zero-evaluator / RESULTS.md
Pawitt's picture
Publish high-variance 1.5M chess position Parquet dataset
5f87c14 verified
|
Raw
History Blame Contribute Delete
3.01 kB

Completed Dataset Results

Artifacts

  • Dataset: data/ccrl-high-variance-1p5m.vpd
  • Training dataset: data/ccrl-high-variance-1p5m.parquet/
  • Parquet manifest: data/ccrl-high-variance-1p5m.parquet/_manifest.json
  • Variance report: data/variance-report.json
  • Source archive: source/ccrl-pgn.tar.bz2
  • Extractor and reader: vpd.py
  • Format specification: FORMAT.md

Provenance

The source is Leela Chess Zero's published CCRL standard dataset: 2.5 million CCRL 40/40 and 40/4 engine games with an 80/20 train/test split. Extraction used 910,462 source games and completed with zero PGN parse errors.

The final VPD1 database contains 1,509,201 unique normalized FEN positions and is 438,558,720 bytes.

The Parquet derivative contains the same 1,509,201 rows in six Hive-style source_split/phase partitions. It is 52 MiB on disk, uses Zstandard level 6 compression, and has 26 row groups capped at 65,536 rows.

Variance result

Overall result: HIGH VARIANCE — PASS

Every declared threshold in data/variance-report.json passed.

Dimension Result
Opening 234,201 (15.52%)
Middlegame 900,000 (59.63%)
Endgame 375,000 (24.85%)
Normalized phase entropy 0.858704
White wins 570,602 (37.81%)
Draws 481,726 (31.92%)
Black wins 456,873 (30.27%)
White to move 758,019 (50.23%)
Black to move 751,182 (49.77%)
Positions in check 103,519 (6.86%)
Positions with castling rights 289,814 (19.20%)
Positions without castling rights 1,219,387 (80.80%)
Material imbalance of at least 3 234,018 (15.51%)

Legal move counts span 0–87, with p10 12, median 33, p90 44, and standard deviation 12.064. Piece counts span 2–32, with p10 10, median 21, p90 30, and standard deviation 7.5146.

Validation

  • SQLite PRAGMA integrity_check: ok
  • Random FEN validation: 1,000/1,000 valid
  • Indexed random retrieval: 1,000 queries in 4.404 ms
  • Mean in-process lookup latency: 0.004404 ms per board
  • Parquet row-count validation: 1,509,201/1,509,201
  • Parquet partition-count validation: 6/6 exact matches
  • Parquet FEN validation sample: 6,000/6,000 valid
  • Parquet file checksums: 6/6 match _manifest.json
  • Parquet compression audit: all 26 row groups use Zstandard
  • Source archive SHA-256: 5f4d7ec86a99ba56fd3e46b1eb35f3ae109890bc77fecc3e605a6174ea717e76
  • Dataset SHA-256: 2f3af2973d1e4cb2d7e0d54e7e7a0f12def44d982915df0c9cab1735f6b138cc
  • Report SHA-256: 759699d7bc704fd265524a6df320c1a0c925f902e29a60b83366d987d3ab507e

Retrieve one board

cd /Users/pawit/Documents/vexilon/tmp/vex-position-dataset
PYTHONDONTWRITEBYTECODE=1 .venv/bin/python vpd.py sample \
  data/ccrl-high-variance-1p5m.vpd

Example output:

{"id":257499,"fen":"8/4p3/p6p/1p3k1K/1n3P2/1P4P1/r3RN2/8 w - - 0 1","phase":"endgame","result":"1-0"}

The source result is provenance metadata. It is not the depth-zero static WDL label; lc0 can add that label in a subsequent pass.