coder3101/Mistral-Small-3.2-24B-Instruct-2506-heretic

Model Overview

This model, coder3101/Mistral-Small-3.2-24B-Instruct-2506-heretic, is a 24 billion parameter instruction-tuned language model based on mistralai/Mistral-Small-3.2-24B-Instruct-2506. Its primary differentiator is its decensored nature, achieved using the Heretic v1.1.0 tool. This modification drastically reduces refusals, with the model exhibiting only 5 refusals out of 100, compared to 97/100 in the original model.

Key Capabilities

• Decensored Responses: Significantly fewer refusals compared to the base model, allowing for broader utility.
• Improved Instruction Following: Enhanced ability to follow precise instructions, as indicated by higher scores on Wildbench v2 (65.33%) and Arena Hard v2 (43.1%).
• Robust Function Calling: Features a more robust function calling template, making it effective for tool use and complex task execution.
• Vision Reasoning: Capable of processing and reasoning with image inputs, demonstrated through examples like analyzing battle scenarios.
• Reduced Repetition Errors: Shows a 2x reduction in infinite generations on challenging prompts, improving output quality.
• Strong STEM Performance: Maintains strong performance in STEM benchmarks, with improvements in MMLU Pro (69.06%), MBPP Plus (78.33%), HumanEval Plus (92.90%), and SimpleQA (12.10%).

When to Use This Model

This model is particularly well-suited for applications requiring:

• Unfiltered Responses: Use cases where the base model's refusal rate is prohibitive.
• Complex Instruction Execution: Scenarios demanding high accuracy in following detailed instructions.
• Tool Integration: Applications leveraging function calling for external tool interaction.
• Multimodal Tasks: Projects involving both text and image inputs for reasoning.
• Code Generation and STEM Tasks: Where strong performance in programming and scientific domains is critical.