ARGUARD SHARED TASK: Harmful Content Detection in Arabic Memes and LLM Prompts

ArGuard is a comprehensive shared task designed to advance the detection of harmful content in Arabic, focusing on two distinct but related domains: multimodal memes and LLM prompts. The task addresses the limitations of previous research, which often treated these settings in isolation or relied on coarse-grained labels. By providing a unified evaluation framework, ArGuard enables researchers to compare diverse modeling strategies, including encoder-based architectures, vision-language models (VLMs), and decoder-only LLMs.

The task was divided into two tracks. Track A evaluated multimodal hateful meme detection, requiring systems to jointly interpret image and text content. Subtask A1 focused on binary hate detection, while A2 required fine-grained classification across ten categories (e.g., mocking, dehumanization, sarcasm). Track B addressed LLM prompt safety, with B1 performing binary classification (safe vs. unsafe) and B2 assigning prompts to one of seven specific harm categories, such as self-harm or bullying. Participants utilized a wide array of techniques, including LoRA-based fine-tuning of VLMs, ensemble learning, class-aware loss functions, and threshold calibration to handle label imbalances and distribution shifts.

Results from the 35 participating teams highlight a consistent performance gap between binary and fine-grained classification. In the meme track, while the best system achieved a macro-F1 of 0.823 for binary detection, performance dropped to 0.419 for fine-grained classification. Similarly, in the prompt safety track, binary detection reached high macro-F1 scores (0.984), whereas fine-grained harm categorization proved more difficult (0.790). The analysis suggests that while OCR text often provides the primary signal for meme detection, multimodal fusion and VLM-based ensembles offer significant advantages. For prompt safety, both decoder-only LLMs and ensembles of Arabic-specific encoders proved highly effective, provided that systems were calibrated to account for differences between training and test data distributions.