ishagarg1103/gambit-zero

Gambit Zero — Human Chess Dataset A

Status: complete-50k-filtered-evaluation-v1. Only completed, checksum-verified monthly artifacts listed in the manifest are usable. Building datasets are not the accepted 50K pilot.

Repository visibility is public, as requested by the owner. Run: human-v3-pilot-20261025-dataset-d-v1. Parent run: human-v3-pilot-20261025. This repository stores dataset artifacts, not model weights. No M2 model or self-play training is included.

Size and splits actually published in this catalog

Split	Games	Training positions
train	35000	2654850
validation	7500	569124
test	7500	569626

Final real-data targets are 35,000 training games (5,000 each January–July 2026), 7,500 validation games (August), and 7,500 test games (September). A game never crosses splits. The table above reports this catalog, not promised target counts.

Sources, selection and verification

Real sources: public CC0 Lichess monthly standard-rated PGN archives, https://database.lichess.org/. Exact URLs, published hashes and verification status are in source_provenance.json and each completed month manifest. Published archive hashes alone do not mean the complete sources were verified. Synthetic smoke artifacts are explicitly marked and are not Lichess samples.

Streaming consumes original compressed bytes in bounded frames without retaining full archives. Month checkpoints are accepted only after byte-count/EOF and full SHA256 verification against the official monthly checksum. Sampling uses utc-day-rating-bottomk-sha256-v1, seed 20261025, full UTC calendar-day strata, deterministic SHA256 priorities and bounded reserves. At least 60% of accepted games per day have both player ratings >=2000; all remaining games are >=1800. The lower rating defines the band. Exhausted reserves/coverage gaps stop the run.

Quality and targets

Standard rated chess from the initial position, base clock >=300 seconds independent of increment, valid complete results and legal moves. Exclude variants, abandoned/aborted/rules-infraction games, malformed metadata and games that cannot be completely replayed under the 512-ply horizon. Time forfeits retain policy targets and default value weight 0. Generic Normal endings are not misrepresented as proven resignation/agreement. Labels are noisy human outcomes, not strategic ground truth. No Stockfish/[%eval] labels are consumed.

Replay every accepted trajectory from the initial corrected Pgx state, preserving history. Full legal actions, position/history, terminal status and reward parity are checked against python-chess offline at every ply. Actual final checkmate is checked against Pgx; non-board PGN results do not require environment termination. Rules: tinychess-auto-fide-pgx-v1, no voluntary claims, fivefold/150-halfmove automatic draws; truncations never supply WDL labels. Remaining dead-position and hash-collision limitations are inherited from the accepted rules contract.

Schema and encoder

human_v3; schema.json defines every physical field. Policy uses Pgx's native 4,672-action order, codec pgx-2.6.0-chess-v2-uci-v1. Encoder pgx-v2-history8-ep-context8-v1: planes float32 (8,8,120), context float32 (8,), legal mask bool (4672,). Feature bytes are little-endian float32; masks are 584 bytes, little bit order. See encoder_contract.md for channel/context order. WDL class order is win/draw/loss (0/1/2), scalar +1/0/-1 from side to move. Game metadata/action sequences are provenance; the primary training inputs are the materialized positions configuration.

Evaluation integrity and limitations

Dataset D dataset-d-v1-20261029 is frozen independently of overlap/model performance. Exact matching training positions are physically excluded; accepted game identities and held-out chronological data remain intact. Original data revision: d409fce8bd444c2d456ef4c343dbc9b5e25981eb. Puzzle solutions are evaluation only and are never Dataset A training labels. Complete compact game actions remain original provenance: replaying them as training requires the same frozen evaluation exclusions. Match pieces, side to move, castling and legally capturable EP; clocks/history are ignored conservatively. See dataset_manifest.json for the frozen suite receipt, revised monthly manifests, postfilter overlap and the original position-disjoint diagnostic subset. That original subset is conservative for the filtered training set, not claimed maximal.

Upload, reproducibility and recovery

Small game-aligned Parquet shards are grouped by kind, split and source month. Each immutable filename includes its full SHA256. Only a completed month is listed in the catalog. Uploads use authenticated huggingface_hub; files are read back at an immutable revision, streaming without a download cache, and compared by SHA256 and byte count before the Mac copy is removed. An upload/read-back failure retains the local file and halts generation of further shards.

A second feature build can consume the verified compact selected-PGN input from Hugging Face, one bounded selection file/game at a time, and compare regenerated shard hashes without a second local dataset. Reproduction is accepted only when actually run; completed reproduction manifests record results. Store the upload journal and completed-month receipts for recovery. Interrupted archive months restart from byte zero; verified completed months are reused. No blind compressed byte-range game access is assumed.

Usage without downloading the whole dataset

Authenticate using hf auth login; never embed a token in a notebook or commit. Install datasets in your consumer environment, then use its streaming mode:

from datasets import load_dataset
# Use the configuration listed in this card; smoke data is synthetic only.
rows = load_dataset("ishagarg1103/gambit-zero", name="human_v3",
                    split="train", streaming=True)
row = next(iter(rows))

Decode a position row (NumPy):

import numpy as np
planes = np.frombuffer(row["planes"], dtype="<f4").reshape(8, 8, 120)
context = np.frombuffer(row["context"], dtype="<f4")
legal = np.unpackbits(np.frombuffer(row["legal_mask"], dtype=np.uint8),
                      bitorder="little").astype(bool)
assert legal[int(row["action"])]

Use policy_weight/value_weight separately, retain perspective, and pin the Hub revision in any experiment. Outcome value loss must respect the timeout mask. This dataset makes no claim of optimal moves, model strength or frontier-GPT wins.

Dataset D revision

{"test": {"games": 7500, "positions": 569626}, "train": {"games": 35000, "positions": 2654850}, "validation": {"games": 7500, "positions": 569124}}

Rewritten shards were filtered twice and every retained field/feature byte was compared with the original. M1 final review and the separate GPU benchmark disposition remain required. No model training performed.