Browse Teach Curate

Multi-Dimensional Attack-Defense Data Construction for LLM Safety Alignment

A data-centric technique for improving the safety of large language models. Safety-alignment training data is enriched by expanding the diversity of adversarial attack instructions across many intent categories and by regenerating higher-quality safe responses, then filtering both with a perplexity- and LLM-score-based safety reward model before supervised fine-tuning.

Data Text

A Method for Enhancing the Safety of Large Model Generation Based on Multi-dimensional Attack and

Proposes a data-centric method for improving the safety of large language models by building a safety-alignment supervised fine-tuning (SFT) dataset. Adversarial 'attack' instructions drawn from Chin…