aiml-sarcasm-harassment

Benchmark veiled harassment detection by running the provided Python scripts against JSON datasets to evaluate a local transformer-based classifier. This workflow validates data integrity and uses explicit context checks across scenarios like workplace bullying, online dogpiling, and microaggression.

Veiled harassment detection covers workplace bullying, online dogpiling, gaslighting, microaggression, and passive threats. It uses anchor-based evaluation constraints to ensure realism and deniability in generated harassment samples for AI safety model behavior research.

Yes, you can use a local transformer-based classifier with explicit context checks to evaluate gaslighting and passive threats. The benchmarking workflow enforces no placeholders and validates data integrity to ensure accurate AI safety model behavior assessment.

Evaluate sarcasm detectors on veiled harassment by executing the provided benchmark.py against sarcasm_cases.json. The harness validates data integrity, enforces no placeholders, and uses anchor-based constraints to ensure realistic deniability in AI safety datasets.

Yes, the benchmark harness enforces no placeholders and validates data integrity in generated harassment samples. It applies anchor-based evaluation constraints to maintain realism and deniability across workplace bullying, gaslighting, and microaggression contexts.

Extend the AI safety benchmark to new harassment contexts by modifying the included JSON datasets and Python scripts. The ready-made evaluation harness supports adding new scenarios while maintaining anchor-based constraints, data integrity, and explicit context checks for model behavior.