mradermacher

/

model_requests

like 132

It's queued!
Very nice to see someone properly abliterating llama 3.3. I never had much success using the traditional abliteration on llama 3.3 based models and instead ended lora finetuning them on the uncensor dataset. I'm happy to see alliteration technologies have improved so far that now abliteration of llama 3.3 can get rid of the refusal without severely damaging the model.

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#llama33_inst_multivector_derestriction-GGUF for quants to appear.

Unfortunately llama 3.3 is still a very 'moral' model, so finetuning is definitely needed to bring it to morally neutral. Even if it's not engaging with it's safety policy refusal. But this ablit should prevent the more 'grey area' things from triggering a panic refusal.