https://huggingface.co/grimjim/gemma-3-12b-it-biprojected-abliterated and https://huggingface.co/grimjim/gemma-3-12b-it-projection-abliterated

#1497
by mingyi456 - opened

Apparently there is a new refinement to the abliteration technique, and these models show impressive results on the UGI leaderboard.
https://huggingface.co/grimjim/gemma-3-12b-it-biprojected-abliterated
https://huggingface.co/grimjim/gemma-3-12b-it-projection-abliterated

I just had a nice read of https://huggingface.co/blog/grimjim/projected-abliteration

However, for the purposes of abliteration, where the refusal direction is to be ablated to allow a model to comply with a harmful prompt, removing a push away from compliance is ungrounded as compliance is the goal. Removing the component which comprises a push away from correct compliance, however, has no theoretical justification.

I hope they mean that they still want the model to comply with the users prompt and not comply with any safety garbage. Let's quant them and give it a try. I already tested hundreds of uncensored/alliterated models and created many myself so this is one of the few occasions it is likely worth it for me to share my opinion about it once I extensively tested it.

They are queued! :D

You can check for progress at http://hf.tst.eu/status.html or regularly check the model summary page at https://hf.tst.eu/model#gemma-3-12b-it-biprojected-abliterated-GGUF and https://hf.tst.eu/model#gemma-3-12b-it-projection-abliterated-GGUF for quants to appear.

Turns out mradermacher already did https://huggingface.co/grimjim/gemma-3-12b-it-projection-abliterated a week ago but the other is now inside the queue.

Sign up or log in to comment