gemma4uncensor. hands-on study of how to reduce model refusals (“uncensor” in the sense of weight edits, not retraining) for Gemma 4 E2B IT by estimating a per-layer refusal-related direction from activations and projecting it out of o_proj and down_proj while preserving row norms

github.com/kvsnoufal/gemma4uncensor

Vaya's read on this project

Problem, audience, market, and the verdict — sign in to see it.

Updates

No recent activity.