---
license: cc-by-4.0
language:
  - en
tags:
  - connectomics
  - drosophila
  - connectome
  - text-generation
  - tiny-model
datasets:
  - fernandofernandes/fly-connectome-49k
  - roneneldan/TinyStories
pipeline_tag: text-generation
---

# Fly Wordbrain — rank 32

**A 2,343,125-parameter language model on a frozen fruit-fly connectome — 22.5× smaller
than the model it derives from, and still ahead of it on held-out text.**

This is the smaller sibling of
[fly-wordbrain-rank64](https://huggingface.co/fernandofernandes/fly-wordbrain-rank64),
which is the recommended baseline. Everything about the architecture is identical except
the readout rank: 49,393 → **32** → 1,024 instead of → 64 →.

Research and receipts: [fly-wordbrain](https://github.com/fernando-neto-ai/fly-wordbrain)

## What halving the readout again costs

200 fresh TinyStories, 45,059 next-token targets, selecting no checkpoint.

| Model | Parameters | Audit CE | Audit top-1 | ΔCE vs reference (95% CI) |
|---|---:|---:|---:|---|
| ngxson reference | 52,756,661 | 3.9882 | 31.38% | — |
| rank 64 | 3,956,469 | 3.2802 | 33.37% | −0.708 [−0.738, −0.679] |
| **rank 32 (this)** | **2,343,125** | **3.3091** | **32.58%** | **−0.679** [−0.711, −0.647] |

Relative to rank 64: **1,613,344 fewer parameters** for **0.029 nats** and **0.79
accuracy points**. On the selection-validation split that gap reads as 1.42 points, and
its generated text has more visible failures on generic prompts. Take this one when size
matters more than the last point of quality.

| Component | Parameters |
|---|---:|
| Encoder (embedding + 8 input projections) | 482,816 |
| Neuron gain, recurrent gain, bias | 148,179 |
| Output LayerNorm | 98,786 |
| Factorized readout (rank 32) | 1,613,344 |
| **Total** | **2,343,125** |

**No synaptic weight was modified.** The frozen-buffer digests in `manifest.json` are
byte-identical to the reference model's and to rank 64's — no edge changed, none rewired.

The 49,393 neurons' own dynamics *are* trained, though: 148,179 learned gain / recurrent
gain / bias values, part of the reference architecture rather than an addition here.
Against the seed-rebuilt initialization, `gain` moves **60.57%** in relative L2, `rec_gain`
**13.04%**, and the effective per-neuron scaling `gain × rec_gain` **56.18%**. `rec_gain`
rescales a neuron's whole incoming sum by one factor, so it cannot change the relative
strengths or signs of individual synapses.

Both retained selectors land on update 16,600 here, so `min-ce.safetensors` and
`max-accuracy.safetensors` hold the same weights under two different selection receipts.

## Caveats

The comparison against the released reference is **not** controlled — its training
stories and trainer are unknown. The controlled finding is that constraining the readout
rank acts as a regularizer against overfitting 1,000 short stories; it says nothing about
whether the fly's wiring is a good prior for language. A randomised-graph control has since
measured that directly: rewiring the connectome at random, with every degree and weight
preserved, costs **0.0100 nats [+0.0026, +0.0175]** — about 1% of this architecture's margin
over the reference. The wiring contributes something, and very little of it.

The connectome is not duplicated here; it lives in
[fly-connectome-49k](https://huggingface.co/datasets/fernandofernandes/fly-connectome-49k),
with the complete 166,700-neuron graph at
[fly-connectome-malecns-166k](https://huggingface.co/datasets/fernandofernandes/fly-connectome-malecns-166k).
Loading instructions and the full step function are on the
[rank-64 card](https://huggingface.co/fernandofernandes/fly-wordbrain-rank64).

## 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`](https://huggingface.co/ngxson/fly-llm-hf) architecture. Code MIT.
TinyStories is not redistributed.