Datasets:

abdelkader-dev
/

XO_matches

language: en
license: mit
tags:
  - tic-tac-toe
  - xo
  - synthetic
  - board-game
  - classification
task_categories:
  - text-classification
size_categories:
  - 10K<n<100K
pretty_name: FemtoXO Dataset (Synthetic XO Games)
configs:
  - config_name: default
    data_files: data.jsonl

FemtoXO Dataset – Synthetic Tic‑Tac‑Toe Games

FemtoXO Dataset is a large collection of board states and moves from randomly played Tic‑Tac‑Toe games.
It was created to train the FemtoXO model, a tiny Transformer that plays as X.

Dataset Summary

Each row represents one board position where it is X’s turn to play, along with the move X made in that turn.
The dataset is entirely synthetic and was generated programmatically by simulating 10,000 full random games and recording every X‑turn state.

Total samples: ≈ 90,000 (varies slightly due to different game lengths)
Format: JSON Lines (.json)
Language: not applicable (board symbols)

Data Structure

Each line is a JSON object with two fields:

Field	Type	Description
`board`	string (length 9)	Board state: `.` = empty, `X` = player X, `O` = player O
`move`	integer (0–8)	The index of the cell (0‑based, row‑major) chosen by X

Example:

{"board": "X..O.....", "move": 4}

Data Splits

The data is provided as a single file (data.json) containing all samples.
During training we typically split it into:

Train: 90%
Validation: 10%

Generation Process

Start with an empty 3×3 board (all .).
Players alternate turns (X first, then O), each choosing a random legal move.
Before every X move, save the current board state and the chosen move.
Game ends on a win or a draw (board full).

The complete generator script is available in the FemtoXO repository under src/train.py.

Usage

You can load the dataset directly with 🤗 Datasets:

from datasets import load_dataset

dataset = load_dataset("abdelkader-dev/XO_matches")
print(dataset['train'][0])

Known Limitations

Random strategy: Moves are chosen uniformly, so the dataset does not contain optimal/Minimax play. A model trained on this data will learn only to avoid immediate mistakes but not to force a win/draw optimally.
No board rotation/augmentation: All boards are in fixed orientation. You can apply data augmentation (rotations/reflections) during training to improve robustness.

Citation

If you use this dataset, please cite:

@dataset{femto-xo,
  author    = {Abdelkader Hazerchi},
  title     = {FemtoXO Dataset: Synthetic Tic‑Tac‑Toe Games},
  year      = {2025},
  url       = {https://huggingface.co/datasets/abdelkader-dev/XO_matches}
}

License

MIT – feel free to use, modify, and share.