Directional Robustness Asymmetry in Indonesian Sarcasm Detection: A Cross-Platform Evaluation of IndoRoBERTa on X and Reddit

Sarcasm can reverse sentence polarity, making it one of the hardest problems in natural language processing and a persistent obstacle to reliable sentiment analysis on social media, where models are often deployed on platforms they were not trained on. Research on Indonesian sarcasm has largely stayed within a single domain or tested transfer in only one direction, leaving the cross-domain robustness of the IndoRoBERTa family on the IdSarcasm benchmark unclear. This study measures and compares the cross-domain robustness of IndoRoBERTa-small and IndoRoBERTa-base, with IndoBERT as a baseline, across X (Twitter) and Reddit, using the F1 gap (ΔF1) after cleaning anonymization artifacts, adding a data-size control, and running Stratified 5-Fold cross-validation. The results reveal a one-sided asymmetry: X → Reddit transfer is far more fragile (ΔF1 0.345–0.400) than Reddit → X (0.135–0.240), failing by missing sarcasm (false negatives) rather than over-flagging it (false positives). Practically, training on the context-rich domain (Reddit) and choosing IndoRoBERTa-base give the most reliable cross-platform sarcasm detection, with direction-specific mitigation recommended for robust deployment.