Zoe-Mango/Vis-Skill-Dataset

Procedural Exposure Bias in Skill-Augmented Language Model Agents

🤗 Environment data: ALFWorld & WebShop

This dataset holds the ALFWorld and WebShop files read by the code of our paper, so you don't need to fetch them from the original sources or rebuild the WebShop search index. It contains no new annotations or model outputs.

🚀 Usage

git clone https://github.com/NEUIR/Vis-Skill.git && cd Vis-Skill
hf download Zoe-Mango/Vis-Skill-Dataset --repo-type dataset --local-dir data   # `hf` comes with: pip install -U huggingface_hub
bash scripts/setup_data.sh      # checks the files and links the WebShop data into experiments/webshop/

Then follow the GitHub README to run the experiments. ScienceWorld is not included: it is installed with pip install scienceworld (official).

📦 Included Files

• alfworld/ (76 MB): the valid_unseen split (= eval_out_of_distribution, 134 evaluated games) and logic/, taken unchanged from the official ALFWorld release (alfworld 0.4.2).
• webshop/data/: the full WebShop product files (items_shuffle.json, items_ins_v2.json, items_human_ins.json; 1,181,436 product records) from the WebShop authors, plus the two prompt files from SkillNet.
• webshop/search_engine/indexes/ (3.3 GB): the prebuilt Lucene index loaded by the WebShop simulator. It has 1,181,370 documents, because the simulator drops 6 records whose ASIN is longer than 10 characters and 60 products have no text to index.
• Reference copies, not read by the code: webshop/web_agent_site/, webshop/test_indices.json (the 200 evaluated sessions) and the 1,000-product development index (indexes_1k/, resources_1k/). The code uses its own copies in the GitHub repository.

⚖️ License

This dataset is released under the CC BY 4.0 license. The files come from MIT-licensed projects; please also cite them: ALFWorld (Shridhar et al., ICLR 2021), WebShop (Yao et al., NeurIPS 2022; product data scraped from amazon.com by the WebShop authors) and SkillNet.

📝 Citation

@misc{yang2026proceduralexposurebiasskillaugmented,
      title={Procedural Exposure Bias in Skill-Augmented Language Model Agents}, 
      author={Qing Yang and Yuqi Xiong and Pengcheng Huang and Zhipeng Xu and Zhenghao Liu and Zulong Chen and Yu Gu and Ge Yu},
      year={2026},
      eprint={2610.26062},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2610.26062}, 
}

📬 Contact

Please open an issue on GitHub or email yangqing_neu@outlook.com.