TESS-Computer/csgo-vla-stage1-16hz

---
license: mit
task_categories:
- robotics
- image-to-text
tags:
- VLA
- gaming
- counter-strike
- behavioral-cloning
- imitation-learning
size_categories:
- 1M<n<10M
---

# CS:GO VLA Stage 1 Dataset (16Hz)

Vision-Language-Action dataset for Counter-Strike: Global Offensive, converted from the [TeaPearce CS:GO dataset](https://huggingface.co/datasets/TeaPearce/CounterStrike_Deathmatch).

## Overview

- **Frame rate:** 16Hz (native, 1 action per frame)
- **Total samples:** ~5.5M frames
- **Split:** train (~5M) / test (~500K) following [Diamond](https://github.com/eloialonso/diamond) split
- **Map:** Dust2 deathmatch

## Action Format

```
<|action_start|> mouse_x mouse_y [keys] <|action_end|>
```

**Examples:**
```
<|action_start|> 0 0 <|action_end|> # idle
<|action_start|> 5 0 W <|action_end|> # walking forward
<|action_start|> -200 50 W A L <|action_end|> # strafing + shooting
```

## Schema

| Column | Type | Description |
|--------|------|-------------|
| `id` | string | Unique sample ID |
| `episode_id` | string | Source HDF5 file |
| `frame_idx` | int32 | Frame number (0-999) |
| `action` | string | Text-formatted action |
| `kill_flag` | int32 | 1 if player got a kill |
| `death_flag` | int32 | 1 if player died |
| `split` | string | "train" or "test" |
| `image_bytes` | bytes | JPEG screenshot |

## Usage

```python
from datasets import load_dataset

# Load full dataset
ds = load_dataset("TESS-Computer/csgo-vla-stage1-16hz")

# Filter by split
train_ds = ds.filter(lambda x: x['split'] == 'train')
test_ds = ds.filter(lambda x: x['split'] == 'test')
```

## Related

- [5Hz chunked variant](https://huggingface.co/datasets/TESS-Computer/csgo-vla-stage1-5hz) - 3 actions per sample
- [Diamond World Model](https://github.com/eloialonso/diamond) - For evaluation
- [Original Dataset](https://huggingface.co/datasets/TeaPearce/CounterStrike_Deathmatch)