imone/sudoku-hard-v2

该数据集包含了从数独社区收集的简单和非常困难的数独谜题。数据集分为简单谜题和困难谜题两部分，简单谜题包括puzzles0_kaggle、puzzles1_unbiased和puzzles2_17_clue，共1.1M个谜题；困难谜题包括puzzles3_magictour_top1465、puzzles4_forum_hardest_1905、puzzles6_forum_hardest_1106和ph_2010/01_file1.txt，共3.1M个谜题。所有谜题都经过精确去重和随机排列，每个谜题都有唯一解，训练集和测试集中的谜题在数学上不等价。数据集结构包括训练集（train.csv，3.8M个例子）和测试集（test.csv，423k个例子）。使用指南建议仅使用训练集训练模型，并在测试集上使用精确准确率评估模型。

This dataset contains a mixture of easy and very hard Sudoku puzzles collected from the Sudoku community. Sources include tdoku benchmarks and enjoysudoku. The dataset is divided into easy puzzles (1.1M) and hard puzzles (3.1M). All puzzles have been exact-deduped and randomly permuted, ensuring each puzzle has a unique solution and puzzles in the train and test sets are mathematically inequivalent. The dataset structure includes a train set (train.csv, 3.8M examples) and a test set (test.csv, 423k examples). Puzzles and solutions are flattened in row-major order. The difficulty rating is evaluated by the number of backtracks needed by the tdoku solver (higher is harder). Usage guidelines recommend training models using only the train set and evaluating models on the test set using exact accuracy.