gary-neuron-emergent

The emergence-trained cortex behind neuron-db — a 1.13M-parameter, pure-NumPy GPT that learned to read its context window and copy values out of it. It is the optional "thinking" tier that sits on top of the neuron-db store: the store retrieves a small working set, this cortex generates the answer over it.

What it is

• arch: gpt-numpy — 8 layers, 4 heads, dim 96, 384-token context, vocab 2048
• params: 1,128,384
• trained: step 33,597 (curriculum: copy-from-window + masked answer loss + abstention)
• emergence: in-context QA probe reached 5/6 — it copies unseen values from the window (how many participants? -> 84,512, what is the wifi password? -> vekam73) and learned to abstain ("i don't know right now.") when the answer isn't present.
• val: answer-token loss 0.20; chat perplexity held (val_soda 2.2)

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, gpt_numpy as G            # gpt_numpy.py from gary-neuron-chat / neuron-db
from tokenizers import ByteLevelBPETokenizer

d = snapshot_download("gary23w/gary-neuron-emergent")
z = np.load(f"{d}/cortex.npz", allow_pickle=True)
P = {k[2:]: z[k] for k in z.files if k.startswith("P/")}
CFG = dict(E=96, H=4, L=8, BLK=384)
tok = ByteLevelBPETokenizer(f"{d}/petite_vocab.json", f"{d}/petite_merges.txt")

# feed it a working set + a question, greedily decode:
ids = tok.encode("U: the launch is on Friday\nG: noted.\nU: when is the launch?\nG:").ids
# ... forward + argmax loop -> "Friday"

Or point neuron-db's bridge at a checkout:

export NEURON_MODEL_DIR=<this download dir>
python -c "from neuron_db.bridge import GaryNeuronBridge; b=GaryNeuronBridge(); ..."

Honest notes

• It was trained on a ~2k everyday-token vocabulary in a U:/G: fact format. It excels at copying normalized facts out of a bounded window; it is not a general chatbot.
• For exact recall you don't need it at all — neuron-db's store returns the value deterministically. This cortex is for generation/association over the working set.

MIT. Part of the gary-neuron family by gary23w.