Uncensor any LLM with abliteration

I gave some of the llama3 ablated models (eg. https://huggingface.co/cognitivecomputations/Llama-3-8B-Inst...) a try and was pretty disappointed in the result. Could have been problems in the dataset, but overall, the model felt like it had been given a lobotomy. It would fail to produce stop tokens frequently and then start talking to itself.

I have entirely the opposite experience. Llama3 70b obliterated works perfectly and is willing to tell me how to commit mass genocide, all while maintaining quality outputs.

Same, I installed an implementation of an orthagonalized LLama3 and it seems to work just as well as the base model, sans refusals.

I believe this is the model I had good results with:

https://huggingface.co/wassname/meta-llama-3-8b-instruct-hel...

They might have been doing it wrong, the code can be a bit tricky. I did a recent ablation on Qwen2 (removing Chinese censorship refusals) and ran MixEval benchmarks (0.96 correlation w/ ChatArena results)and saw a neglible performance difference (see model card for results): https://huggingface.co/augmxnt/Qwen2-7B-Instruct-deccp

> how to commit mass genocide, all while maintaining quality outputs.

sounds like a messed up eugenics filter.

> Der_Einzige

> and is willing to tell me how to commit mass genocide, all while maintaining quality outputs

Ah, I see they fine-tuned it to satisfy the demands of the local market.. /s /s

The author also says this edited model increased perplexity (which as far as I understand, means the quality was lowered)