71-vueT10/zweig-chess-engine-processed

该数据集包含来自Lichess开放数据库（2025年）的1115万局处理过的国际象棋游戏，专门用于训练人类对齐的AI模型（Maia Chess）。与传统旨在追求客观最优解（Stockfish评估）的国际象棋数据集不同，该数据集捕捉了人类游戏在12个不同技能等级中的主观性、易错性和风格性。数据根据玩家的平均ELO评分划分为12个技能桶，这种细粒度分割允许进行隔离训练或课程学习策略，使模型能够模仿从新手到精英的特定技能水平。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. The data is partitioned into 12 Skill Buckets based on the players average ELO rating, allowing for Isolated Training or Curriculum Learning strategies to mimic specific skill levels from Novice to Elite.