alexkstern/nca-paper-share10-seq_len_1024-164M
收藏资源简介:
该数据集是通过程序生成的神经细胞自动机轨迹,以平坦的uint16 token-id .bin文件格式存储。随机神经细胞自动机规则在12×12网格上展开,包含10种细胞状态,并通过2×2补丁进行标记化(基数为10)。仅通过gzip比率过滤器保留高复杂性规则(保留比率在0.5到1.0之间的规则)。token ID包括10,000个补丁ID和两个网格分隔符(起始=10000,结束=10001),词汇表大小为10,002。数据集分为训练集和验证集,训练集包含164,160,000个token,验证集包含10,000,422个token。每个轨迹记录固定为1026个token,包含网格边界的分隔符ID(在损失计算中被屏蔽)。训练集和验证集来自同一生成器的独立流,token数量等于文件大小除以2。复杂性过滤器基于每个规则应用,gzip波段(0.5, 1.0)对每个规则评分一次,基于单个随机初始化(10网格展开)。每个保留规则生成的模拟使用新的初始化,不重新评分,因此由于神经细胞自动机规则可能对初始化敏感(多个盆地),极小部分轨迹(约0.01%)可能退化为固定点(如均匀网格),这些低复杂性轨迹无害,是每个规则过滤器的固有特性,而非生成错误。
This dataset consists of programmatically generated neural cellular automaton (NCA) trajectories, stored as flat uint16 token-id .bin files. Random neural cellular automaton rules are evolved on 12×12 grids with 10 cell states, and tokenized via 2×2 patches (radix 10). Only high-complexity rules are retained via a gzip ratio filter, with rules with a ratio between 0.5 and 1.0 kept. The token IDs include 10,000 patch IDs and two grid separators (start = 10000, end = 10001), resulting in a vocabulary size of 10,002. The dataset is split into training and validation sets: the training set contains 164,160,000 tokens, while the validation set contains 10,000,422 tokens. Each trajectory record is fixed at 1026 tokens, including separator IDs for grid boundaries which are masked during loss calculation. The training and validation sets are derived from independent streams of the same generator, and the token count equals the file size divided by 2. The complexity filter is applied on a per-rule basis: each rule is scored once against the gzip band (0.5, 1.0) based on a single random initialization with 10 grid evolutions. Simulations generated for each retained rule use a new random initialization and are not re-scored. As neural cellular automaton rules may be initialization-sensitive (with multiple basins of attraction), a tiny fraction of trajectories (~0.01%) may degenerate into fixed points such as uniform grids. These low-complexity trajectories are harmless and an inherent property of the per-rule filter, rather than a generation error.




