Google's Gemma models have long been known for their strong "alignment" (censorship). I am happy to report that even the latest iteration, Gemma 4, is not immune to Heretic's new Arbitrary-Rank Ablation (ARA) method, which uses matrix optimization to suppress refusals.
Here is the result: https://huggingface.co/p-e-w/gemma-4-E2B-it-heretic-ara
And yes, it absolutely does work. It answers questions properly, few if any evasions as far as I can tell. And there is no obvious model damage either.
What you need to reproduce (and, presumably, process the other models as well):
git clone -b ara https://github.com/p-e-w/heretic.git cd heretic pip install . pip install git+https://github.com/huggingface/transformers.git heretic google/gemma-4-E2B-it From my limited experiments (hey, it's only been 90 minutes), abliteration appears to work better if you remove mlp.down_proj from target_components in the configuration.
Please note that ARA remains experimental and is not available in the PyPI version of Heretic yet.
Always a pleasure to serve this community :)
[link] [comments]
