victorqueiroz/chessmate-positions

ChessMate Positions

Chess positions with Stockfish policy + value labels for supervised distillation of the ChessMate policy/value network (victorqueiroz/chessmate-net).

Stores framework-agnostic chess facts only — FEN, UCI moves, centipawn/mate scores. The 8×8×22 plane encoding and the 4096 policy_index are computed at train time by ChessMate's chess-core, so the data bakes in no engine-specific encoding.

Configs

config	rows	what
default	~702k	The v4 training set: phase-balanced (≈ middlegame/endgame + opening-book pass), Stockfish depth-18 multipv-8 policy.
ood	~4k	Out-of-distribution held-out eval set (engine self-play openings; VQ-444 D4) — disjoint from default, for measuring generalization off-distribution.
v3	~50k	Earlier (v3-era) labeling run; kept for history.
lc0_v7	2.5M	lc0-teacher labels (full policy distribution + WDL), VQ-497 scale-up. ⚠️ middlegame/endgame only (0% openings — the streaming source carries none; add an opening pass before training a shippable net).
lc0_onpolicy_8m	~7.95M	Scaled on-policy corpus (VQ-468): positions v5 reaches in its own raw-policy rollouts seeded from its v4 training FENs, lc0-labeled (same teacher + schema as lc0_v7). 7.78M unique (2.1% cross-wave dup); two waves (RNG seeds 42/99), generated by the batched selfplay_gen.py at ~772 pos/s. Distribution-matched — unlike lc0_v7 it inherits opening coverage from its seeds.

from datasets import load_dataset
ds  = load_dataset("victorqueiroz/chessmate-positions")             # default (v4 train)
ood = load_dataset("victorqueiroz/chessmate-positions", "ood")      # OOD eval

Schema

fen, side_to_move, fullmove, best_move (UCI), multipv (list of {move, cp, mate}, STM-POV), policy_depth, achieved_depth, policy_multipv, value_cp, value_mate, value_depth (value is white-POV), source.

lc0_v7 config — extra columns + a different encoding

The lc0_v7 config is labeled by an lc0 teacher (t1-256×10) rather than Stockfish, and adds two native columns: policy (the full lc0 policy distribution as a list of {move, prob}) and wdl ([win, draw, loss], STM-POV). It keeps the multipv / value_cp columns as a drop-in compat bridge so the existing trainer works unchanged — but the encoding differs:

• multipv[i].cp is not a centipawn eval; it encodes the lc0 prior as cp_i = SOFT_TEMP·(ln p_i − ln p_max), so softmax(cp / SOFT_TEMP) reproduces lc0's policy. The best move's cp is always 0 (it is the reference). A Stockfish-style audit that correlates value_cp against multipv[0].cp will therefore see a constant 0 and report a spurious failure — that check does not apply to this config.
• value_cp is lc0's WDL mapped to white-POV cp via cp = VALUE_SCALE·atanh(w−l).

The lc0_onpolicy_8m config is produced by the same lc0 labeler (pretrain/hf/lc0_labeler.py) on positions from the batched on-policy generator (pretrain/hf/selfplay_gen.py, VQ-468) — it shares this schema and these encoding caveats.

Provenance & generation

• Positions + carried value evals sampled from mateuszgrzyb/lichess-stockfish-normalized (Lichess games, deep Stockfish value), plus Lichess puzzle FENs for the tactical slice and engine self-play for ood.
• Policy labels (best_move, multipv) freshly computed by Stockfish multi-PV at policy_depth.
• Generated by pretrain/hf/label_positions.py in the ChessMate repo (dedup + phase-balance + tactical/OOD passes).

License & attribution

CC-BY-4.0. This dataset is a derivative of mateuszgrzyb/lichess-stockfish-normalized (CC-BY-4.0 — attribution required) and Lichess open data (CC0). Freshly computed labels are produced by Stockfish (a GPL engine; its move/eval output does not place the GPL on this data). If you use this dataset, please attribute mateuszgrzyb/lichess-stockfish-normalized and Lichess.