0xA50C1A1/Mistral-Small-3.2-24B-Instruct-2506-SOM-MPOA
0xA50C1A1/Mistral-Small-3.2-24B-Instruct-2506-SOM-MPOA is a 24 billion parameter instruction-tuned language model, a decensored version of Mistral-Small-3.2-24B-Instruct-2506, offering a 32K context window. This model is specifically modified to reduce refusals and improve instruction following, repetition error reduction, and function calling capabilities compared to its base model. It excels in multimodal tasks, including vision reasoning, and is suitable for applications requiring robust instruction adherence and tool use.
Loading preview...
Model Overview
This model, 0xA50C1A1/Mistral-Small-3.2-24B-Instruct-2506-SOM-MPOA, is a 24 billion parameter instruction-tuned language model. It is a decensored variant of the original Mistral-Small-3.2-24B-Instruct-2506, created using the Heretic v1.2.0 tool. The primary differentiation of this version lies in its significantly reduced refusal rate, dropping from 79/100 to 2/100 compared to the original model, while maintaining a KL divergence of 0.0413.
Key Capabilities
- Decensored Output: Engineered to provide responses with fewer refusals, offering a less restricted conversational experience.
- Enhanced Instruction Following: Demonstrates improved ability to follow precise instructions, leading to more accurate and desired outputs.
- Reduced Repetition Errors: Generates less repetitive or infinite responses, particularly in challenging and long prompts, with a 2x reduction in infinite generations.
- Robust Function Calling: Features a more robust function calling template, excelling in tool-use tasks.
- Multimodal Reasoning: Supports vision capabilities, allowing for reasoning based on image inputs, as demonstrated in examples like Pokémon battle strategy and image-based calculations.
- Strong Performance: Maintains or slightly improves upon the Mistral-Small-3.2-24B-Instruct-2506's benchmarks in areas like instruction following (Wildbench v2: 65.33%, Arena Hard v2: 43.1%), STEM tasks (MMLU Pro: 69.06%, MBPP Plus - Pass@5: 78.33%, HumanEval Plus - Pass@5: 92.90%), and certain vision tasks (ChartQA: 87.4%, DocVQA: 94.86%).
Good for
- Applications requiring a less restrictive and more direct response generation.
- Scenarios demanding high accuracy in following complex instructions.
- Use cases involving function calling and tool integration, especially with vLLM.
- Multimodal applications that benefit from vision reasoning capabilities.
- Tasks where reducing repetitive or circular responses is critical.