stoofvlees-data
收藏资源简介:
Stoofvlees训练数据集是为连续赢得2023年和2024年世界计算机国际象棋锦标赛的Stoofvlees引擎生成的国际象棋训练数据。该数据集包含多个大规模文件,总数据量从数百兆字节到上百吉字节不等,具体包括:完整对局数据(无提前裁定或认输)、核心训练数据以及为数据混合实验而转换的Leela Chess Zero训练测试数据。数据中的位置经过二次采样,长残局和和棋局的比例相对较低。每个数据样本是一个固定长度为4545字节的记录,以压缩的二进制格式存储。记录内容详尽,包含:46个8x8棋盘位平面(从行棋方视角表示,包含棋子位置、攻击图等信息)、基于蒙特卡洛树搜索访问计数的策略目标、18个浮点数标量特征(如静态评估、王车易位权、重复局面标志等)、搜索评估分数、从该位置出发的最终对局结果(输、和、赢),以及距离对局结束的剩余步数。该数据集主要用于训练国际象棋AI引擎(如强化学习)以及进行数据混合实验。不同文件遵循不同的开源许可证(CC-BY-4.0或ODbL)。
The Stoofvlees training dataset is chess training data generated for the Stoofvlees engine, which consecutively won the 2023 and 2024 World Computer Chess Championships. This dataset comprises multiple large-scale files, with total data volume ranging from hundreds of megabytes to hundreds of gigabytes. Specifically, it includes complete game data (without early adjudication or resignation), core training data, and Leela Chess Zero training and test data converted for data mixing experiments. The positions in the dataset have been subsampled, with relatively low proportions of long endgames and drawn games. Each data sample is a record with a fixed length of 4545 bytes, stored in a compressed binary format. The record content is comprehensive, including 46 8x8 chessboard bit planes (represented from the perspective of the moving side, containing piece positions, attack maps, etc.), strategy targets based on Monte Carlo Tree Search visit counts, 18 floating-point scalar features (such as static evaluation, castling rights, repetition flags, etc.), search evaluation scores, the final game result from this position (loss, draw, win), and the remaining number of moves until the end of the game. This dataset is mainly used for training chess AI engines (such as reinforcement learning) and conducting data mixing experiments. Different files follow different open-source licenses (CC-BY-4.0 or ODbL).




