themantralab/gol-emergence-pipeline

GoL Emergence Discovery Dataset

1.5 million Conway's Game of Life initial conditions (16×16 seeds), with full reproducibility metadata plus retained behavioural characterisation. Released by Mantra Labs as the foundation dataset for the GoL Emergence Discovery System — a research program for unsupervised discovery of emergent behaviour in cellular automata.

GitHub (design + code): themantralab/gol-emergence-pipeline

⚠️ Architecture pivot (2026-05-27). The consuming model pivoted from a four-stage learned-transition design with behavioural-signal supervision to an encode/decode-only world model that learns a structured hypershell latent geometry (the GoL engine is the simulator; the model never predicts dynamics). The dataset is unchanged — but how the model relates to each file changed. The model's only direct input is now seeds.npy. The behavioural signatures, class labels, and cluster files are retained for external diagnosis (verifying that emergent latent clusters align with known behaviour) and the FFT reference for future explorer use — none of them are training targets anymore. See DATASET.md and design/ for details.

---

File classification (post-pivot)

File	Shape	Size	Role for the current model
seeds.npy	(1.5M, 16, 16) uint8	~230 MB	CONSUMED — the model's only direct input
seeds.json	—	< 1 KB	CONSUMED — RNG seeds for reproducible trajectory regeneration
lifespans.npy	(1.5M,) int32	5.8 MB	Future — stratified sampling toward long-lived patterns
buckets.npy	(1.5M,) int32	5.8 MB	Future — density-band stratification
sig_reference.npy	(1.5M, 1290) float32	~7 GB	Future — explorer novelty basis (phase-invariant FFT)
grids.npy	(1.5M, 128, 128) uint8	23 GB	Future/convenience — f₀ cache, regenerable from seeds
labels.npy	(1.5M,) str	115 MB	Diagnosis only — old class labels; never a training target
signatures_norm.npy	(1.5M, 257, 10) float32	15 GB	Diagnosis only — per-frame behavioural descriptors
sig_mean.npy / sig_std.npy	(10,) float32	< 1 KB	Diagnosis — normalisation stats for the signatures
n_seeds.npy	scalar	< 1 KB	Metadata

Full rationale in DATASET.md.

---

Dataset at a Glance

Property	Value
Seeds	1,500,000
Grid size	128 × 128 (16 × 16 seed embedded at offset (24, 24))
Timesteps	257 (T = 0 … 256)
Sampling	Density-stratified (4 bands, 0.03–0.30)
RNG seed	3750551643
Generated	2026-04-30
Rule	Conway's B3/S23, fixed-zero boundary

Heuristic class distribution (diagnostic, not a training target)

Class	Count	Share
still_life	815,485	54.4%
oscillator	365,313	24.4%
dying	307,915	20.5%
glider	11,287	0.75%

---

Quick Start

from huggingface_hub import hf_hub_download, snapshot_download
import numpy as np

REPO = "themantralab/gol-emergence-pipeline"

# The model's only direct input
seeds = np.load(hf_hub_download(REPO, "seeds.npy", repo_type="dataset"))  # (1.5M, 16, 16)

# Download everything (≈45 GB)
# snapshot_download(repo_id=REPO, repo_type="dataset", local_dir="./gol-data")

# Diagnostic artifacts (retained, not used for training)
labels = np.load(hf_hub_download(REPO, "labels.npy", repo_type="dataset"))
glider_idx = np.where(labels == "glider")[0]

---

Retained behavioural signatures (diagnostic)

signatures_norm.npy holds a (257, 10) per-frame descriptor trajectory for each seed — population, centre-of-mass displacement, spatial variance, motion energy, connected-component count, and temporal self-similarity at lags 2/4/8/16. These were the supervision targets of the previous design; under the current encode/decode world model they are kept only to evaluate whether the unsupervised latent geometry recovers known behavioural structure. Use sig_mean.npy / sig_std.npy to invert the normalisation.

---

Reproducing the Dataset

The full corpus regenerates deterministically from seeds.npy + seeds.json (RNG seed 3750551643) under B3/S23 with a fixed-zero boundary. Generation tooling will accompany the fresh world-model implementation on GitHub.

---

Citation

@misc{koegler2026gol,
  author    = {Koegler, Maxwell},
  title     = {{GoL Emergence Discovery Dataset}},
  year      = {2026},
  publisher = {Mantra Labs},
  url       = {https://huggingface.co/datasets/themantralab/gol-emergence-pipeline}
}

---

License

Data and figures: CC BY 4.0 — Code: MIT (see GitHub repo)