ygkla/zweig-chess-engine-processed

Related Projects

• 🧠 Maia Chess / Cognitive Chess Coach
  https://github.com/ygkali/zweig-chess-engine

This dataset is designed to train human-aligned chess models such as Maia-style ELO-conditioned neural networks.

♟️ Maia Chess 2025: Human-Aligned Chess Dataset

"Predicting human moves, not engine moves."

This dataset contains 11.15 million processed chess games from the Lichess Open Database (Year 2025), specifically curated and segmented to train Human-Aligned AI models (Maia Chess).

Unlike traditional chess datasets that aim for objective optimality (Stockfish evaluation), this dataset captures the subjective, error-prone, and stylistic nature of human play across 12 distinct skill levels.

📊 Dataset Structure & Distribution

The data is partitioned into 12 Skill Buckets based on the players' average ELO rating. This granular segmentation allows for Isolated Training or Curriculum Learning strategies, enabling the model to mimic specific skill levels from "Novice" to "Elite".

ID	Filename	ELO Range	Population %	Description	Skill Level
01	train_01.pgn	400 - 1050	8.3%	Novice	🌱 Beginner
02	train_02.pgn	1051 - 1200	9.6%	Beginner	🌱 Beginner
03	train_03.pgn	1201 - 1325	11.6%	Casual Player	♟️ Casual
04	train_04.pgn	1326 - 1425	11.4%	Lower Intermediate	♟️ Casual
05	train_05.pgn	1426 - 1500	9.1%	Intermediate	⚔️ Intermediate
06	train_06.pgn	1501 - 1575	9.1%	Upper Intermediate	⚔️ Intermediate
07	train_07.pgn	1576 - 1650	8.5%	Advanced Intermediate	⚔️ Intermediate
08	train_08.pgn	1651 - 1750	10.2%	Club Player	🏅 Advanced
09	train_09.pgn	1751 - 1875	9.8%	Strong Club Player	🏅 Advanced
10	train_10.pgn	1876 - 2100	8.8%	Expert	🏆 Expert
11	train_11.pgn	2101 - 2400	3.3%	Master	👑 Master
12	train_12.pgn	2401 - 3000	0.3%	Elite / Super-Human	🤖 Super-Human

⚠️ Data Scarcity Alert: The Elite bucket (ID 12) represents only 0.3% of the total population. Models trained on this bucket may require transfer learning from lower buckets to converge effectively.

🚀 Usage

You can load this dataset in Streaming Mode to avoid downloading the entire 10GB+ archive. This is ideal for Colab or low-disk environments.

from datasets import load_dataset

# Example: Load only the Grandmaster games (Bucket 12)
dataset = load_dataset(
    "ygkali/zweig-chess-engine-processed",
    data_files="train_12.pgn",
    streaming=True
)

print("Streaming games...")
for i, game in enumerate(dataset['train']):
    print(f"Game {i+1}: {game['text'][:50]}...")
    if i == 5: break

##

ZWEIG Chess Engine Dataset is a general-purpose human chess dataset, while Maia Chess models are downstream consumers of this data.

##

🛠️ Methodology
1. Source

    Origin: Lichess Open Database (Jan 2025 - Dec 2025).

    Time Controls: Blitz, Rapid, Classical (Bullet excluded to reduce noise).

2. Preprocessing Pipeline

    Parsing: python-chess library used for PGN parsing.

    Filtering: Games with fewer than 10 moves were discarded.

    Bucketing: Games were assigned to buckets based on the average ELO of White and Black players: (White_ELO + Black_ELO) / 2.

⚖️ License & Citation

This dataset is derived from Lichess (CC0) and is released under the MIT License.
Kod snippet'i

@dataset{maia_chess_2025,
  author = {ygkali},
  title = {Maia Chess 2025: Human-Aligned Chess Dataset},
  year = {2025},
  publisher = {Hugging Face},
  url = {[https://huggingface.co/datasets/ygkla/zweig-chess-engine-processed](https://huggingface.co/datasets/ygkla/zweig-chess-engine-processed)}
}


***