The Open Chess Plies (OCP)

A large-scale open chess dataset for machine learning: 358K+ engine-annotated positions from 32 open-source engines — classical + Chess960/DFRC self-play, elite human openings, and curated UHO systems. Every ply ships eval, WDL, depth, mate + full game metadata. Monthly versioned releases, MIT licensed.

Quickstart

import datasets
ds = datasets.load_dataset("rkaluzny/OCP", split="train")
print(ds[0])
# {'fen': 'r1bq1knr/pp3ppp/2nb4/1B1p4/3p4/1P3N2/...', 'eval': 119,
#  'wdl': [0.3936, 0.3417, 0.2647], 'depth': 22, 'white': 'Clover 9.1', ...}

Raw JSONL without dependencies:

import json
with open("ocp-2025-06.jsonl") as f:
    for line in f:
        row = json.loads(line)  # one position per line

Polars (fast filtering):

import polars as pl
df = pl.read_ndjson("ocp-2025-06.jsonl")
print(df.filter(pl.col("depth") >= 20).head())

Releases

Release	Positions	Format	Download
OCP-2025-06	monthly snapshot	JSONL	data/ocp-2025-06.jsonl
OCP-2026-05	150k	JSONL	data/ocp-2026-05.jsonl
Viewer (train)	358k rows	Parquet + SQL	Dataset Viewer on Hugging Face

Files are immutable — old releases never change. Cite the dated file you used. Direct link pattern: https://huggingface.co/datasets/rkaluzny/OCP/resolve/main/data/ocp-YYYY-MM.jsonl?download=true

Schema

One JSON object per line. eval is centipawns from White's perspective; wdl is [white win, draw, black win].

Field	Type	Description
fen	string	Position
eval	int	Engine eval, cp, White POV
mate	int | null	Mate distance, if found
wdl	float[3]	Win/draw/loss probabilities
white / black	string	Engine names + versions
tc	string	Time control, e.g. 60+0.1
depth	int	Search depth of the annotation
gameply	int	Ply number within the game
result	string	1-0, 0-1, 1/2-1/2, *
event / round	string	Tournament + round
date	string	YYYY.MM.DD
game_id	int	Groups plies from the same game

Example row:

{
  "fen": "8/p3rpk1/1p1p2p1/3Bb2p/2P4P/1P3RP1/P1nB1P2/5K2 w - - 65 68",
  "eval": 107,
  "mate": null,
  "wdl": [0.3774, 0.3585, 0.2641],
  "white": "Clover 9.1",
  "black": "stormphrax 7.0.0",
  "tc": "60",
  "depth": 24,
  "gameply": 134,
  "result": "1-0",
  "event": "V1",
  "round": "2",
  "date": "2026.05.22",
  "game_id": 19
}

Why OCP exists

Traditional chess datasets often suffer from excessive repetition of equal positions, overrepresentation of opening theory, limited structural diversity, and noisy human annotations. OCP addresses this with hybrid generation — engine self-play (Cute Chess round-robins, 1–4 threads, AVX2, UHO + DFRC openings) blended with elite human opening games — and deliberate retention of positions with initiative, imbalance, attacks, and asymmetry.

OCP is not a rating list, a benchmark suite, or an opening book. Note: CPU-only self-play means GPU-oriented nets (notably Lc0) play below their reputation — see the project FAQ.

Intended uses

NNUE training, eval calibration, WDL modeling, representation learning, policy/value experiments, opening-structure analysis, distillation.

Limitations

Experimental and evolving: eval noise, opening-source imbalance, duplicated motifs, and (for now) modest scale vs. billion-position corpora. Filter and validate for your use case.

Citation

@dataset{kaluzny_ocp_2026,
  author = {Kaluzny, Robin},
  title = {The Open Chess Plies (OCP)},
  year = {2026},
  url = {https://huggingface.co/datasets/rkaluzny/OCP},
  note = {Large-scale open chess dataset for machine learning and chess engine research, MIT licensed. If used, please cite the release file, e.g. ocp-2025-06.}
}