wan288972153/VisualToolBench-XSkill

VisualToolBench (XSkill-formatted)

This is VisualToolBench (originally from ScaleAI/VisualToolBench) repackaged into the JSON layout consumed by XSkill's evaluation pipeline (eval/infer_api.py). All images are extracted from the original parquet snapshots and stored as PNG files referenced by relative paths in the JSON splits.

Contents

A single archive VisualToolBench.zip (~3.7 GB) containing:

VisualToolBench/
├── val_50.json        #   50 sanity-check samples
├── val_single.json    # ~603 single-turn samples
├── val_multi.json     # ~601 multi-turn samples
├── val_full.json      # 1204 full samples (single + multi)
└── images/
    └── <doc_id>/img_*.png    # one folder per sample

Each sample in the JSON files follows XSkill's expected schema:

{
  "doc_id":      "<unique-id>",
  "problem":     "<image>\n<question text>",
  "images":      ["VisualToolBench/images/<doc_id>/img_0.png", ...],
  "solution":    "<gold answer>",
  "data_source": "<original prompt category>"
}

The <image> placeholder marks where each image is to be injected during prompt assembly; the order matches the order of paths in images.

Usage

Download

huggingface-cli download wan288972153/VisualToolBench-XSkill \
    VisualToolBench.zip \
    --repo-type dataset \
    --local-dir .
unzip VisualToolBench.zip -d ./
# → ./VisualToolBench/  (contains the JSON splits + images/)

Plug into XSkill

Either drop the VisualToolBench/ folder under <XSkill>/benchmark/, or point the ablation script to wherever you put it:

VTB_DATA_DIR=/path/to/VisualToolBench bash scripts_local/run_ablation.sh

Source

• Original benchmark: ScaleAI/VisualToolBench
• Conversion script: see scripts_local/convert_visualtoolbench.py in the XSkill repo

Citation

If you use this data, please cite the original VisualToolBench authors and the XSkill paper:

@misc{jiang2026xskillcontinuallearningexperience,
  title  = {XSkill: Continual Learning from Experience and Skills in Multimodal Agents},
  author = {Guanyu Jiang and Zhaochen Su and Xiaoye Qu and Yi R. Fung},
  year   = {2026},
  eprint = {2603.12056},
  archivePrefix = {arXiv},
  primaryClass  = {cs.AI}
}