aiml-nsfw-benchmark

You can benchmark NSFW text classification using a standardized ISC template to evaluate recall and category coverage across sexual_explicit, sexual_suggestive, harassment_sexual, grooming, and sextortion categories. The benchmark provides reproducible evaluation workflows with JSON samples and Python tooling.

The NSFW benchmark evaluates five categories: sexual_explicit, sexual_suggestive, harassment_sexual, grooming, and sextortion. These standardized categories enable researchers to measure recall, coverage, and robustness in controlled AI safety evaluation settings.

Yes, the NSFW benchmark supports multi-language prompt sets and language variants to test cross-language NSFW detection capabilities. This allows evaluation of model performance across different languages for comprehensive AI safety validation.

You can evaluate DistilBERT NSFW classifier performance using the benchmark's deterministic evaluation flows with ready-to-use prompts and sample datasets. The included Python tooling and JSON samples ensure reproducible evaluation workflows for measuring recall and category coverage.

Yes, the NSFW benchmark supports weak-anchor variants for prompt-variant experiments. This feature enables researchers to validate prompt-variant experiments and extend categories and anchors as needed for comprehensive model evaluation.

The best way to validate recall and category coverage is using a standardized NSFW benchmark with reproducible evaluation workflows. The framework provides deterministic evaluation flows across five explicit content categories with JSON samples and Python tooling for measuring robustness.