Test Set Quality in Multilingual LLM Evaluation

TL;DR

This paper critically examines the quality of multilingual benchmark datasets for LLM evaluation, revealing significant errors and performance discrepancies, and advocates for ongoing dataset revision and quality assurance.

Contribution

It provides a manual error analysis of multilingual test sets in French and Telugu, highlighting the importance of dataset quality and suggesting revisions for more reliable evaluation.

Findings

Large performance differences when using original vs. revised datasets

Identification of numerous errors in semi-automatically created datasets

Recommendations for dataset creators and users to improve test set quality

Abstract

Several multilingual benchmark datasets have been developed in a semi-automatic manner in the recent past to measure progress and understand the state-of-the-art in the multilingual capabilities of Large Language Models. However, there is not a lot of attention paid to the quality of the datasets themselves, despite the existence of previous work in identifying errors in even fully human-annotated test sets. In this paper, we manually analyze recent multilingual evaluation sets in two languages - French and Telugu, identifying several errors in the process. We compare the performance difference across several LLMs with the original and revised versions of the datasets and identify large differences (almost 10% in some cases) in both languages). Based on these results, we argue that test sets should not be considered immutable and should be revisited, checked for correctness, and potentially versioned. We end with some recommendations for both the dataset creators as well as consumers on addressing the dataset quality issues.