Humor recognition is a major challenge in Natural Language Processing (NLP) due to its subtle and context-dependent nature. Despite advances, Large Language Models (LLMs) still struggle with this task, especially in Brazilian Portuguese, where no dedicated benchmarks exist. This paper presents HuNeBR, a new benchmark of 475 annotated humorous texts from Northeastern Brazilian comedians. The benchmark evaluates LLMs on three tasks: identifying punchlines, classifying texts into eight comic styles, and explaining humor. This is the first benchmark to evaluate LLMs on the in-depth interpretation of humorous texts in Brazilian Portuguese, going beyond the binary tasks of traditional humor benchmarks. Both general-purpose and Portuguese-specialized LLMs were evaluated under zero-shot and few-shot settings. The findings indicate that LLMs perform very well at identifying punchlines, show inconsistent results in classifying comic styles, and produce humor interpretations that mostly align with human judgments. Among the models assessed, general-purpose multilingual systems like GPT-4 and Gemini 2.5 Flash achieved the top overall performance, whereas Sabiá 3.1, a model specialized in Brazilian Portuguese, demonstrated competitive results across all three tasks, highlighting the value of locally trained models in capturing linguistic and cultural subtleties.

Large Language Models (LLMs) are typically evaluated using standardized benchmarks to enable consistent performance measurement and model comparison. However, the reliability of these benchmarks can be undermined by data contamination, which occurs when evaluation items are inadvertently included in training corpora. While this issue has been investigated primarily in high-resource languages such as English and Chinese, its impact on underrepresented languages — such as Brazilian Portuguese — remains understudied. In this paper, we present one of the first systematic investigations of benchmark data contamination (BDC) in an underrepresented language setting, using Brazilian Portuguese as a case study. Using validated methodologies from the literature, we evaluate specialized and multilingual models across four benchmarks: BLUEX, ENEM Challenge, OAB Exams, and HealthQA-BR. Our approach applyes TS-Guessing to detect contamination via memorized knowledge, alongside a 50-character n-gram similarity strategy to identify benchmark items leaked into training data. Our results provide consistent evidence of contamination, revealing that models with stronger memorization and retrieval abilities tend to achieve artificially inflated benchmark scores. Our contributions include: (i) classifying models according to their contamination risk, (ii) identifying the benchmarks most affected by data leakage, and (iii) reporting contaminated training corpora.