| --- |
| license: apache-2.0 |
| base_model: |
| - mistralai/Mistral-Small-24B-Base-2501 |
| language: |
| - en |
| pipeline_tag: text-generation |
| tags: |
| - character_roleplay |
| - creative_writing |
| - roleplay |
| --- |
| # 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](https://huggingface.co/DarwinAnim8or/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](https://arxiv.org/abs/2505.00047); 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. |
