Dialect-aware LLM safety classifiers from the DIA-GUARD project — fine-tuned on 48 English dialects for safe/unsafe content detection.