Model reports

SciClaimEval

原始来源英文简介

1,664 claims from 180 papers in machine learning, NLP and medicine, verified against tables and figures. Refuted cases keep the authentic claim and edit the evidence instead, so a refuted item can differ from a supported one only in a small change to a table or figure. The authors report that figure-based verification remains particularly challenging for all models, so a blended score depends on the table/figure mix, and table and figure numbers are not interchangeable.