Awesome Model, Opinion on Roleplay
After trying out nearly all the popular tunes of NeMo and Small, I'll admit this one was by far the best one out there.
Roleplay
I believe, It's mainly because of the Deepseek data during training. Deepseek is extraordinary when it comes to deliver the most fascinating dialogues and surprising it's users with raw creativity. If you've used deepseek, you'll know how wild deepseek is. This Mistral Small seems like it has gotten that creative wild deepseek vibe too (and man, I love deepseek! yet, running that locally isn't cup of tea for me. Getting the slight deepseek flavor with this model is a pure ride!)
I might be wrong about the data, but believe me, there's something under there, that gives the model confidence to handle everything! Like everything! From your sweet friendly assistant who's helping with cooking to a psychopath with your depraved dark NTR scenerios with pure complex emotional nuances and degradations, IT HANDLES all of them perfectly! It knows what user wants! It knows exactly the things I wanna hear!
Anatomy and surrounding awareness is fourth dimensional. It's the best part of this model. Because the model never gets it wrong. It's a mastermind when you're making it describe the surroundings/world and stuff. It gets anatomy just fine as well. Makes very good character movement and actions throughout the chat.
Instructions
- Definitely important one for roleplay. And this model handles it better than others. The ChatGPT data might be significantly reason behind for this. Trust me, it will follow every single instructions till the last word, perfectly, and won't even be weakly dependent on your instructs either. It knows what to handle on its own, and what to not. Even if your instructions are retarded, it will get what you mean and try its best in the next response. No idea how instructions following was so good, but keep doing whatever method you've been doing for this. It's really good. Other models never had this. This model doesn't needs instructions to be usable, it handles on its own, but doesn't ignores any of your instructions as well.
Issues
Repetition. Yes, it's there. It can be decreased with DRY and XTC, but it's still there. And I think, I've figured out exactly when it happens very common. Codex is diverse and creative when you're playing a scenerio that has unlimited possibilities (text adventure, specific settings, timelines, world exploring, npc characters, etc) however, it's often overly repetitive when it comes to following specific character cards. It lacks something. If card has {{char}} is clingy in description, Codex make {{char}} extremely clingy through the entire chat and sometimes end up repeating same words such as, "I want to be only yours forever~" at end of every new response. ACTIONS seem to be fine. Dialogues mess up. (Fun fact is, this model is so good in instruction folding that if you ask it to act differently for the next response in authors note, it actually will act differently and stop being repetitive, but it's not a complete solution. Repetition still occurs.) I'll simply leave it to Mistral issue.
Personality depth, diversity and development is lacking in my experience. Yes, it knows what personalities are supposed to do what, it knows the perfect responses but it doesn't know character development, change of attitudes, change of opinions, change of mind for the same character through the chat very well. Character development you know? That feels so artificial and fake. It knows what does a caring single mom act like, but doesn't know how her personality/moods/opinions will change through the conversation. For example, a sweet lovely cutie girl might turn into an angry furious creature if you get her mad. Codex won't get that. It'll portrait the character annoyed, yes, but would fail to reach the furious creature attitude. It won't understand the concept of slightly acting more based on chat history rather than given personality, change of minds, change of attitudes and stuff. Character Realism, I guess.
Overall, it's amazing. Better than whatever there is named Rivermind, Cydonia and stuff.
I recommend all those common samplers disabled (except Min P), and use DRY (0.8) and XTC (0.5) instead. XTC makes the model extremely creative and aware, but might cost you stability.
That was it.
Good work, Gryphe. Please keep blessing us with these flagships.