datasets_1
收藏资源简介:
该数据集是一个专门为将棋AI学习而构建的棋盘局面数据集。数据来源于三类原始局面:一是来自nodchip的三个公开数据集;二是源自AobaZero的学习棋谱;三是基于nodchip数据集,使用dlsuisho的策略网络筛选出概率高于10%的候选着法并进行局面扩展后生成的局面。所有局面均经过静止搜索处理,并采用dlsuisho进行了评估值的重新标注。数据集已进行去重和随机打乱处理,确保样本独立性和分布均匀性。数据规模总计为49,594,855,063个局面,适用于将棋AI的监督学习、策略优化或价值评估等任务。
This dataset is a board position dataset specifically constructed for shogi AI training. Its data comes from three categories of original positions: 1) three public datasets from nodchip; 2) learning game records derived from AobaZero; 3) positions generated via expanding the nodchip datasets, where candidate moves with a probability exceeding 10% were screened using the dlsuisho policy network for position expansion. All positions were processed through static search and re-annotated with evaluation values by dlsuisho. The dataset has undergone deduplication and random shuffling to ensure sample independence and uniform distribution. The total number of positions in the dataset is 49,594,855,063, which is applicable to tasks such as supervised learning, policy optimization, and value evaluation for shogi AI.
数据集概述
- 数据集名称:将棋AI学习用局面数据集
- 总局面数:49,594,855,063 个
数据来源
该数据集由以下三种类型的局面构成:
- nodchip氏公开数据集:来自三个公开数据集的局面。
- AobaZero学习棋谱:源自AobaZero学习棋谱的衍生局面。
- 局面生成扩展:基于nodchip氏三个数据集中的局面,使用dlsuisho的策略,对概率在10%以上的候选手进行展开生成。
数据处理
- 评估值替换:对每个局面,使用静止探索和dlsuisho进行评估值的重新标注。
- 去重与打乱:已完成同一局面的去重,并对数据进行随机打乱。




