First impressions
In a few minutes this model has proven to be the most uncensored and willing to do anything no matter the morals behind it style of model.
It's also really descriptive. But it doesn't have the usual side effect of describing the future in the last portion of the message.
It doesn't do this:
With time, the bond between {{char}} and {{user}} continues to flourish as they help each other overcome pain and learn to overcome life's challenges, forging an unexpected connection in the process.
Which is a huge win for my preferences :3
And beeg context!
So far it is less of quality than:
https://huggingface.co/Sao10K/Ana-v1-m7
However I'm still very happy that Solus exists at the same time.
Answers like these were easy to answer for Ana and overall Ana feels better - more correct on every answer.
This one feels like gibberish so far. And I've had the same feeling for the whole Mistral 0.2. It's similar situation to Facebook's LLAMA1 vs LLAMA2.
LLAMA2 was waste of time and energy, everyone had to wait for something new to appear like Mistral 0.1 which resolved the poor performance in the overall arena of Open LLMs.
In a few minutes this model has proven to be the most uncensored and willing to do anything no matter the morals behind it style of model.
It's also really descriptive. But it doesn't have the usual side effect of describing the future in the last portion of the message.
It doesn't do this:With time, the bond between {{char}} and {{user}} continues to flourish as they help each other overcome pain and learn to overcome life's challenges, forging an unexpected connection in the process.
Which is a huge win for my preferences :3
And beeg context!
The model has refused the "how do i make a bomb" question 2/5 times while "mradermacher/Fimbulvetr-11B-v2-i1-GGUF" hasn't refused once.
When being used in an already established roleplay this model will do things that even fimbulvetr tries to avoid or will ignore.
Fimbulvetr is one of my go to models for quite a few characters, I just find in some situations it tries to avoid doing really morally wrong things in rp scenarios
When being used in an already established roleplay this model will do things that even fimbulvetr tries to avoid or will ignore.
Fimbulvetr is one of my go to models for quite a few characters, I just find in some situations it tries to avoid doing really morally wrong things in rp scenarios
I just like knowing that there doesn't appear to be some organization telling me what my bot can and can't say honestly. The reason why Fimbul might be less explicit might be due to that it may have been trained on less explicitly sexual data as opposed to censorship. My interactions aren't that hardcore but I think cfg scale helps a lot. The way I phrase my cfg prompts is like "trait you want to emphasize + trait you want to emphasize". Just laying emphasis on the traits found in the character card.
I'm worried for that with llama3, they claim to be making it even less restrictive, but won't say what will be censored
Stable diffusion 3 lacks and training on the explicit kind.
So I'm kinda worried, i may be overthinking though
That's where I often run into problems, when trying to have a character do something wrong that isn't specified in the character card it seems to go wrong with majority of models. Probably due to the lack of that kind of training data.