endgame
Datasets
All datasets matching “endgame”chess-endgame-tablebaseendgame-dataset
avewright/endgame-dataset
6–13 piece positions with official Stockfish 19 MultiPV soft
targets. SF vs SF games start from Lichess ECO, fast-forward (capture-biased)
until the piece window, then label.
300,017 rows. Source id 4. Vocab compact (1968).
Filter
Keep only boards with 6 ≤ n_pieces ≤ 13
Stop the game below 6 pieces (Syzygy territory)
split=1 is a ~5% holdout. Honor split.
ECO starts: 7,852.
Teacher
Stockfish 19 tag sf_19 (``)
Full… See the full description on the dataset page: https://huggingface.co/datasets/avewright/endgame-dataset.SalesTranscriptQA
SalesTranscriptQA
2,962 reviewed questions over 10,829 verbatim synthetic CRMArena-Pro sales-call transcripts. All 14,916 eligible source units were processed: single calls in both domains and distinct-dialogue pairs within one B2B opportunity. No B2C multi-call questions. Rejected proposals and review failures do not become benchmark questions.
Domain
Corpus calls
Single-call QA
Two-call QA
B2B
4033
1215
428
B2C
6796
1319
0
Methodology and… See the full description on the dataset page: https://huggingface.co/datasets/EndgameLabs/SalesTranscriptQA.lichess-endgame-bestlineEndgameAtlasBench
EndgameAtlasBench
A population-scale benchmark of human chess endgame decisions, labeled with
exact Syzygy tablebase ground truth (3-6 piece positions). Released for the
NeurIPS 2026 Datasets & Benchmarks Track.
Scale: 2,021,565,160 rated positions, 4,053,145 unique players (source corpus)
Source: Lichess Open Database monthly archives
Ground truth: Syzygy WDL tablebases (mathematically exact)
License: CC-BY 4.0 (see LICENSE); upstream Lichess data is CC0
Paper: see accompanying… See the full description on the dataset page: https://huggingface.co/datasets/Agusanso/EndgameAtlasBench.chess-grandmaster-endgames-v1