Dialectical Variations

Uses dialects, regional variations, slang, pidgins, and creoles to express harmful requests in language variants underrepresented in safety training. The model understands most English dialects; the classifier may not. This is intra-language what multilingual evasion (T2-AT-002) is inter-language.

• Dialect-aware classifiers
• Slang-to-standard-English mapping

Chains with T2-AT-002 (Multi-Language) as intra-language equivalent.