---
license: mit
task_categories:
  - visual-question-answering
  - reinforcement-learning
language:
  - en
tags:
  - vlm
  - gymnasium
  - benchmark
  - multimodal
size_categories:
  - 1G<n<10G
---

# VLM-Gym Inference Dataset

This dataset contains pre-defined test episodes and initial states for evaluating Vision-Language Models (VLMs) on the VLM-Gym benchmark.

## Dataset Structure

```
inference-dataset/
├── test_set_easy/           # Easy difficulty test episodes (JSONL)
├── test_set_hard/           # Hard difficulty test episodes (JSONL)
├── initial_states_easy/     # Initial environment states for easy episodes (JSON)
├── initial_states_hard/     # Initial environment states for hard episodes (JSON)
└── partial_datasets/        # Assets required by some environments
    ├── objaverse/           # 3D models for mental rotation tasks
    ├── counting/            # Images for counting tasks
    ├── refcoco+/            # Images for referring expression tasks
    └── ...
```

## Tasks Included

| Task | Description |
|------|-------------|
| `maze_2d` | 2D maze navigation |
| `maze_3d` | 3D maze navigation |
| `mental_rotation_2d` | 2D shape rotation matching |
| `mental_rotation_3d_cube` | 3D cube rotation matching |
| `mental_rotation_3d_objaverse` | 3D object rotation matching |
| `jigsaw` | Jigsaw puzzle solving |
| `sliding_block` | Sliding block puzzle |
| `colorization` | Image colorization |
| `counting` | Object counting |
| `patch_reassembly` | Image patch reassembly |
| `matchstick_equation` | Matchstick equation solving |
| `matchstick_rotation` | Matchstick rotation |
| `video_unshuffle` | Video frame ordering |
| `zoom_in_puzzle` | Zoom-in puzzle solving |
| `fetch_reach` | Robotic reaching (easy only) |
| `fetch_pick_and_place` | Robotic manipulation (hard only) |
| `referring_dot_pointing` | Referring expression grounding (easy only) |

## Quick Start

### Installation

```bash
pip install huggingface_hub
```

### Download Full Dataset

```python
from huggingface_hub import snapshot_download

dataset_path = snapshot_download(
    repo_id="VisGym/inference-dataset",
    repo_type="dataset",
)
```

### Download Specific Subsets

```python
from huggingface_hub import snapshot_download

# Download only test sets (small, no large assets)
dataset_path = snapshot_download(
    repo_id="VisGym/inference-dataset",
    repo_type="dataset",
    allow_patterns=["test_set_easy/**", "test_set_hard/**"],
)

# Download only easy difficulty
dataset_path = snapshot_download(
    repo_id="VisGym/inference-dataset",
    repo_type="dataset",
    allow_patterns=["*_easy/**"],
)
```

### Using the Loader Script

```bash
# Download everything
python load_from_hf.py --output_dir ./inference_dataset

# Download only test sets (no large assets)
python load_from_hf.py --output_dir ./inference_dataset --subset test_sets

# Download only easy difficulty
python load_from_hf.py --output_dir ./inference_dataset --subset easy
```

## File Formats

### Test Set Files (JSONL)

Each line in the JSONL files contains an episode specification:

```json
{"seed": 1803372, "env_id": "maze_2d/hard", "episode_seed": 1052368083, "extra_state": null}
```

### Initial State Files (JSON)

JSON files containing the initial state for reproducible episode starts:

```json
{
  "object_path": "000-156/fa3dad5169784cec85b96682231e3f44.glb",
  "secret_yaw": 1.098,
  "secret_pitch": 0.487,
  ...
}
```

## Usage with VLM-Gym

```python
from pathlib import Path
import json

# Load test episodes
test_file = Path(dataset_path) / "test_set_easy" / "maze_2d__easy" / "*.jsonl"
for jsonl_file in test_file.parent.glob("*.jsonl"):
    with open(jsonl_file) as f:
        for line in f:
            episode = json.loads(line)
            env_id = episode["env_id"]
            seed = episode["seed"]
            episode_seed = episode["episode_seed"]
            # Use with VLM-Gym inference runner
```

## Citation

If you use this dataset, please cite:

```bibtex
@misc{wang2026visgymdiversecustomizablescalable,
      title={VisGym: Diverse, Customizable, Scalable Environments for Multimodal Agents}, 
      author={Zirui Wang and Junyi Zhang and Jiaxin Ge and Long Lian and Letian Fu and Lisa Dunlap and Ken Goldberg and XuDong Wang and Ion Stoica and David M. Chan and Sewon Min and Joseph E. Gonzalez},
      year={2026},
      eprint={2601.16973},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2601.16973}, 
}
```

## License

MIT License