Okay, here's a review. Sorta.
So, I played with this model for some time, using GGUF Q4 K M (same as Bagel 34b), using it with some cards and story writing. Also, tried iQ4 XS (with that iMatrix stuff) from another dude, but the results were strange, so I dropped that one soon after. This review will be more from an overall 'feeling' perspective, so I may use this model wrong, or this might be just my taste case, or I may be just an idiot, k? :)
• First of all, this model behaves normally after 4k tokens and uses lorebook more often in generations. But may get confused with semi-large cards with a lorebook, using the name of an event or person but hallucinating the rest.
• This model likes to do stuff like: "She answered in a confident tone, leaving no room for ambiguity." - what she replied is a mystery xD
• More hit or miss, middle ground is rare. Also, I noticed that I need to press retry much more often.
• This model is more... well, toxic and evil. Evil characters are much more brutal, and there is more accent on action rather than logic. Even if they are out of place by logic or scenario.
• Kinda had a similar feel with AiDungeon GPT3, before OpenAi dropped their PG-13 nuke on everything, when my action may be not considered successful: Yeah, you did that, but your character is now in even deeper sh... Again, I used all the same settings as I used Bagel.
• Model has trouble picking the format of the chat despite 500+ tokens of example/first message. It just uses something random, like speaking directly as a char, ignoring novel style, narrator, or tone of voice. Some special instructions with examples like living tattoos, that show current emotions through their shape are ignored completely. Funny when an elegant and wise medieval person starts to chat with "xD" and other emojis.
• I feel like there is more weight on negative traits rather than overall character. So I guess that 'toxic' stuff works well too well here.
Have not tried this model for coding... because even GPT4 writes GDScript and shader code badly for Godot 4.
For songwriting, well, asked to create a sarcastic and a bit toxic song with the ABAB structure just for fun, and got toxic AABB with so-so rhythm.
Overall, I feel that I use this model kinda wrong here. It can generate something interesting, but more "miss" than "hit" in my experience.
I hope that this will be somewhat useful. And have a great day! ✨
Thank you for taking the time to review; I'll see what can be improved for the next iteration.