emiliocantuc/maze-30x30-hard-1k
收藏资源简介:
该数据集名为maze-30x30-hard-1k,用于Tiny Recursion Models (TRM)论文研究。它包含30x30大小的硬迷宫实例,每个迷宫由900个单元格组成(序列长度seq_len=900),词汇表大小为5。数据集包括训练集(8000个示例)和测试集(1000个示例)。每个示例具有三个特征:inputs(uint8类型,表示迷宫输入状态)、labels(分类标签,对应迷宫元素:#墙编码为0、空格/开放单元格编码为1、S起点编码为2、G目标编码为3、o解决方案路径编码为4)和puzzle_index(int32类型,表示谜题索引)。数据集通过官方代码生成,并经过数据增强处理,适用于迷宫求解和递归模型训练任务。
This dataset, named maze-30x30-hard-1k, is used in the Tiny Recursion Models (TRM) paper. It consists of hard maze instances of size 30x30, with each maze comprising 900 cells (sequence length seq_len=900) and a vocabulary size of 5. The dataset includes a training set (8,000 examples) and a test set (1,000 examples). Each example has three features: inputs (uint8 type, representing maze input states), labels (class labels corresponding to maze elements: # wall encoded as 0, space/open cell as 1, S start as 2, G goal as 3, o solution path as 4), and puzzle_index (int32 type, indicating puzzle indices). The dataset is generated using official code with data augmentation and is suitable for maze-solving and recursive model training tasks.




