Judgy: Correcting LLM Judge Bias for Reliable AI Model Evaluation

Introduction

judgy is a Python library developed to address the challenges of using Large Language Models (LLMs) as judges for evaluating other AI models or systems. When LLMs are used in this capacity, their inherent biases and potential errors can significantly impact the reliability and accuracy of the evaluation results. This package provides a robust solution to estimate the true success rate of your system by correcting for these LLM judge biases.

Installation

Getting started with judgy is straightforward. You can install the package using pip:

pip install judgy

Examples

Here's a quick example demonstrating how to use judgy to estimate a system's true success rate with confidence intervals:

import numpy as np
from judgy import estimate_success_rate

# Your data: 1 = Pass, 0 = Fail
test_labels = [1, 1, 0, 0, 1, 0, 1, 0]      # Human labels on test set
test_preds = [1, 0, 0, 1, 1, 0, 1, 0]       # LLM judge predictions on test set  
unlabeled_preds = [1, 1, 0, 1, 0, 1, 0, 1]  # LLM judge predictions on unlabeled data

# Estimate true pass rate with 95% confidence interval
theta_hat, lower_bound, upper_bound = estimate_success_rate(
    test_labels=test_labels,
    test_preds=test_preds, 
    unlabeled_preds=unlabeled_preds
)

print(f"Estimated true pass rate: {theta_hat:.3f}")
print(f"95% Confidence interval: [{lower_bound:.3f}, {upper_bound:.3f}]")

Why Use It

The core problem judgy solves is the unreliability introduced by LLM judge biases. By implementing a bias correction method, judgy helps you obtain a more accurate estimate of your system's true performance. It works by first estimating the LLM judge's accuracy (True Positive Rate and True Negative Rate) using a labeled test set. Then, it applies a correction formula to the observed pass rate from the judge. Finally, it uses bootstrap resampling to quantify the uncertainty and provide a confidence interval, giving you a statistically sound range for your system's true success rate. This ensures that your evaluations are not only corrected for bias but also come with a measure of their statistical robustness.

Links

For more detailed information, documentation, and to contribute, visit the official judgy resources:

• GitHub Repository
• PyPI Package