Stratified Prediction-Powered Inference for Hybrid Language Model Evaluation

TL;DR

This paper introduces StratPPI, a stratified inference method that enhances prediction-powered inference for language model evaluation by providing tighter confidence intervals through data stratification, without assumptions on data distribution.

Contribution

The paper proposes StratPPI, a stratification-based method that improves confidence interval estimation in prediction-powered inference for language models, with theoretical guarantees and empirical validation.

Findings

StratPPI yields significantly tighter confidence intervals than unstratified methods.

The approach is robust without assumptions on data distribution or automatic labeling bias.

Empirical results demonstrate improved evaluation accuracy for language models.

Abstract

Prediction-powered inference (PPI) is a method that improves statistical estimates based on limited human-labeled data. PPI achieves this by combining small amounts of human-labeled data with larger amounts of data labeled by a reasonably accurate -- but potentially biased -- automatic system, in a way that results in tighter confidence intervals for certain parameters of interest (e.g., the mean performance of a language model). In this paper, we propose a method called Stratified Prediction-Powered Inference (StratPPI), in which we show that the basic PPI estimates can be considerably improved by employing simple data stratification strategies. Without making any assumptions on the underlying automatic labeling system or data distribution, we derive an algorithm for computing provably valid confidence intervals for population parameters (such as averages) that is based on stratified sampling. In particular, we show both theoretically and empirically that, with appropriate choices of stratification and sample allocation, our approach can provide substantially tighter confidence intervals than unstratified approaches. Specifically, StratPPI is expected to improve in cases where the performance of the autorater varies across different conditional distributions of the target data.