structural Retrieval for Massive Agent Skills}, 
      author={Dawei Liu and Zongxia Li and Hongyang Du and Xiyang Wu and Shihang Gui and Yongbei Kuang and Lichao Sun},
      year={2026},
      eprint={2604.05333},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2604.05333}, 
}