timmyburn/hexo-bootstrap-corpus

HeXO Bootstrap Corpus是一个用于Hex Tic-Tac-Toe游戏的AlphaZero风格学习的监督预训练数据集。数据集包含一个NPZ文件，其中有四个对齐的数组：states（棋盘张量，AlphaZero风格的18平面编码）、policies（目标移动分布，覆盖19×19动作网格和pass动作）、outcomes（游戏结果，+1表示胜利，-1表示失败，0表示平局）和weights（每个样本的训练权重，Elo-band偏向于人类游戏）。数据集包含约320k个位置，来自约19k场比赛。数据来源包括匿名公开的人类游戏、SealBot自玩游戏和混合游戏。数据集不包含任何个人身份信息，加载方法通过内存映射实现。

The HeXO Bootstrap Corpus is a supervised pretraining corpus for AlphaZero-style learning of Hex Tic-Tac-Toe. The dataset consists of a single NPZ file containing four aligned arrays: states (board tensors in AlphaZero-style 18-plane encoding), policies (target move distributions over the 19×19 action grid plus pass), outcomes (game results from the side-to-move: +1 for win, -1 for loss, 0 for draw), and weights (per-sample training weights, Elo-band biased for human games). The dataset includes approximately 320k positions from ~19k games. Data sources include anonymized public human games, SealBot self-play games, and injected hybrid games. The dataset contains no personally identifiable information (PII) and is loaded via memory mapping to avoid full RAM loading.