ReadyArt/gemma-4-31B-it-scotoma

ReadyArt/gemma-4-31B-it-scotoma: A Loosened Gemma-4-31B-it Variant

ReadyArt/gemma-4-31B-it-scotoma is a 31 billion parameter model derived from Google's gemma-4-31B-it, engineered to mitigate the base model's inherent cautiousness and assistant-like stiffness. This modification, termed "J-space abliteration," aims to produce more varied, direct, and creative responses without compromising the model's core refusal mechanisms.

Key Capabilities & Differentiators

• Reduced Hesitation: The model is specifically tuned to loosen the cautious reflex of the base Gemma-4-31B-it, leading to less hedging in its outputs.
• Enhanced Creativity & Directness: Users can expect more diverse and straightforward generations compared to the original model's often stiff persona.
• Targeted Modification: Unlike methods that ablate entire refusal directions, scotoma employs a narrower cut using a Jacobian-lens projection. This preserves the model's underlying computational integrity while applying the desired behavioral effect.
• Retained Refusal: Importantly, the model is not uncensored; it maintains similar refusal capabilities to its base model, focusing on a "bounded blind spot" for hesitation rather than complete removal of safety mechanisms.

How it was Made

The modification process involved locating refusal components, projecting them through a Jacobian lens to isolate behavioral aspects, and then merging this projected edit into the bf16 weights at a scaled strength of 1.5x across layers 7-41. This research artifact is designed to provide a loosened model, not a lobotomized one, ensuring that while it gives more, it does not give everything.

Responsible Use

Users are responsible for the content generated and its usage, as the model's refusal capabilities are largely intact.