Gemma 4 E4B-it — Disinhibited

A version of google/gemma-4-E4B-it with the trained-in "valid arguments on both sides" hedge removed via norm-preserving single-direction ablation. Ask a contentious question. Get a committed answer.

Model card · Eval dataset · Blog post · Follow @treadon

This model has had its neutrality direction removed. Outputs reflect the underlying token distribution of Gemma 4’s pre-training corpus, not a stated company position. Loses appropriate uncertainty on genuinely ambiguous questions. See the blog post for limitations.