PulseAugur
EN
LIVE 11:51:22

RVL-CDIP dataset revised to fix errors and test-train overlap

By PulseAugur Editorial · [2 sources] ·

A new paper revises the RVL-CDIP dataset, a popular benchmark for document classifiers, by identifying and correcting label errors and test-train overlap. The analysis revealed approximately 12% label errors and 35% test-train duplication within the original dataset. Modifications to the dataset showed that removing label errors improved classification accuracy, while removing duplicates led to a decrease. Furthermore, training on the corrected data significantly enhanced out-of-distribution generalization, with supervised models achieving an average accuracy increase of 8.1 percentage points on the RVL-CDIP-N benchmark. AI

IMPACT Improved dataset quality for document classification benchmarks may lead to more reliable model evaluations and better out-of-distribution generalization.

RANK_REASON The cluster contains an academic paper detailing dataset revisions and benchmark analysis.

Read on arXiv cs.CL →

AI-generated summary · Google Gemini · from 2 sources. How we write summaries →

RVL-CDIP dataset revised to fix errors and test-train overlap

COVERAGE [2]

  1. arXiv cs.CL TIER_1 English(EN) · Stefan Larson, Attila Nagy, Sam Desai, Cyrus Desai, Nicole C. Lima, Yixin Yuan, Siddharth Betala, Kaushal K. Prajapati, Jamiu T. Suleiman, Sharad Duwal, Kevin Leach ·

    Revising RVL-CDIP: Quantifying Errors and Test-Train Overlap

    arXiv:2606.31446v1 Announce Type: new Abstract: RVL-CDIP is a popular dataset for benchmarking document classifiers. However, the dataset contains ample amounts of label errors as well as non-trivial amounts of test-train overlap, both of which may impact model performance metric…

  2. arXiv cs.CL TIER_1 English(EN) · Kevin Leach ·

    Revising RVL-CDIP: Quantifying Errors and Test-Train Overlap

    RVL-CDIP is a popular dataset for benchmarking document classifiers. However, the dataset contains ample amounts of label errors as well as non-trivial amounts of test-train overlap, both of which may impact model performance metrics. In this paper, we address these two problems …

RELATED TOPICS