Skip to content

Generating LLM Explanations

EvalxNLP provides powerful tools to interpret explanation scores and evaluation metrics using natural language generated by Large Language Models (LLMs). These human-readable interpretations enhance the transparency and usability of XAI methods.

The framework supports integration with Together AI's API, using the meta-llama/Llama-3.3-70B-Instruct-Turbo-Free model by default.

✅ Prerequisites

  • Create an API key from Together AI.
  • Pass your API key and model name to the LLMExplanationGenerator and/or EvaluationMetricsExplainer class(es)

📘 Understanding Explanation Scores

Use LLMExplanationGenerator class to interpret importance scores generated from different explainers. generate_and_save_explanations() allows to obtain explanations and save descriptions in JSON, HTML, or both formats. Explanations are saved in the results/explanations/ folder.

You can also visualize the output within notebook cell using display_explanations() function.

from LLMExplanationGenerator import LLMExplanationGenerator
import pandas as pd
from IPython.display import display, HTML

model_name = "meta-llama/Llama-3.3-70B-Instruct-Turbo-Free"
api_key = "YOUR_API_KEY_HERE"

explainer = LLMExplanationGenerator(model_name=model_name, api_key=api_key)

explanations, saved_files = explainer.generate_and_save_explanations(
    exps=exps,
    output_format="both"  # Choose from "json", "html", or "both"
)

# 📺 Display in notebook
explainer.display_explanations(explanations)

📊 Understanding Evaluation Metrics

Use EvaluationMetricsExplainer class to interpret evaluation metrics. explain_results() generates explanations and save_results() saves the output in JSON, HTML, or both formats to results/explanations/ folder. You can also visualize the output within notebook cell using display_results() function.

from EvaluationMetricsExplainer import EvaluationMetricsExplainer

api_key = "YOUR_API_KEY_HERE"
explainer = EvaluationMetricsExplainer(api_key=api_key)

# Provide results from evaluation functions
results = explainer.explain_results(metrics)

# Save results to files
json_path, html_path = explainer.save_results(results)

# 🖥️ Display in notebook
explainer.display_results(results)

📍 The generated reports include metric definitions, score interpretations, and comparisons across explainers in structured HTML or JSON format.