TRITONBENCH-T: NSFW Moderation Benchmarks

• TRITONBENCH-T is a benchmark suite featuring rigorous evaluation protocols and diverse, expert-annotated datasets for NSFW content detection in multimodal environments.
• It aggregates large-scale datasets such as CensorChat and curated collections like P2, ACI, and REDDIX-NET for detailed class-level assessment.
• The suite supports robust, reproducible evaluation of automated NSFW moderation systems in challenging real-world contexts including generative AI and online platforms.

TRITONBENCH-T refers to a prominent and comprehensive suite of benchmarks and datasets for evaluating and comparing methods for the detection, filtering, and moderation of Not Safe For Work (NSFW) content in multimodal (text, image, video) settings. TRITONBENCH-T comprises discrete evaluation sets, protocols, and model evaluation standards intended for rigorous, reproducible, and nuanced assessment of automated NSFW moderation systems, particularly in complex, real-world contexts such as generative AI, online platforms, and adversarial scenarios.

1. Benchmark Datasets and Taxonomies

A key aspect of TRITONBENCH-T is its aggregation of large-scale, high-complexity datasets across textual, visual, and multimodal domains, representing a spectrum of NSFW subtypes, adjudicated by expert and community-based human annotation to facilitate granular class-level evaluation. Notable constituent datasets include:

• CensorChat: Human-machine dialogues parsed into single utterances and multi-turn segments, annotated for NSFW content and rationales using GPT-4 and ChatGPT distillation. Validation/test splits rely on self-criticism for label consistency, with strong protocols for handling ambiguous or context-dependent cases (Qiu et al., 2023).
• P2, ACI, and REDDIX-NET: Curated collections encompassing pornographic, porn-indicative, and clean images, and textual