Mistral-Small-Sisyphus-24b-2503
| Property | Value |
|---|---|
| Parameter Count | 24 billion |
| Base Model | Mistral |
| Training Template | v7-Tekken |
| Model URL | HuggingFace |
What is Mistral-Small-Sisyphus-24b-2503?
Mistral-Small-Sisyphus-24b-2503 is a fine-tuned version of the 24B parameter Mistral base model, developed by allura-org. Initially intended for multi-turn instruction following, the model unexpectedly became oriented towards roleplay due to a configuration quirk during training.
Implementation Details
The model implements a Claude-like system with support for explicit reasoning blocks. It performs well across various temperature settings when used with min-p or top-p sampling methods. The implementation includes support for specialized thought processing through <think> tags.
- Supports both reasoning and non-reasoning system prompts
- Uses v7-Tekken instruct template
- Implements Claude-style interaction patterns
- Features explicit thought processing capabilities
Core Capabilities
- Multi-turn conversation handling
- Structured reasoning using thought blocks
- Temperature-flexible response generation
- Roleplay capabilities (unintended feature)
Frequently Asked Questions
Q: What makes this model unique?
The model's unique feature is its ability to handle both standard instruction-following and explicit reasoning through thought blocks, while maintaining coherence across different temperature settings. Its unintended roleplay capabilities add an interesting dimension to its functionality.
Q: What are the recommended use cases?
The model is well-suited for applications requiring structured reasoning, multi-turn conversations, and potentially roleplay scenarios. It works best when used with temperature control methods like min-p or top-p sampling.
