README.md

---
language:
- en
license: mit
size_categories:
- n<1K
task_categories:
- image-text-to-text
pretty_name: AgentViSS
tags:
- multimodal
- social-simulation
- visual-social-intelligence
- multi-agent-simulation
configs:
- config_name: default
  data_files:
  - split: train
    path: train.jsonl
---

# AgentViSS Dataset

This repository contains the public data package for **AgentViSS: Can Agents
Read the Room? Benchmarking Visual Social Intelligence in Multimodal
Simulation**.

[**Paper**](https://huggingface.co/papers/2606.15152) | [**GitHub**](https://github.com/JunsWan/AgentViSS)

## Files

```text
.
+-- README.md
+-- train.jsonl
+-- data.json
+-- images/
    +-- scenario_001/
    |   +-- group_labeled.png
    |   +-- individual/
    |       +-- character_01.jpg
    |       +-- character_02.jpg
    +-- ...
```

- `train.jsonl` is a viewer-friendly table with 240 AgentViSS records. It
  exposes key fields from `data.json` and keeps image path columns at the end.
- `data.json` contains the canonical full records.
- `images/` contains all image assets referenced by `data.json`.
- Image paths in `data.json` are package-relative. Resolve them from the
  dataset repository root.
- All images are resized proportionally so that the longest edge is at most
  512 pixels.

## Dataset Structure

Each record includes an anonymized scenario id, dialogue type, conflict level,
characters, role-level information, goals, emotions, and image paths.

The Hugging Face Dataset Viewer is configured to read `train.jsonl` rather than
the raw `images/` directory. This keeps the viewer organized around the 240
scenario records instead of displaying the 282 image files as separate rows.

Important image fields:

- `group_image_labeled`: path to the scenario-level group image.
- `individual_images`: mapping from character names to role portrait paths.

Example:

```json
{
  "id": "agentviss_low_01",
  "source_scenario": "scenario_001",
  "group_image_labeled": "images/scenario_001/group_labeled.png",
  "individual_images": {
    "CharacterName": "images/scenario_001/individual/character_01.jpg"
  }
}
```

The fields `base_json_path`, `group_image`, `background`, and
`individual_informations.*.personality` are not included in this public data
package.

## Download

Download the full dataset, including images, with `huggingface_hub`:

```python
import json
from pathlib import Path
from huggingface_hub import snapshot_download

root = Path(snapshot_download(
    repo_id="JunsWan/AgentViSS",
    repo_type="dataset",
))

records = json.loads((root / "data.json").read_text(encoding="utf-8"))
first_group_image = root / records[0]["group_image_labeled"]
```

For the flattened viewer table:

```python
import json
from pathlib import Path
from huggingface_hub import snapshot_download

root = Path(snapshot_download(
    repo_id="JunsWan/AgentViSS",
    repo_type="dataset",
))

with (root / "train.jsonl").open(encoding="utf-8") as f:
    rows = [json.loads(line) for line in f]
```

You can also clone the repository:

```bash
git clone https://huggingface.co/datasets/JunsWan/AgentViSS
```

## Citation

```bibtex
@misc{agentviss2026,
  title = {Can Agents Read the Room? Benchmarking Visual Social Intelligence in Multimodal Simulation},
  author = {Anonymous Authors},
  year = {2026}
}
```