README.md

---
dataset_info:
  features:
    - name: id
      dtype: string
    - name: game
      dtype: string
    - name: trial_id
      dtype: int32
    - name: episode_id
      dtype: int32
    - name: frame_idx
      dtype: int32
    - name: action
      dtype: string
    - name: action_int
      dtype: int32
    - name: score
      dtype: int32
    - name: reward
      dtype: int32
    - name: reaction_time_ms
      dtype: int32
    - name: gaze_positions
      dtype: string
    - name: image_bytes
      dtype: binary
license: mit
task_categories:
  - robotics
  - reinforcement-learning
tags:
  - atari
  - vla
  - vision-language-action
  - imitation-learning
  - human-demonstrations
size_categories:
  - 1M<n<10M
---

# TESS-Atari Stage 1 (5Hz)

Human gameplay demonstrations from Atari games, formatted for Vision-Language-Action (VLA) model training.

## Overview

| Metric | Value |
|--------|-------|
| Source | [Atari-HEAD](https://zenodo.org/records/3451402) |
| Games | 11 (overlapping with DIAMOND benchmark) |
| Samples | ~4M |
| Action Rate | 5 Hz (1 action per observation) |
| Format | Lumine-style action tokens |

## Games Included

Alien, Asterix, BankHeist, Breakout, DemonAttack, Freeway, Frostbite, Hero, MsPacman, RoadRunner, Seaquest

## Action Format

```
<|action_start|> FIRE <|action_end|>
<|action_start|> LEFT <|action_end|>
<|action_start|> RIGHTFIRE <|action_end|>
```

## Schema

| Field | Type | Description |
|-------|------|-------------|
| `id` | string | Unique sample ID: `{game}_{trial}_{frame}` |
| `game` | string | Game name (lowercase) |
| `trial_id` | int | Human player trial number |
| `episode_id` | int | Episode within trial (-1 if unknown) |
| `frame_idx` | int | Frame sequence number |
| `action` | string | Lumine-style action token |
| `action_int` | int | Raw ALE action code (0-17) |
| `score` | int | Current game score |
| `reward` | int | Immediate reward |
| `reaction_time_ms` | int | Human decision time in ms |
| `gaze_positions` | string | Eye tracking data (x,y pairs) |
| `image_bytes` | bytes | PNG image of game frame |

## Usage

```python
from datasets import load_dataset

ds = load_dataset("TESS-Computer/atari-vla-stage1-5hz")

# Get a sample
sample = ds["train"][0]
print(sample["action"])  # <|action_start|> FIRE <|action_end|>

# Decode image
from PIL import Image
from io import BytesIO
img = Image.open(BytesIO(sample["image_bytes"]))
```

## Evaluation

Designed for evaluation in [DIAMOND](https://diamond-wm.github.io/) world models on the Atari 100k benchmark.

## Related

- [15Hz variant](https://huggingface.co/datasets/TESS-Computer/atari-vla-stage1-15hz) - 3 actions per observation for faster gameplay
- [Lumine AI](https://www.lumine-ai.org/) - Inspiration for VLA architecture
- [DIAMOND](https://diamond-wm.github.io/) - World model for evaluation

## Citation

```bibtex
@misc{atarihead2019,
  title={Atari-HEAD: Atari Human Eye-Tracking and Demonstration Dataset},
  author={Zhang, Ruohan and others},
  year={2019},
  url={https://zenodo.org/records/3451402}
}
```