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

# CS:GO VLA Stage 1 Dataset (5Hz Chunked)

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

## Overview

- **Frame rate:** 5Hz (every 3rd frame)
- **Action chunking:** 3 actions per sample (~200ms coverage)
- **Total samples:** ~1.8M chunks
- **Split:** train / test following [Diamond](https://github.com/eloialonso/diamond) split
- **Map:** Dust2 deathmatch

## Action Format

```
<|action_start|> m1_x m1_y [keys1] ; m2_x m2_y [keys2] ; m3_x m3_y [keys3] <|action_end|>
```

**Examples:**
```
<|action_start|> 0 0 ; 0 0 ; 0 0 <|action_end|>                              # idle
<|action_start|> 5 0 W ; 3 0 W ; 4 0 W <|action_end|>                        # walking
<|action_start|> -200 50 W L ; -50 10 L ; 10 0 W <|action_end|>              # flick shot
```

Each chunk contains the exact mouse delta and keys for that frame - no aggregation.

## Schema

| Column | Type | Description |
|--------|------|-------------|
| `id` | string | Unique sample ID |
| `episode_id` | string | Source HDF5 file |
| `chunk_idx` | int32 | Chunk number within episode |
| `frame_idx` | int32 | Starting frame number |
| `action` | string | Text-formatted 3-action chunk |
| `kill_flag` | int32 | 1 if any kill in chunk |
| `death_flag` | int32 | 1 if any death in chunk |
| `split` | string | "train" or "test" |
| `image_bytes` | bytes | JPEG screenshot (first frame) |

## Usage

```python
from datasets import load_dataset

# Load full dataset
ds = load_dataset("TESS-Computer/csgo-vla-stage1-5hz")

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

## Why 5Hz with Chunking?

1. **VLA inference speed:** 62ms (16Hz) is too fast for current VLMs. 200ms (5Hz) is achievable.
2. **No information loss:** Each chunk predicts exactly what the human did for 3 consecutive frames.
3. **World model sync:** Diamond executes `step(a1), step(a2), step(a3)` then returns frame to VLA.

## Related

- [16Hz variant](https://huggingface.co/datasets/TESS-Computer/csgo-vla-stage1-16hz) - 1 action per frame
- [Diamond World Model](https://github.com/eloialonso/diamond) - For evaluation
- [Original Dataset](https://huggingface.co/datasets/TeaPearce/CounterStrike_Deathmatch)