Asilarkness/flybrain-lm

FlyBrain LM — Drosophila whole-brain language model

A character of BPE-level language model whose recurrent core IS the real adult fruit-fly connectome (FlyWire v783, top-48k neurons by out-degree, 6.65M biological synapses). Only synaptic strengths + input/output lenses are trained; wiring topology is frozen fly anatomy.

Status: iter 3000, best (newchat+cot)/2 val 3.329

Files

• model.safetensors — bf16 weights (buffers: Widx topology, params: Wval/Win/emb/lens)
• tokenizer.json — BPE-4k trained on TinyStories
• config.json — architecture sizes
• flybrain.py — standalone inference (torch+safetensors+tokenizers)

Use

from flybrain import load_model, generate
m, tok, cfg = load_model('REPO_ID')
print(generate(m, tok, 'Human: Hi Fly!\nFly:'))