Model Overview

allura-org/Mistral-Small-Sisyphus-24b-2503 is a 24 billion parameter language model, developed by allura-org, featuring a 32,768 token context window. This model is a fine-tuned variant of the 24B base model, primarily optimized for multi-turn roleplay (RP) and general instruct tasks. Despite an initial training dataset mix-up, the model exhibits strong coherence across different temperature settings, particularly when paired with min-p or top-p sampling.

Key Capabilities

• Multi-turn Roleplay: Designed for engaging in extended, coherent roleplaying scenarios.
• Reasoning Blocks: Supports the integration of reasoning blocks, allowing for structured thought processes within responses, typically guided by specific system prompts.
• Coherent Output: Maintains consistency and logical flow in generated text across a range of sampling parameters.
• Instruct Template: Utilizes the v7-Tekken instruct template, aligning with the original instruct model's format.

Recommended Usage

This model is particularly well-suited for applications requiring interactive and sustained conversational capabilities. For reasoning tasks, it is recommended to prefill <think>\n to encourage structured thought processes and to avoid including previous thought blocks in new requests. While initially intended for a broader dataset, its performance leans towards effective roleplaying and general conversational instruction.