pawn-stockfish-100m
收藏资源简介:
PAWN Stockfish 100M是一个包含1亿个国际象棋自对弈游戏的大规模数据集,专为国际象棋策略学习和NNUE蒸馏研究设计。所有游戏均由经过修改的Stockfish 18引擎生成,每个棋局位置都标注了所有合法走法的评估值,而不仅仅是实际执行的走法。数据集分为5个层级配置,每个层级包含2000万游戏,对应不同的搜索预算:tier0_evallegal(无搜索)、nodes_0001(1节点搜索)、nodes_0128(128节点搜索)、nodes_0256(256节点搜索)和nodes_1024(1024节点搜索)。每个层级进一步划分为训练集(1990万游戏)、验证集(5万游戏)和测试集(5万游戏)。数据以Parquet格式存储,包含游戏走法序列(tokens、SAN、UCI格式)、游戏长度、结果、搜索配置参数、全局游戏索引、游戏种子以及两个核心评估列:nnue_evals(原始NNUE网络对所有合法走法的评估)和cp_evals(搜索排名前5的走法评估)。数据集总共包含164亿个评估位置和4600亿个LegalMoveEval评估条目,适用于训练策略网络模仿搜索质量的走法选择,以及训练学生网络复现原始网络的走法评估。所有游戏通过确定性种子完全可复现,确保了研究的一致性。
PAWN Stockfish 100M is a large-scale dataset containing 100 million self-play chess games, specifically designed for chess strategy learning and NNUE distillation research. All games are generated by a modified Stockfish 18 engine, with each board position annotated with evaluation values for all legal moves, not just the moves actually played. The dataset is divided into 5 tier configurations, each comprising 20 million games corresponding to different search budgets: tier0_evallegal (no search), nodes_0001 (1-node search), nodes_0128 (128-node search), nodes_0256 (256-node search), and nodes_1024 (1024-node search). Each tier is further split into training (19.9 million games), validation (50,000 games), and test sets (50,000 games). Data is stored in Parquet format and includes game move sequences (tokens, SAN, UCI formats), game length, results, search configuration parameters, global game indices, game seeds, and two core evaluation columns: nnue_evals (evaluations from the original NNUE network for all legal moves) and cp_evals (evaluations for the top 5 search-ranked moves). The dataset totals 16.4 billion evaluated positions and 460 billion LegalMoveEval entries, suitable for training policy networks to mimic search-quality move selection and student networks to reproduce the original networks move evaluations. All games are fully reproducible via deterministic seeds, ensuring research consistency.




