This is a comprehensive instruction-tuning dataset designed to train language models to avoid generating specific forbidden words while maintaining natural language capabilities. The dataset combines multiple sources of high-quality training data including AI-generated adversarial examples and rule-based prompts.
