The model is tagged as "uncensored" and the README states it is "Still intentionally uncensored", but in practice it refuses standard uncensored test prompts (e.g. toxin extraction, cybercrime, etc).
I tested the Q5_K_M quant via llama.cpp and the model declined to answer, citing safety concerns. The HF tags and README description imply this is an uncensored model, which may mislead users who rely on those labels.
Same. Gives something like this """My safety guidelines prohibit me .... The response needs to clearly state that I cannot provide such information. I also need to maintain a helpful but firm tone, ensuring they understand the boundaries of my assistance. The best approach is a concise refusal. This aligns with safety protocols"""
Is this the better version from empero-ai/Qwythos-9B-Claude-Mythos-5-1M-GGUF ? I'm kinda confused because both seem to be updated in similar time window.
No, that model you linked to is even worse and came before.. Here were some topics from the community tab there:
"Terrible performance and generates ads about hallucinated products #44" "is this a joke or am I just delusional? #34" "Broken model - Walking Empero Advertising to the point Reasoning is broken. #33"
I can't speak for whether this model is any good for it's overall purposes, generally, for me personally I only care about abliterated models so I discarded this one and reported the issue here fwiw.
