quintic/go19x19

Go 19x19: mixed-strength games with one fixed teacher

This release contains 3,200 complete games and 1,356,897 pre-action positions. The fixed teacher, kata1-tf3-b11c768-s11500M-d6163M, plays against eight KataGo checkpoint strata, with 400 games per stratum. One stratum is teacher self-play. The same teacher labels both players' positions. These are synthetic engine games, not human games or self-play from the research architectures.

The release contains both complete game archives and the exact packed train/validation corpus used in the supervised CNN/dense-transformer/MoE comparisons. The packed corpus has 3,058 games and 1,297,737 positions; it excludes test games. Its original manifest SHA-256 is preserved exactly: 8e1ab17423f083137367adec93d215d7dcea2440dfafa1c448f0cea834470855.

Splits

Split	Games	Positions
train	2,906	1,233,366
validation	152	64,371
test	142	59,160

D4-canonical first-eight-move opening families determine the split. The hash bucket assigns 0-4 to test, 5-9 to validation and 10-99 to training. Family IDs and game IDs remain unchanged. No family crosses a split. Later transpositions can still be shared. The fixed research view includes terminal training and validation games up to 768 plies; all collected games fit this limit. The held-out test games are archived for preservation and have not been used for these architecture comparisons.

Behavior and supervision

Rules: 19x19, komi 7.5, positional superko, area scoring, multi-stone suicide allowed, no tax/button/handicap, and friendlyPassOk=false. Games terminate naturally; there is no resignation action. Every collected game passed full-history board/legality validation during generation.

Each player searches with 16 visits. For the first 16 plies, behavior samples 75% normalized root-edge visits plus 25% raw policy; afterward it chooses the first reported root move. Full move histories are supplied to KataGo. Root noise and symmetry pruning are disabled. This small search budget does not represent maximum-strength KataGo play.

raw_policy and raw_value come from the fixed teacher's raw neural outputs on every ply. On an opponent turn, a one-visit teacher query supplies those raw labels. search_policy and search_value are eligible only where search_policy_valid is true. actions contains actual mixed-strength behavior and is a different target from the raw teacher policy.

Raw value and score labels use the player-to-move perspective. The metadata field white_score is the terminal game score from White's perspective, including komi, and must not be confused with the raw value target. Checkpoint names, URLs, hashes, rules and behavior settings are in contract.json.

Files and restoration

• raw/part-*.tar: uncompressed containers of individually compressed NPZ games.
• index/part-*.jsonl: metadata, archive members and checksums for each game.
• packed/manifest.json and packed/shard-*/: exact historical feature/target arrays.
• manifest.json: complete release inventory and validation counts.
• SHA256SUMS: checksums of every release file other than itself.
• read_games.py: streaming game reader requiring only NumPy.

Pin an immutable Hub revision when comparing models:

hf download quintic/go19x19 --type dataset --revision RELEASE_COMMIT --local-dir go19x19
cd go19x19
sha256sum -c SHA256SUMS

For the exact research corpus, install the gozero library from the research source and load the packed directory:

from gozero.corpus_sequence_batches import Dataset

data = Dataset('go19x19/packed',
    '8e1ab17423f083137367adec93d215d7dcea2440dfafa1c448f0cea834470855')

For complete game records:

from read_games import games

for metadata, arrays in games('go19x19', split='train'):
    actions = arrays['actions']
    teacher_policy = arrays['raw_policy']
    teacher_value = arrays['raw_value']

Opening the test split requires allow_test=True in the reader.

Array schema

For a game of T plies, coordinates are row-major from the top left. Action 361 is pass; stones use 0 empty, 1 Black and 2 White. Black moves first.

Raw array	Shape
stones	T x 19 x 19
legal	T x 362
raw_policy, search_policy, root_edge_visits	T x 362
raw_value, search_value, raw_score	T
search_policy_valid, raw_score_valid, teacher_visits, actions	T

The packed view uses KataGo V7 features: 22 binary spatial channels and 19 float32 global features. Spatial and legal arrays are packed least-significant bit first. The included manifest declares every shape, dtype and checksum. expert_offsets separates complete games; games stores stable identities, opening families, trajectory hashes, split, opponent, expert color and length.

Privacy and provenance

Deployment ordinal fields were removed from raw game metadata, and archive headers use anonymous ownership and fixed timestamps. No machine names, account paths, credentials or operational inventories are included. All numeric NPY member bytes in each raw game were verified unchanged. The index records both the new NPZ checksum and its original checksum. The packed corpus remains byte-for-byte identical, including its original manifest.

The packed manifest's source_game_sha256 references original game payloads; use original_npz_sha256 in the release index to map those identities to the sanitized game archives. Packaging did not change actions, labels, features, legal masks or split assignments. Test records were preserved without model evaluation or selection. This release supplies data, not trained weights or an assertion of playing strength.