Prima-24B
A larger character roleplay model trained on the custom "Actors" dataset, the largest I've ever made! This model was made to expand on the things I learned from TinyRP, and to overcome certain limitations I found from it; also on an entirely new dataset made just for this model.
This model does have a bit of AI-style writing, but is overall more reliable in it's outputs than the smaller Trouper-12B; also tolerates mismatched templates more. That said, I do think that it is worth releasing this model, especially since it's larger size will likely help it's knowledgebase / longer conversations.
IF you want less purple prose, and more emotionally intimate characters, do check out Trouper-12B.
-> You can find it here! Trouper-12B
Looking for feedback, so please do share if you got any!
Key Features
- Reliable: Consistent behavior without meta-breaks or template issues
- Grammatically consistent: Zero perspective confusion errors
- Long context: Better handling of 50+ turn conversations
- Action-oriented: Natural energy for adventure/action roleplay scenarios
- Zero-fuss setup: More forgiving of template variations
Recommended Settings
Use chat completion mode
- Temperature: 0.7 (tested and validated)
- Template: Mistral-V7-Tekken or ChatML (critical for proper formatting and stop behavior, ChatML may perform better)
- Context: Handles longer turn conversations effectively
- Prompt Preprocessing: Semi-strict, no tools
Strengths
- Reliability: No meta-narration breaks, consistent stopping behavior
- Template Flexibility: Works with various Mistral templates
- Long Context: Maintains quality over extended conversations
- Adventure Energy: Better at action-oriented, dynamic scenarios
- Accessibility: Easier to set up and use than Trouper
Comparison to Trouper-12B
Prima-24B and Trouper-12B are trained on identical data but offer different trade-offs:
| Aspect | Prima-24B | Trouper-12B |
|---|---|---|
| Prose Style | Slightly more elaborate | Direct and concrete |
| AI Slop | Moderate (some patterns) | Minimal |
| Reliability | Excellent | Good (template-sensitive) |
| Long Context | Better (24B) | Good (12B) |
| Inference Speed | Slower (24B) | Faster (12B) |
| Setup Difficulty | Easy | Moderate (template critical) |
| Action RP | Excellent | Good |
| Emotional RP | Good | Excellent |
Choose Prima-24B if: You want reliability, long context, or action-oriented RP
Choose Trouper-12B if: You want best prose quality and don't mind template setup or the occaisional regeneration of a reply.
Known Characteristics
- Prose Style: Tends toward slightly more elaborate descriptions (some users may perceive as "AI-ish")
- Repetitive Descriptors: May occasionally reuse phrases like "blue eyes" + descriptor
- Purple Prose: Occasional tendency toward flowery language (not excessive)
- Structural Patterns: Generally good variety, but slightly more predictable than Trouper-12B
None of these are critical flaws - just characteristics to be aware of. Temperature adjustments (0.8-0.9) may help increase variety.
Got Feedback?
Issues, questions, or feedback welcome! Particularly interested in:
- Long conversation quality (20+ turns)
- Template compatibility findings
- Comparison with other RP models
Feel free to make a post in the Community tab here!
Why train on a base model?
According to this paper: Base Models Beat Aligned Models at Randomness and Creativity; and to avoid any possible "GPT-isms", I decided to train on a base model. Think of it as more mallable clay vs re-shaping something that was already formed to be something else.
This is what led to the behavior observed in this model, where the model just legitimately doesn't understand being an "assistant" outside of being a character that is an assistant. SO while the model is probably not useful outside of RP, it is also not intended to be.