Datasets:

mireklzicar
/

cellarc_100k_meta

CellARc 100k Meta is the metadata-rich companion to the lightweight CellARc split, retaining every field emitted during generation while staying Hugging Face friendly.

Data quick facts

Vocabulary: cell values use integer labels 0–5 (alphabet size up to 6).
Max list length: every train, query, or solution list has at most 21 entries (median 13).
Flattened full task: concatenating all lists in an episode yields at most 252 integers (median 156).
JSON keys: id, train[].input, train[].output, query, solution, meta, rule_table.

Episode schema sample

{
  "id": "…",
  "train": [
    {"input": [0, 1, …], "output": [2, 0, …]},  # each list length ≤ 21
    …
  ],
  "query": [3, 0, …],      # length ≤ 21
  "solution": [4, 5, …],   # length ≤ 21
  "meta": { … },           # full provenance + analytics
  "rule_table": { … }      # base64-encoded CA lookup
}

Dataset details

cellarc_100k_meta augments the lightweight cellarc_100k release with the complete per-episode metadata that was captured during dataset generation. The Parquet assets (data/*.parquet) are identical between both packages so models can be trained interchangeably; the distinction lives in the JSONL representation:

cellarc_100k/data/*.jsonl → minimal supervision fields (id, train, query, solution).
cellarc_100k_meta/data/*.jsonl → full records containing the metadata, coverage diagnostics and rule tables used to produce each episode.

cellarc_100k_meta/
├── data/
│   ├── train.{jsonl,parquet}
│   ├── val.{jsonl,parquet}
│   ├── test_interpolation.{jsonl,parquet}
│   └── test_extrapolation.{jsonl,parquet}
├── data_files.json
├── dataset_stats.json
├── features.json          # identical to the lightweight package
└── README.md              # this file

Because the JSONL retains the full metadata, the files are substantially larger (≈4.6 × the lightweight versions) but remain line-delimited for streaming.

Additional fields

Each JSON line mirrors the original generation payload:

meta: structured description of the episode (alphabet_size, radius, steps, window, train/query span descriptors, construction, family, Langton λ statistics, entropy bins, coverage metrics, morphology signals, seeds, wrap flags, etc.).
rule_table: explicit lookup table for the cellular automaton rule encoded as base64 (matching the copy nested inside meta).
id: injected fingerprint alias for convenience (meta["fingerprint"]).

The nested objects (e.g. coverage, morphology, train_spans) remain untouched so downstream tooling can parse them without modification.

For the canonical supervision tensors (train, query, solution) refer to the cellarc_100k README.

Working with the metadata

from datasets import load_dataset

data_files = {
    split: f"artifacts/datasets/cellarc_100k_meta/data/{split}.jsonl"
    for split in ("train", "val", "test_interpolation", "test_extrapolation")
}
full = load_dataset("json", data_files=data_files)

episode = full["train"][0]
print(episode["id"], episode["meta"]["lambda"], episode["coverage"]["fraction"])
print(len(episode["meta"]["train_spans"]), "train windows, context =", episode["meta"]["train_context"])

Use cases for the Meta package include:

Analyzing curriculum difficulty via λ/entropy bins or morphology descriptors.
Filtering subsets by construction (cycle, unrolled, hybrid) or family (random, totalistic, …).
Reconstructing the exact CA rule during evaluation (via rule_table).

Shared statistics

dataset_stats.json, features.json and the Parquet files are byte-for-byte identical to those shipped with cellarc_100k, ensuring consistent training/evaluation splits and schema definitions across both packages.

Reproduction & license

Generated via:

python scripts/build_hf_dataset.py --overwrite

License: inherits the repository’s license (specify here if a dedicated dataset license is selected). Please cite the CellARc project when using this dataset.

cellarc_100k_meta