multiclass llm eval

[mike0sv]

PR Description: Implement Multiclass Classification Descriptors

This PR introduces a new module for multiclass classification prompt templates and evaluations in the evidently library. It includes two evaluation classes: RelevanceLLMEval and BrandSafetyLLMEval, which leverage the newly created MulticlassClassificationPromptTemplate to classify inputs based on specified criteria.

Key Features:

• Multiclass Classification Prompt Template: A structured way to create templates for classification tasks, allowing configuration of category criteria, uncertainty handling, and output formatting.
• Evaluation Classes:
  • RelevanceLLMEval evaluates how well answers address questions.
  • BrandSafetyLLMEval assesses text compliance with brand safety guidelines.

Examples:

1. Relevance Evaluation:

   relevance_eval = RelevanceLLMEval()
   template = relevance_eval.get_template()
   print(template.get_blocks())

   This evaluates how well an answer responds to a given question based on categories like "Irrelevant," "Partially Relevant," and "Relevant."

2. Brand Safety Evaluation:

   brand_safety_eval = BrandSafetyLLMEval()
   template = brand_safety_eval.get_template()
   print(template.get_blocks())

   This assesses text compliance with brand safety and tone using categories like "Fully Compliant," "Partially Compliant," and "Incompliant."

This addition enhances the flexibility and capability of the evidently library in handling complex evaluation scenarios for machine learning models.