Overview
Aratako/MistralPrism-24B is a 24 billion parameter language model developed by Aratako, built upon the mistralai/Mistral-Small-3.1-24B-Instruct-2503 base model. It has been further refined for role-playing applications by merging it with several other models, including AlexHung29629/Mistral-Small-3.1-24B-Instruct-2503-text, TheDrummer/Cydonia-24B-v3, LatitudeGames/Harbinger-24B, Gryphe/Pantheon-RP-1.8-24b-Small-3.1, aixonlab/Eurydice-24b-v3, and Aratako/Mistral-Small-3.1-24B-RP.
Key Capabilities
- Enhanced Role-Playing: Specifically designed and optimized for engaging in detailed character role-play, allowing users to define character settings and dialogue situations via system prompts.
- Context Length: Supports a context length of 32,768 tokens, making it suitable for longer, more complex role-playing narratives, though long context performance is noted as untested.
- Chat Template Adherence: Utilizes a specific chat template (
<s>[SYSTEM_PROMPT]system prompt[/SYSTEM_PROMPT][INST]user message 1[/INST]assistant message 1</s>[INST]user message 2[/INST]) for optimal performance, compatible with tokenizer.apply_chat_template.
Intended Use Cases
- Interactive Storytelling: Ideal for applications requiring dynamic and immersive character interactions.
- Role-Playing Games (RPGs): Can serve as a robust engine for generating character responses in text-based RPGs.
- Creative Writing Assistance: Useful for scenarios where consistent character voice and detailed situational responses are needed.
Limitations
- The model's vision capabilities have been removed, meaning it does not support image inputs.
- While trained with a 32k context length, its performance with long contexts is currently unverified.
- It is suggested to use the model with a low temperature (e.g., 0.15 or lower) for better output quality, similar to its base model, though this is also unverified for MistralPrism-24B itself.