Abstract

For two decades, the HAREM corpus has served as the foundational benchmark for Portuguese Named Entity Recognition (NER), establishing its evaluation paradigm. Virtually all major progress has been measured against its fixed train/test split. This paper presents the first systematic audit of this split, revealing 153 overlapping (contaminated) sentences. We re-evaluate 13 NER models (ranging from CRFs to Transformers) on both the original and a new, decontaminated version of the corpus. Our statistical analysis reveals that decontamination has a significant (p < 0.05) and positive impact on the majority of models. We find that performance gains are most pronounced in the F1_textmacro score (up to +4 points), demonstrating that the contamination primarily harmed generalization on rare entity types. Furthermore, our audit reveals clear evidence of overfitting in some models that benefited from data leakage. We conclude that even minor contamination can distort performance metrics and mask true model generalization. We release our decontaminated benchmark to ensure more reliable future evaluations.