fernandofernandes/fly-wordbrain-shuffled-10k

Fly Wordbrain — randomly rewired connectome (control)

⚠️ These weights were trained on a graph that does not exist in any dataset. Loading them onto the measured fly connectome gives a model that was never trained, and any score you get from that is meaningless. Rebuild the control graph first — see Using it correctly.

This is a control arm, not a model to use. It exists to answer the one question six stages of ablations could not: does the fruit fly's actual wiring matter, or would any sparse recurrent layer of the same shape do?

It is fly-wordbrain-rank64's 10,000-story arm in every respect — same 3,956,469 trainable parameters, same corpus, same 16,800 updates, same seed 42, same optimizer and schedule — trained on the same connectome with every edge rewired at random.

What it answered

200 held-out stories, 45,059 next-token targets, a population that selected no checkpoint:

	Parameters	Audit CE	PPL	Top-1
released reference	52,756,661	3.9882	53.96	31.38%
this control (rewired)	3,956,469	2.9593	19.29	36.52%
measured connectome	3,956,469	2.9493	19.09	37.43%

Paired over the same stories (10,000 whole-story resamples, seed 1729), rewired minus measured:

	Δ	95% CI
cross-entropy	+0.0100 nats	[+0.0026, +0.0175]
top-1 accuracy	−0.91 points	[−1.20, −0.63]

Both intervals exclude zero, so the fly's specific wiring carries real signal — worth about 1% of the model's advantage. The measured-connectome arm beats the released reference by 1.039 nats; the anatomy accounts for 0.0100 of that. A fly brain rewired at random still beats the reference by 1.029 nats.

One asymmetry: the wiring is worth 15.1% of the accuracy gain but 0.96% of the cross-entropy gain. It sharpens the top-1 pick more than it improves the distribution.

What "rewired at random" preserves

A shuffle that also damaged the degree distribution would prove nothing — a null result could be blamed on the damage. So the control holds every connectivity statistic fixed:

preserved exactly	destroyed
every neuron's in-degree (row offsets never written)	which particular neuron connects to which
every neuron's out-degree
every synaptic weight, and each neuron's incoming weight multiset
the input-injection and readout interfaces

Measured on the real graph: 99.9964% of 9,050,172 edges rewired, 118,705 (1.31%) surviving at both endpoints by chance, 0 self-loops, 0 duplicate edges, in-degree and out-degree identical.

Using it correctly

The graph is a deterministic function of the published connectome and one integer:

# scripts/train_connectorch_control.py in the repository
indices, repairs = control_indices(offsets, source, "shuffle", 1729)

Rebuild it from fly-connectome-49k, install it on the reference before constructing the model — the model snapshots its graph and refuses any later edit — then load these weights. manifest.json records control_graph_w_indices_sha256; if your rebuilt graph does not hash to it, stop.

Synaptic weights are byte-identical to the measured connectome's. Only the index array moved.

Files

min-ce.safetensors and max-accuracy.safetensors (both update 16,800), manifest.json, graph-control.json (the rewiring receipt written at training time), and experiment-config.json.

Caveats

One seed and one shuffle. A second seed would show how much of the 0.0100 nats is run-to-run noise. Not converged — 16,800 updates is 1.83 passes over 10,000 stories, so the trainer records debug_stopped with debug: true; that flags the cap, not a failure. No zero-edge control, so nothing here bounds what the recurrence itself contributes: both arms keep all 9,050,172 edges.

Full write-up: Stage 7.

Licence

Weights CC BY 4.0 — derived from MaleCNS v1.0 (FlyEM / HHMI Janelia, University of Cambridge, MRC LMB, Google Research) and the ngxson/fly-llm-hf architecture. Code MIT. TinyStories is not redistributed.