jaagli/othello_shuffle

Dataset Card

Dataset Description

The dataset is derived from real Othello game records collected from EOTHELLO. It combines textual move sequences with corresponding visual board states, enabling joint modeling of language and vision in a structured, rule-based environment.

Each game consists of a sequence of 60 ± 2 moves on average, with one board image generated after every move. This results in a total of approximately 25,000 games and 1.56 million board images.

It provides two synchronized modalities:

• Text modality – move tokens representing board positions (e.g., “C4”, “E6”).
• Visual modality – RGB images depicting the full Othello board state after each move.

Repository:

multimodal-othello

Statistics

Split	Number of Games	Number of Images	Avg. Images per Game
Training	20,525	1,247,852	~ 60.8
Validation	1,282	78,141	~ 60.9
Test	3,850	233,975	~ 60.8
Total	25,657	1,559,968	~ 60.8

Intended Usage

The dataset is intended for academic research for training and/or evaluating language models.

Citation

Paper: What if Othello-Playing Language Models Could See?

BibTeX:

@article{chen2025if,
  title={What if Othello-Playing Language Models Could See?},
  author={Chen, Xinyi and Yuan, Yifei and Li, Jiaang and Belongie, Serge and de Rijke, Maarten and S{\o}gaard, Anders},
  journal={arXiv preprint arXiv:2507.14520},
  year={2025}
}