Qwen models are not my friends
I am not trying out the first Qwen model, and it seems the Qwen is cannot remain consistent in a given scene, while the llama models expand the action of the given scene, the Qwen models tend to take the dialogue or the action in a direction that is unrealistic in the given situation.
(Example:
User:
"I watch you shave your legs naked in the bathroom"
llm's answer:
"Do you like, what you see? How about we go to kitchen and drink a tea now until it gets cold".)
If I set Qwen model settings to give a more detailed description of the environment, sometimes tends to speak for the user and sometimes immediately started describing full intimate scene.
Of course, this may just be my personal issues (compared to Llama models).
Sao, thanks for all the work you do for the finesse of the models.
For me, this model somehow switched me and char with the first character I tried. After that, I tried some characters that I know work with everything and all the model did was act and talked as me. Kunou also has problems with talking/acting as me, but at least it wasn't happening this much.
Also, some refusals that could be swiped, but still ended up starting the swipe with some sassy "I'm against this, but let's continue then" reply.
Overall, I also agree that qwen models are not my friends.
I'm having repetition issues with ALL Qwen models as well (like it did in the past with smaller models)... Is there something to do about it? NeMo works perfect without DRY or Repetition Penalty...
QWEN produces a lot of slop compared to mistral and llama models
I haven't tested this one, but my favorite Qwen writer is Qwentile 32b