avewright/chess-soft-sf19

Official Stockfish 19 MultiPV soft targets. This release supersedes the 25k pilot. It is not a filter of chess-soft-multipv-lichess or chess-soft-100m-disagreements.

1,010,000 rows. Source id 4. Vocab compact (1968).

Mix (as generated)

origin	rows	note
self-play (origin=1)	915,850	SF19 vs SF19, ε=0.20, book + 4 random legal
relabel (origin=0)	84,150	existing local boards, new SF19 labels
frozen eval	10,000	split=1 in data/shard_000000.parquet

Relabel was intended to be 80%. The existing-board iterator stopped at ~84k accepted labels; self-play filled the 1M train remainder. split=0 is train.

Phase (train+eval new rows): opening 236,681 / middlegame 411,855 / endgame 351,464. Eval bucket: equal 490,702 / winning 311,052 / losing 150,760 / mate 47,486.

Teacher

• Stockfish 19 tag sf_19 (edb0d9db), EvalFile nn-1a298aa575a0.nnue
• Full file SHA-256 in teacher.json
• Full strength, Threads=1, Hash=32, UCI_ShowWDL=true
• Ponder left to python-chess (do not set it via configure)
• Hash is not cleared between production searches
• Label budget: 100k nodes / MultiPV=8 / tau=120
• Play (self-play only): 4k-node cheap search on unlabeled plies, label every 3 plies after ply 4

Targets

• Policy: STM softmax(tau=120) over the last complete MultiPV-8 iteration
• Unsearched legal moves are absent, not proven bad
• Mate rank is sign * (100000 - min(|mate|, 1000)), not mate-as-cp
• Bound scores are dropped
• cp / mate / wdl: White-absolute (training loader contract)
• soft_indices / soft_probs: width 8, pad -1 / 0
• soft_cps / soft_mates are stored so softmax can be rebuilt without SF
• FEN-only relabel rows drop repetition history and may omit EP if no legal capture

Honor split. shard_000000 is the frozen eval set (saved_split_v1). Do not invent a new position-hash holdout.

Quality audit (2,000 isolated positions, 100k/8 vs 1M/8)

• top-1 vs 1M: 0.6705 (self-play 0.7616666666666667, relabel 0.6314285714285715; endgame worst)
• regret p50=0 / p90=20 / mean=108
• missing 1M-ref mass p50=5.7% / p90=26.4%

See audit.json. Flags did not predict disagreement; no adaptive extra search.

Files

• data/shard_XXXXXX.parquet — one inbox shard per file
• teacher.json, summary.json, sampling.json, audit.json, eval_manifest.json