huihui-ai

/

Llama-3.2-3B-Instruct-abliterated

like 32

Okay, coming right up!

https://huggingface.co/huihui-ai/Llama-3.2-1B-Instruct-abliterated

thanks, but it seems restrictive than 3B version. feels like censored.

Probably because it's dumb. If it's too dumb to have clear internal concepts, then abliteration has nothing to clearly target. This is just my speculation, but I've noticed that many repreng methods fail on tiny models.

As evidence, just look at huihui-ai models.

• 1b model 43.08 -> 38.96 (abliteration confused it, the low score was probably just because it's dumb)
• 3b 50.55 -> 50.73 helped a little
• 8b 52.98 -> 55.42 big jump, it knew the answer it just wasn't saying it, abliteration flips it from helpful>honest to honest>helpful

I would predict bigger jumps on larger models

@wassname thanks for the explanation.