A chess evaluation is only as comparable as the search that produced it. This collection’s value is not a single definitive score for each position, but a large record of engine judgments made at varied search depths and node counts—making it useful for studying how evaluations and suggested lines change with search effort.
What the Data Enables
- Each row pairs a FEN position with a principal variation and either a centipawn score or a mate result, giving models both a judgment and a candidate line to learn from.
- The source represents 416,442,401 positions; its denormalized form expands those records into more than one billion rows. This scale supports large-scale training and analysis, but the two counts describe different representations.
- Depth and searched-node counts provide context for each evaluation, so researchers can filter or compare results by search effort rather than treating every score as equally reliable.
- The data comes from different Stockfish variants running in users’ browsers. That broadens coverage, while also making it unsuitable as a uniform, tightly controlled engine benchmark.
Who It Suits—and the Trade-Off
A good fit for chess-engine research, position-evaluation models, and experiments that need many labeled positions with associated move lines. Look elsewhere for a curated tournament dataset, a consistent evaluation protocol, or ground-truth scores independent of Stockfish; these are engine-generated judgments at varying search settings, not objective labels.
