Synthetic Chess Reasoning Data For Language Model Training

A collection of text-based datasets designed for training and evaluating reasoning in large language models, using chess as a domain. The data was created by lucasdino and was last updated in April 2026. It is described as being balanced across several axes, though specific row counts and file sizes are not provided.

Use Cases

• Fine-tuning language models on chess-related reasoning schemas mentioned in the paper.
• Training reinforcement learning agents using the described distributions of chess moves and positions.
• Benchmarking model performance on synthetic text tasks involving chess strategy and logic.
• Analyzing how reasoning evolves from data through fine-tuning, as referenced in the associated research.

Strengths

• Data is explicitly balanced across several axes, as noted in the TLDR.
• Dataset size is categorized as between 10 million and 100 million entries on the Hugging Face platform.

Limitations

• Specific row count, column names, and sample data are unavailable, hindering initial assessment.
• The data is synthetic, which may limit its applicability to real-world chess gameplay or reasoning scenarios.
• Geographic and temporal coverage of the underlying chess games is unknown.

Provenance

Source

lucasdino on Hugging Face.

Collection Method

Synthetic generation, likely for research purposes as described in the linked paper.

Freshness

Last updated on 2026-04-03.