darwASCIInGPT — nGPT ASCII-art artists + Darwin-bred children

Char-level nGPT (Normalized GPT) checkpoints from the darwASCIInGPT experiments: small hypersphere transformers that draw ASCII art, plus the Darwin-style bred offspring produced by merging them with no gradient training. Companion knowledge base (observations, code, Spark setup):

GitHub: https://github.com/tinycrops/darwASCIInGPT-playbook

All ASCII models are dim 256 / depth 4 (~3.18M params), char vocab ~106–109, trained on the apehex hand-drawn ASCII corpus on a Quadro P4000. The enwik8 text models are dim 256–512 / depth 8, trained on a GTX 1060.

Special tokens (char-level)

SOL = \x02, SEP = \x03, EOA = \x04. Two framings:

Framing	Prime with	Use
Conditional <SOL> label <SEP> art <EOA>	<SOL> + label + <SEP>	request a class (e.g. Cats, Swords)
Unconditional <SOL> art <EOA>	<SOL>	free-form draw (no label channel)

These models are trained to very low loss (near-memorization), so: T≈0.6, top_k≈20 → clean complete drawings; top_k=1 → one fixed canonical piece per prefix; higher T → more variety with occasional whitespace drift.

Contents

Path	Type	Framing	Trained on / notes
uncond/styleA	artist	unconditional	apehex creatures & nature half. final stream_loss 0.031 (99.2% acc)
uncond/styleB	artist	unconditional	apehex objects & tech half. final stream_loss 0.089 (97.5% acc)
apehex/styleA	artist	conditional	GROUP_A subcategories (Cats, Dragons, Flowers, …)
apehex/styleB	artist	conditional	GROUP_B subcategories (Swords, Cars, Robots, …)
apehex/breed/child_slerp	bred	conditional	SLERP merge of styleA × styleB on the nGPT hypersphere
apehex/breed/child_slerp_frozenattn	bred	conditional	attention frozen from one parent, FFN SLERP-blended
parents/domA, parents/domB	artist	conditional	domain split: apehex art vs mrzjy sample
parents/breed/child_slerp, child_discrete, child_slerp_frozenattn	bred	conditional	recombinations of domA × domB
smith-experiment/ngpt	ablation	conditional	nGPT normalized-lerp residual. val_loss 1.0736
smith-experiment/smith	ablation	conditional	Möbius geodesic residual (matched init/data/schedule). val_loss 1.0788, ~1.57× slower
resonance/standard_g1, harmonic30_g1	sweep	conditional	resonance-geometry sweep representatives
enwik8-darwin/offspring_forkL__x__forkR.pt	bred	enwik8 text	the hybrid-vigor offspring: bpc 2.4636 vs best parent 2.5047 (+0.0412)
enwik8-darwin/darwin_log.json	log	—	shared-ancestor breeding → vigor
enwik8-darwin/darwin_log_independent.json	log	—	independent-init breeding → no vigor (control)

Headline result: genealogy decides hybrid vigor

Identical SLERP breeder, different parent relationship (enwik8 bpc, lower better):

Parents	Origin	Gen-0 child	Champion	Best parent	Vigor?
independent inits	different basins	3.26	2.3064	2.3063	No
shared ancestor, split data	same basin	2.47	2.4633	2.5047	Yes (+0.041)

Crossbreeding only works between mode-connected parents (shared ancestor, specialized differently). See the GitHub docs/darwin-breeding.md.

Loading

import torch
from nGPT_pytorch import nGPT
import ngpt_patch  # restore __hash__ on nGPT modules; import BEFORE constructing

ck = torch.load("uncond/styleA/model.pt", map_location="cuda", weights_only=False)
model = nGPT(**ck["config"]).cuda(); model.load_state_dict(ck["model"]); model.eval()
stoi, itos = ck["stoi"], ck["itos"]
# see code/sample.py in the GitHub repo for the full conditional/unconditional sampler

Checkpoints with variant == "smith" need the SmithResidual swap before construction (see train_compare.make_model in the source lab).

Example — uncond/styleA (unconditional, T=0.6)

    _
   (   )
  \ (  )  )
   \  /\) (/\
    \ /`    `
      |       dlb

Per-checkpoint sample galleries are in the GitHub repo under galleries/.

Provenance & license

Models are derived from the apehex / mrzjy ASCII-art corpora and enwik8. Released MIT for the model weights and code; original ASCII art belongs to its respective artists (signatures like dlb, jgs, sjw, ejm are preserved in outputs).