Module holography

Each of the 248 targeted modules was tested alone; a single module's slice of the weight change classifies a trait's Big Five factor at 0.638 on average against 0.202 chance, where all 248 together reach 0.76.

Module holography

The test

Each targeted module was tested on its own: can that module's slice of the weight change, compressed to 32 dimensions, say which Big Five factor a trait belongs to? qwen35/build_findings_page.py, finding 1:

Personality is holographic. One module out of 248 knows almost as much as all of them.

The numbers

qwen35/analysis/module_holography.json, top-level keys:

So all 248 modules together reach 0.76 on a five-way factor classification; a single module averages 0.638 and ranges 0.43 to 0.74; chance is 0.202. Adding 247 more modules buys about 0.12. The trait is not stored in one place; it is smeared across all of them, redundantly.

#modules is a 248-entry list, each with i, mod (the full PEFT module path), acc, norm, cls and layer. Module classes present: linear-attn, MLP, full-attn. The best single modules stored are base_model.model.model.layers.15.mlp.down_proj and base_model.model.model.layers.15.self_attn.o_proj, both at acc = 0.74; the weakest is base_model.model.model.layers.6.linear_attn.in_proj_a at acc = 0.43.

build_findings_page.py's reading of the per-class and per-depth pattern: "MLP modules edge out attention, and later layers edge out earlier ones, but every class and every depth carries the signal."

Provenance

No producing script for module_holography.json exists in this repo. The file's own keys and qwen35/build_findings_page.py (which renders it as D.holo and D.holo_modules via qwen35/analysis/distil_data.json) are the whole record. The 0.202 chance figure matches the shuffled-null value used elsewhere for five-way factor classification (qwen35/analysis/umap_test.json#factor(5-way).shuffled_null = 0.195, #chance = 0.2), and the 32-dimensional compression is the k = 32 sketch, so the method is legible from context; the code is not on disk. The file was written 2026-08-29 and findings_page is one of the older built pages, so treat the framing as historical even though the numbers themselves are not contradicted anywhere.

Related: Polarity, bipolarity and the trait graph, Weight-space geometry of the 134-adapter zoo, UMAP, sphere and other layouts.