AntimLabs/combined-game-sft

---
license: mit
task_categories:
- reinforcement-learning
- text-generation
tags:
- game-ai
- sft
- multi-task
---

# Combined Game SFT Dataset

Multi-game SFT dataset for training game-playing language models.

## Dataset Details

- **Total Examples**: 7,188
- **Games**: 9 different environments
- **Format**: Standard SFT with `prompt`, `completion`, and `source` columns
- **Balanced**: Each game capped at 1,200 examples max

## Game Distribution

| Game | Examples | % |
|------|----------|---|
| taxi | 1,200 | 16.7% |
| snake | 1,200 | 16.7% |
| cliff-walking | 1,000 | 13.9% |
| frozen-lake | 1,000 | 13.9% |
| flappybird | 710 | 9.9% |
| minigrid-lockedroom | 620 | 8.6% |
| pacman | 500 | 7.0% |
| blackjack | 480 | 6.7% |
| minigrid-lavagap | 478 | 6.6% |

## Source Datasets

- AntimLabs/FlappyBird-SFT
- AntimLabs/cliff-walking-sft
- AntimLabs/taxi-sft-new
- AntimLabs/MiniGrid-LockedRoom-SFT
- AntimLabs/blackjack-sft
- AntimLabs/Frozen-Lake-SFT
- AntimLabs/Minigrid_LavaGap
- AntimLabs/Pacman-Array-SFT
- gokul8967/snake-sft-8x8

## Usage

```python
from datasets import load_dataset

dataset = load_dataset("AntimLabs/combined-game-sft", split="train")

# Filter by game
pacman_data = dataset.filter(lambda x: x["source"] == "pacman")
```