LLM-as-a-judge for sarcasm detection using supervised fine-tuning of transformers

This research conducts a systematic comparative study of large pre-trained language models (LLMs) for sarcasm and irony detection. While pretrained transformers often struggle to capture sarcastic intent, we fine-tune multiple domain-specific models and assess their adaptability across diverse review contexts. Using two complementary open-source datasets, news headlines and product reviews, we evaluate four transformer-based models (RoBERTa-large, RoBERTa-base, DistilBERT-base-uncased, and DistilBERT-SST2) under consistent experimental conditions. The framework emphasizes algorithmic transparency, reproducibility and structured evaluation through group-aware data splits, label smoothing and macro-F1-based early stopping. Results indicate that DistilBERT-SST2 achieves the strongest and most stable performance (macro F1 = 0.8784) and demonstrates resilience across five distinct review domains (movies, general, product/service/tech). The research provides an interpretable and reproducible baseline for evaluating fine-tuned LLMs in sarcasm detection and identifies key patterns of domain sensitivity and transfer.