Datasetpublic

mcmcmcmc/maia3-training-data

Maia3 Training Data Preprocessed chess position data for training Maia3 human move-prediction models. Positions are stored as precomputed features, not raw FENs, so training reads and expands them without re-tokenizing every epoch. Contents path rows size lichess_parquet/train_YYYY-MM_precomputed.parquet (31 files, 2023-01 … 2025-07) 334,438,119 ~33 GB allie_data/test_precomputed.parquet 884,049 69 MB allie_data/2022-test-annotated.jsonl — 45 MB… See the full description on the dataset page: https://huggingface.co/datasets/mcmcmcmc/maia3-training-data.

sourceHugging Facecc0-1.0updated 16d agoView on Hugging Face

1likes321downloads

Dataset Card

Maia3 Training Data

Preprocessed chess position data for training Maia3 human move-prediction models. Positions are stored as precomputed features, not raw FENs, so training reads and expands them without re-tokenizing every epoch.

path	rows	size
`lichess_parquet/train_YYYY-MM_precomputed.parquet` (31 files, 2023-01 … 2025-07)	334,438,119	~33 GB
`allie_data/test_precomputed.parquet`	884,049	69 MB
`allie_data/2022-test-annotated.jsonl`	—	45 MB
`valid_2019-01_history_7_precomputed.parquet`	1,000,000	104 MB

Schema

All parquet files share one schema, with maia3_history: "7" in the file metadata (boards stored per position; a model may use fewer).

column	type
`piece_ids`	`list<list<uint8>>`	per-board piece ids, 64 squares per board, oldest board first
`castling`	`list<uint8>`	castling rights per board, bit-packed
`ep_square`	`list<int8>`	en-passant target square per board, `-1` if none
`move_idx`	`int16`	index of the played move in the fixed 4352-move vocabulary
`legal_idx`	`list<int16>`	indices of all legal moves in that vocabulary
`self_elo`	`int32`	Elo of the side to move (900–2600)
`oppo_elo`	`int32`	Elo of the opponent (900–2600)
`value_target`	`int8`	game result from the mover's view — 0 loss, 1 draw, 2 win

Boards are stored from the mover's perspective. move_idx is always present in legal_idx.

How it was produced

Training shards — monthly Lichess standard rated dumps, one parquet per month, via maia3-preprocess with --history 7 --balance. Game and position filtering happens at preprocess time; see maia3/preprocess.py in the repo. Reproduce with:

bash

JOBS=3 data/download_and_rename_datasets.sh

Validation set — the ALLIE test set of Zhang et al. (2025): drop the opening plies, then truncate each game at the first position where the mover has less than the time threshold remaining. 2022-test-annotated.jsonl is the annotated source; the parquet is built from it with:

bash

python data/allie_data/process_allie_data.py \
  data/allie_data/2022-test-annotated.jsonl \
  data/allie_data/test_precomputed.parquet --history 7

Two validation sets

allie_data/test_precomputed.parquet is the primary one and what the configs point at. valid_2019-01_history_7_precomputed.parquet is a complement, not a replacement — same schema, so either loads with no code change:

	ALLIE test set	2019-01 held-out month
positions	884,049	1,000,000
source month(s)	2022	2019-01
relative to training data (2023-01 … 2025-07)	before	well before
clock filtering	openings dropped, truncated at low time	none
measures	in-distribution move prediction	generalization across time

Usage

bash

hf download mcmcmcmc/maia3-training-data --repo-type dataset --local-dir data

That lands the files where the training configs already expect them, so:

bash

maia3-train --config configs/maia3-5m.yaml

To read a shard directly:

python

import pyarrow.parquet as pq
t = pq.read_table("lichess_parquet/train_2023-01_precomputed.parquet")

License

Lichess game data is released under CC0 — these derived shards carry the same terms. The ALLIE test set follows the terms of Zhang et al. (2025).