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 paper_neurips.pdf

Bundle contents

File	Purpose
LICENSE	CC-BY 4.0 license text
DATASHEET.md	Datasheet for Datasets (Gebru et al.)
endgameatlasbench_croissant.json	Croissant 1.0 descriptor for the source corpus
endgameatlasbench_benchmark_croissant.json	Croissant descriptor for the held-out benchmark split
endgame_atlas_bench_test.csv	Full benchmark pool: 35,006 unique FENs with human reference move + metadata
endgame_atlas_bench_test_20k.csv	Headline split: 19,747 unique FENs (20K seed=42 sample). Paper's Table 6 numbers are reported on this split.
sample.csv	3-row example of the source-corpus schema (per-move records)
sample_benchmark.csv	3-row example of the benchmark-split schema (per-FEN records)
human_classical_err_by_type.json	Per-type knowledge-difficulty reference vector (M1); 384 types, 2000-2400 Elo classical
human_exec_gap_by_type.json	Per-type execution-difficulty reference vector (M2); 384 types, bullet - classical at 2000-2400 Elo
difficulty_overall_ranking.csv	All 387 endgame types ranked by classical/bullet/blitz/rapid error rate and execution gap (paper App.~A)
overreach_drawn_cells.csv	Per-type×TC overreach ratios in tablebase-drawn positions, 2400+ vs 1600-2000 (paper §7, Table 6)
irreversibility_by_elo.csv	Game-level irreversibility rate by Elo tier × time control (paper §7.1)
increment_by_elo.csv	Within-TC error rate by increment bucket (0s vs 1-2s vs 3+s) and Elo tier (paper §4.3)
endgame_atlas_bench/	pip install-able evaluation package (incl. probe_human helper)

Two schemas

EndgameAtlasBench is distributed in two complementary forms:

1. Source corpus (per-move records) — every human move in every in-scope endgame position from the source corpus. Schema documented in endgameatlasbench_croissant.json. Use this for training, exploratory analysis, and to recompute aggregate human baselines.

2. Benchmark split (per-FEN records) — a deduplicated, held-out evaluation split, shipped as endgame_atlas_bench_test.csv (35,006 FENs) plus the seed=42 endgame_atlas_bench_test_20k.csv headline split (19,747 FENs). Each row is one FEN with the human reference move and metadata; tablebase ground-truth columns are filled locally with probe_human (see below). Schema in endgameatlasbench_benchmark_croissant.json, consumed by endgame_atlas_bench.evaluate.

Quick start: scoring a system

Install the eval package from the bundle root:

pip install -e .

Produce a predictions CSV with two columns (fen, move) — one row per position in the benchmark split:

fen,move
"8/8/8/8/4k3/8/4K3/4R3 w - - 0 1",e1e7
...

Score it:

import json
from endgame_atlas_bench import evaluate

human_K = json.load(open("human_classical_err_by_type.json"))
human_E = json.load(open("human_exec_gap_by_type.json"))

metrics = evaluate(
    predictions_csv="my_predictions.csv",
    ground_truth_csv="endgame_atlas_bench_test_20k_full.csv",  # see below
    human_classical_err_by_type=human_K,
    human_exec_gap_by_type=human_E,
    move_outcome_table=None,
)
print(metrics)

Completing the benchmark with Syzygy

The released CSV ships FENs, the human's move, and metadata. The four tablebase-derived columns (wdl, wdl_after_optimal, wdl_after_human, is_optimal_human) are filled in locally by the bundled probe_human helper. This avoids redistributing the 69 GB Syzygy WDL set; the tablebases are public domain and freely available.

# 1) Fetch the 3-6-piece Syzygy WDL tablebases (~69 GB, public domain)
mkdir -p ~/syzygy
rsync -av --include="*.rtbw" --include="*.rtbz" --exclude="*" \
      rsync://tablebase.lichess.ovh/tables/standard/3-4-5/ ~/syzygy/
rsync -av --include="*.rtbw" --include="*.rtbz" --exclude="*" \
      rsync://tablebase.lichess.ovh/tables/standard/6/ ~/syzygy/

# 2) Probe the FEN and the human's move on every row (~1-2 min for 20K)
python -m endgame_atlas_bench.probe_human \
       --in  endgame_atlas_bench_test_20k.csv \
       --out endgame_atlas_bench_test_20k_full.csv \
       --syzygy ~/syzygy

The _full.csv output has all four tablebase columns filled in; evaluate.py consumes it directly.

Scoring non-imitation systems

evaluate.py reports M0 (move-match accuracy) directly from the predictions CSV. Outcome-based metrics (%Err, M1, M2, M3, M4) need wdl_after for the predicted move. If your system always plays the human's reference move, the _full.csv above is enough; otherwise, probe the predicted move with the same Syzygy set and pass {fen: {move: wdl_after}} as move_outcome_table= to evaluate.

Citation

See the citeAs field of endgameatlasbench_croissant.json for the canonical BibTeX entry.

Maintenance

Issues, schema clarifications, and reproduction requests: github.com//endgame-atlas-bench (link populated post-review).