gary-neuron-emergent

The dispatcher cortex behind neuron-db — now a ~6.97M-parameter, pure-NumPy GPT that reads its context window and decides what to do with each turn. It's the optional "thinking" tier on top of the neuron-db store: the store retrieves a small working set, this cortex routes over it.

🌐 Live site & in-browser demos: https://gary23w.github.io/neuron-db/

v5 — a larger, better-grounded model than the original 1.13M emergent checkpoint. Same pure-NumPy format; new weights, tokenizer, and config.

What it does — one route per turn

route	meaning
ANSWER <value>	the working set has it — copy the value straight out
ESCALATE	the working set can't answer — hand up to a larger host model
FETCH <topic>	a live-world question — go to the web
STORE <fact>	a declarative — remember it

It also chains multi-fact lookups (e.g. "who is the manager of Aurora's owner?").

What it is

• arch: gpt-numpy — 8 layers, 8 heads, dim 256, 512-token context, vocab 2048
• params: 6,973,952
• trained: ~23,784 steps (12 epochs), final train loss 0.006

Held-out results

Fresh values (a different random stream than training), greedy decode. ANSWER accuracy vs. the size of the working set — the hard axis, since more facts in the window means more distractors:

working set	1	6	12	18
ANSWER correct	100%	96%	98%	76%

Route-action accuracy: ESCALATE 96% · FETCH 100% · STORE 100% · multi-hop ANSWER 100%.

Files

cortex.npz            float32 weights (P/ keys; gpt-numpy layout)
petite_vocab.json     byte-level BPE vocab (2048)
petite_merges.txt     BPE merges
config.json           E/H/L/BLK, vocab, param count, trained step

Use

from huggingface_hub import snapshot_download
import numpy as np

d = snapshot_download("gary23w/gary-neuron-emergent")
z = np.load(f"{d}/cortex.npz")
P = {k[2:]: z[k] for k in z.files if k.startswith("P/")}   # tensor name -> float32 array
CFG = dict(E=256, H=8, L=8, BLK=512, vocab=2048)
# pre-LN transformer, learned positions, tanh-GELU, weight-tied output;
# greedy-decode over a "U: <fact>\nG: noted.\n... U: <question>\nG:" working set.

The same weights also ship int8-quantized and baked into the neuron-db Rust crate (and its WASM build), where route() does recall + dispatch in one call — see the repo.

Honest notes

• Trained on a ~2k everyday-token vocabulary in a U:/G: fact format. It excels at routing and copying normalized facts out of a bounded window; it is not a general chatbot.
• Large working sets are the hard axis: ANSWER accuracy stays high to ~12 facts and tapers by ~18.
• For exact recall you don't need the model at all — neuron-db's store returns the value deterministically. The cortex is for the route decision and generation/association over the set.

MIT. Part of the gary-neuron family by gary23w. Site: https://gary23w.github.io/neuron-db/