SaRoHead: Detecting Satire in a Multi-Domain Romanian News Headline Dataset

News headlines mainly aim to summarize an event in a news article in a few words. Media outlets may use them either to summarize objectively or to attract attention. For the latter, they may employ stylistic devices such as sarcasm, irony, and exaggeration, the key elements of satire. Consequently, even a headline may reflect that tone. Romanian approaches typically detect non-conventional tones (i.e., satire and clickbait) by combining the article and the headline, so headline-only analysis remains largely unexplored. To address this gap, we introduce SaRoHead, a dataset of 20,676 Romanian news headlines spanning social, politics, and sports domains, and analyze satirical tone in headlines alone. We benchmark standard machine learning algorithms, transformer-based classifiers, and large language models (LLMs), each with five random seeds. Bidirectional transformer models outperform standard machine learning approaches. Among BERT-based settings, the meta-learning Reptile approach performs best (pooled F1 90.42 for Romanian BERT). LLMs struggle under few-shot in-context learning, but parameter-efficient fine-tuning makes them competitive, with LoRA RoGemma and RoLlama3 achieving the highest pooled F1-scores (92.7 and 92.0). Because all satirical headlines originate from a single outlet, we treat the reported scores as an upper bound and restrict our claims to this benchmark setting.