bartowski

/

Qwen2.5-Coder-32B-Instruct-abliterated-GGUF

like 16

I suspect that abliteration made it actually work better. I loaded abli of the q4_k_m and in comparison to the non uncensored quant it seems to be outputting more natural results. I am confused a bit

Isn't that the entire point of doing that ? So its not so restricted and does 'better'?

I mean, I thought abliteration is supposed to lift censorship. But it seems to work better even in scenarios where censorship theoretically shouldn't be happening

that is interesting, i wonder if there's something about the abliteration process that happens to prune away some of its circular tendencies