NewSonnet/flychess-stockfish-500k

Flychess Stockfish 500k

A reproducible, sharded dataset of 500,000 legal chess positions with Stockfish depth-3 move and bounded value labels for training and evaluating Flychess FlyNet.

This repository is the canonical 500k dataset release. It contains binary training shards, integrity sidecars, a machine-readable manifest, and the schema files needed to interpret the arrays without guessing.

Release identity

• Dataset: Flychess Stockfish 500k
• Canonical release prefix: campaign/depth3/segment-000-500k-fresh-20260921
• Samples: 500,000
• Shards: 5 x 100,000 positions
• Feature width: 851
• Action vocabulary: 4,544 deterministic UCI move shapes
• Teacher: Stockfish, depth 3
• Position sampler: uniform_legal_playout_per_global_index
• Plies per sampled position: 4-60
• Generation seed: 20260921
• Source commit: 1efceab534244096b9af309262913abdf0be2982
• Schema version: flynet.sharded_dataset/v1

The manifest is authoritative for exact counts, shard boundaries, checksums, and provenance. The three schema files in the release prefix are authoritative for array interpretation.

Repository layout

campaign/depth3/segment-000-500k-fresh-20260921/
|- manifest.json
|- feature-schema.json
|- label-schema.json
|- action-vocabulary.json
|- shard-000000.npz
|- shard-000000.json
|- ...
|- shard-000004.npz
|- shard-000004.json

Each NPZ shard contains four arrays:

Array	Shape	Stored dtype	Meaning
features	[N, 851]	float16	FlyNet board encoding
action_indices	[N]	uint16	Index into action-vocabulary.json
values	[N]	float16	Bounded side-to-move teacher value in [-1, 1]
fens	[N]	Unicode string	Exact FEN for the sampled position

The per-shard JSON sidecar records the sample range, byte count, storage contract, and SHA-256 digest for the corresponding NPZ file.

Loading a shard

Use allow_pickle=False when reading the files:

from pathlib import Path
import numpy as np
from huggingface_hub import snapshot_download

repo_id = "YOUR_HF_NAMESPACE/flychess-stockfish-500k"
prefix = "campaign/depth3/segment-000-500k-fresh-20260921"

root = Path(snapshot_download(
    repo_id=repo_id,
    repo_type="dataset",
    allow_patterns=[f"{prefix}/*"],
))

with np.load(root / prefix / "shard-000000.npz", allow_pickle=False) as shard:
    features = shard["features"]
    action_indices = shard["action_indices"]
    values = shard["values"]
    fens = shard["fens"]

print(features.shape, action_indices.shape, values.shape, fens.shape)

The source project provides the compatible FlyNet encoder and training loader. The schema files make the storage contract independently inspectable.

Feature contract

The 851 features are ordered as follows:

1. 768 piece-square indicators: white then black, six piece types, 64 squares.
2. One side-to-move scalar: +1 for White and -1 for Black.
3. Four castling-right indicators.
4. A 64-element en-passant-square one-hot region.
5. Halfmove and fullmove counters with the documented clipping and scaling.
6. Twelve normalized piece-count features.

See feature-schema.json for exact half-open index ranges and semantics.

Provenance and limitations

• Positions are generated from legal chess playouts; they are not human game records.
• Stockfish supplies the move and bounded value labels. Depth 3 is a low-cost teacher setting, not ground truth or a claim of playing strength.
• This release is a versioned publication of a verified 500,000-position snapshot. The manifest records the source prefix and source commit; the upload itself does not claim that all positions were newly generated during publication.
• The data encodes the FlyNet feature contract. It is an engineering dataset, not a biological connectome or evidence of biological cognition.
• The dataset is stored as NumPy NPZ rather than Apache Arrow/Parquet, so the full feature arrays are intended for programmatic loading rather than automatic row preview in the Hub Dataset Viewer.

Licensing and attribution

The dataset contains generated labels produced with Stockfish; it does not redistribute Stockfish source code or a Stockfish binary. Stockfish is distributed under GPLv3. Review the upstream Stockfish license and the Flychess source provenance before redistributing or incorporating this dataset into another release.

• Stockfish: https://github.com/official-stockfish/Stockfish
• Flychess source project: https://github.com/EF-Code/flychess

No broad permissive license is asserted for this generated artifact beyond the applicable third-party terms and the provenance recorded in manifest.json.