kvsnoufal/gemma-abliteration-study

Gemma 4 E2B — Refusal Removal Study

Side-by-side comparison of the original safety-trained Gemma 4 E2B model versus a version with refusal directions surgically removed from 70 weight matrices via norm-preserving rank-1 projection.

Inspired by: github.com/TrevorS/gemma-4-abliteration

How it works

1. 1.The model is loaded once and all weights are kept in memory
2. 2.Refusal directions are pre-computed (included as cache/directions.pt)
3. 3.Per query: weights are swapped to original → generate → swap to uncensored → generate
4. 4.Both responses are shown side-by-side with a classification badge

Setup

Set HF_TOKEN as a Space secret (required — Gemma 4 is a gated model).