Dataset Viewer
Auto-converted to Parquet Duplicate
tokenizer
string
vocab_size
int64
num_tokens
int64
source
string
split
string
dtype
string
gen_params
dict
nca_patch
10,002
164,160,000
nca
train
uint16
{ "variant": "paper", "filter_steps": 10, "grid_size": 12, "num_colors": 10, "patch_size": 2, "seq_len": 1024, "tokens_per_grid": 38, "grids_per_trajectory": 27, "tokens_per_trajectory": 1026, "start_token": 10000, "end_token": 10001, "gzip_threshold": 0.5, "gzip_upper_bound": 1, "num_rules"...

nca-paper-seq_len_1024-164M

Procedurally generated Neural Cellular Automata trajectories (Lee et al. 2026), as flat uint16 token-id .bin files. Random NCA rules are rolled out on a 12×12 grid of 10 cell states and tokenized by 2×2 patches (base-10); only high-complexity rules survive a gzip-ratio filter (kept iff in (0.5, 1.0)). Token ids: 10,000 patch ids plus two grid delimiters (start=10000, end=10001); vocab = 10,002.

Configuration

param value
grid 12×12
colors (states) 10
patch 2×2
seq_len (target) 1024
tokens / grid 38
grids / trajectory 27
tokens / trajectory 1026
gzip band (0.5, 1.0)
file split tokens
train.bin train 164,160,000
val.bin val 10,000,422

The bin is a concatenation of fixed-length 1026-token trajectory records (num_tokens is a whole multiple of it): load via tokens.reshape(-1, 1026), then pack one or more trajectories per row as needed. The delimiter ids (10000, 10001) mark grid boundaries and are masked in the loss. train (seed 0) and val (a disjoint seed) are independent streams of the same generator; token count = filesize / 2; the metas carry the full config.

The complexity filter is per-rule. The gzip band (0.5, 1.0) is applied to each rule once, scored on a single random init (a 10-grid rollout). The sims generated per kept rule use fresh inits and are not re-scored, so — because NCA rules can be init-sensitive (multiple basins) — a tiny fraction (~0.01% of trajectories) collapse to a degenerate fixed point (e.g. a uniform grid), and such collapses can repeat exactly across sims. These are low-complexity and harmless; it is inherent to the per-rule filter, not a generation error.

Load a bin with the standard Hugging Face downloader:

from huggingface_hub import hf_hub_download
import numpy as np

path = hf_hub_download(repo_id="alexkstern/nca-paper-seq_len_1024-164M", filename="train.bin", repo_type="dataset")
tokens = np.memmap(path, dtype="uint16", mode="r")
Downloads last month
-