🜲 Gemma-4-31B-it SRT-Sunstone

A read-out that reads images — trained only on words

SRT Sunstone is a live demonstration that a read-out head trained on text alone can interpret images — with zero image training. It is a 12.3 M side-channel head on a frozen google/gemma-4-31B-it, taught one thing from prose: to tell apart communities of language use. Point it at a picture and it names what the image means and the discourse it evokes.

This is the semiotic claim of the SRT (Semiotic-Reflexive Transformer) program made concrete: the read-out interprets a sign by its place in a shared system of meaning, independent of whether the sign arrived as a word or an image.

---

Train once, read everywhere (July 2026)

The program's consolidated finding places this demo inside a larger result: the structure it reads is substrate-invariant. The cross-modal gap inside a frozen multimodal LLM is anisotropic-linear, so a single trained 44 MB linear head turns the same frozen states into full image-text retrieval. And the head keeps working when everything about the deployment changes:

Trained head: RiverRider/srt-sunstone-linear-head. This Space remains the zero-training exhibit: everything below uses only per-modality centering, no trained alignment at all.

---

What you can do here

Pick any category in the gallery. For each one you see:

• The words it means — the image, placed in a shared word space, retrieves its own words: bicycle → bicycle, rose → flower, dog → pet, car → car. This is the precise, load-bearing result.
• The caption it retrieves — the same image state picks a full sentence out of 10,000 COCO captions, open vocabulary, zero training: horse → "a person riding a horse". The image's place in meaning space is precise enough to select a whole description, not just a word.
• The discourse it evokes — the nearest of 35 discourse communities the head learned from text: cars into the automotive community, deer and mushrooms into gardening, cats and dogs into the cozy-domestic communities. A flavour, not a label.

---

Why this is a semiotic read-out, not an image classifier

An image classifier is trained on labelled images and learns a fixed pixel→label map; it only handles the labels in its training set. This read-out is different in kind: its training data was text alone, used to separate discourse communities in the residual stream of a frozen gemma-4-31B. It can read a picture because gemma-4 already fuses image and word into one representational stream, and the read-out taps the shared interpretant — the meaning a sign carries, whatever form it arrived in. It does not classify the image; it tells you what the image means to a system that learned meaning from words.

---

The evidence (held-out)

The head was selected by held-out community accuracy (not training loss), which peaks at step 2250 of 3000 (0.535, an 18.7× lift over chance).

---

How the head was trained

• Backbone: google/gemma-4-31B-it, frozen (never modified).
• Head: 12.3 M parameters (a discourse-community encoder plus MAH divergence heads), reading the residual stream at an early text-tower layer.
• Data: 150 K discourse passages across 35 communities.
• Objective: supervised-contrastive on community id (grouped sampling) plus self-supervised divergence and chain terms. No images, no labels on images, no fine-tuning of the base model.

---

About this Space (precomputed, CPU)

gemma-4-31B is 62.5 GB, far too large to run inside a Space. Every read-out and thumbnail here was precomputed once on a GPU box (scripts/build_crossmodal_gallery.py) with per-modality centering to remove backbone anisotropy. This Space serves the cache, so it runs on free CPU with no model download.

---

Scope & limitations

• A semiotic read-out, not a classifier: community labels reflect discourse association, not object taxonomy, and the head knows only 35 communities — far too few to label the visual world precisely.
• Image↔word retrieval (0.27) is well above chance but lossy: the 64-dim discourse code discards most fine visual detail.
• Assignments are aggregated per category; single 32×32 images are individually noisier.

Links

• Model (read-out head): RiverRider/Gemma-4-31B-it-SRT-Sunstone
• Trained linear head (retrieval): RiverRider/srt-sunstone-linear-head
• Code & paper: https://github.com/space-bacon/SRT
• Base model: google/gemma-4-31B-it

License: Apache-2.0. Base model governed by Google's gemma-4 license.