Scoring outputs for a small-scale, reviewer-defensible pilot that probes where multilingual LMs systematically fail on minimal pairs, and whether those failures reflect grammatical competence or a surface confound (sentence length / subword tokenisation). Harness: github.com/suchirsalhan/xblimps.

What was scored

Published BLiMP-family benchmarks, one common scoring/confound/sampling pipeline, 10 pairs per paradigm (the… See the full description on the dataset page: https://huggingface.co/datasets/xBLiMPs/xblimps-pilot-results.