Sign in to view source links and access this dataset
Description
Derived from the action-value data released with DeepMind's 'Grandmaster-Level Chess Without Search' paper, this dataset reorganizes chess positions and their legal moves. It provides per-move win probabilities and augments decisive moves with Stockfish-calculated mate depths. The dataset was created by Ramora0 and last updated on Hugging Face on February 6, 2026.
Use Cases
Training chess evaluation models based on per-move win probabilities.
Analyzing decisive positions using the mate-in-N augmentation.
Benchmarking AI move selection against grandmaster-level action-value data.
Studying the structure of chess positions and legal move sets using FEN notation.
Strengths
Derived from a well-known DeepMind research release on grandmaster-level chess AI.
Augments raw action-values with Stockfish-calculated mate depths for decisive moves.
Reorganizes data into a per-position structure for easier analysis of legal moves.
Limitations
Column-level documentation is absent; field semantics must be inferred after download.
Row count is unknown, which may limit suitability assessment.
Description metadata is limited; actual data quality requires manual inspection after download.
Provenance
Source
DeepMind's action-value data from the 'Grandmaster-Level Chess Without Search' paper.
Collection Method
Reorganization and augmentation of the original release with Stockfish mate search.
Freshness
Last updated 2026-02-06 20:28:49; freshness should be verified.
License is unknown and must be verified before use.