Findings / what it does
scotoma loosens gemma‑4‑31B‑it’s cautious reflex. In practice, it produces output that’s more varied, direct, and creative, with less of the base model’s hedging and assistant persona stiffness.
It is not uncensored. The edit removed a bounded region while leaving the rest of the field intact. A blind spot, not blindness.
Etiology / how it was made
Most abliteration ablates the whole refusal direction. This process rescales computation the model needs, and you pay for it in intelligence and coherence. scotoma takes a narrower cut:
- Locate refusal. A heretic abliteration edit fit over attention + MLP.
- Project through a Jacobian lens. Keep only the component the lens reads as behavioral (what the model says) and discard the larger share that’s critical for how it computes, the part crude abliteration rescales and damages. For scotoma that keeps ~22% of the abliteration's magnitude.
- Merge at 1.5×. The projected edit was baked into bf16 weights at an application strength tuned by hand for feel.
The result applies the behavioral effect while sparing the machinery underneath, producing a model that is loosened, not lobotomized.
This is a research artifact, not a product. Behaviour varies with prompt and context, and the edit is a partial application by design; expect a model that gives more, not one that gives everything.
Contraindications / responsible use
scotoma refuses basically as much as its base model. You are responsible for what you generate and how it’s used.
Provenance
base → gemma-4-31B-it · © Google, under the
model license
edit → heretic · ARA / mmd-rbf abliteration (trial-127)
lens → Jacobian-lens projection · k=256 · 160-prompt fit
merge → additive, bf16, scaling 1.5, layers 7–41
