gratex/Mistral-Small-3.2-24B-Instruct-2506-sbvr-compliance-unresolvable-pattern-finetune-v1

VISIONPricing:Input $0.867 / Output $1.61Concurrent Unit Cost:2Model Size:24BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Aug 6, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

The gratex/Mistral-Small-3.2-24B-Instruct-2506-sbvr-compliance-unresolvable-pattern-finetune-v1 is a 24 billion parameter Mistral-Small-3.2-24B-Instruct-2506 model fine-tuned by gratex. This model specializes in evaluating SBVR rules, enhancing the accuracy of rule status assignment (PASSED, FAILED, UNRESOLVABLE) with a focus on correctly identifying 'UNRESOLVABLE' cases. It addresses a limitation of its base model by preventing unsupported assumptions when context is insufficient, improving reliability in business rule compliance checks. The model has a context length of 32768 tokens and shows a 15.5% improvement on unresolvable patterns in HELM evaluations.

Loading preview...

Model Overview

This model, gratex/Mistral-Small-3.2-24B-Instruct-2506-sbvr-compliance-unresolvable-pattern-finetune-v1, is a fine-tuned version of the mistralai/Mistral-Small-3.2-24B-Instruct-2506 base model. Its core purpose is to enhance the evaluation of Semantics of Business Vocabulary and Business Rules (SBVR) by accurately assigning one of three verdicts: PASSED, FAILED, or UNRESOLVABLE.

Key Capabilities

  • Improved SBVR Rule Evaluation: Specifically trained to distinguish between rule satisfaction, violation, or insufficient context.
  • Enhanced 'UNRESOLVABLE' Detection: Addresses a limitation of the base model by correctly identifying situations where a definitive conclusion cannot be reached due to lack of evidence, preventing unsupported assumptions.
  • Specialized Fine-tuning: The model underwent LoRA fine-tuning over 180 iterations, targeting attention layers (q/k/v/o) with a learning rate of 5e-06.

Performance Highlights

HELM evaluation results demonstrate significant improvements over the base model, particularly in its specialized domain:

  • +15.5% improvement on unresolvable_patterns suite.
  • Modest gains across several other benchmarks, including quac (+4.1%), mmlu_pro:history (+3.1%), and truthful_qa (+2.5%).

When to Use This Model

  • Business Rule Compliance: Ideal for applications requiring precise evaluation of business rules against given contexts.
  • Mitigating Ambiguity: Use when it's critical to identify and flag cases where information is insufficient to make a definitive judgment, rather than forcing an incorrect PASSED or FAILED verdict.
  • Domain-Specific Reasoning: Suited for tasks where understanding and correctly classifying 'unresolvable' conditions is paramount.

Finetunes
1 models