200,000 synthetically generated conversations for training safety classifiers, covering harmful (100k), borderline (40k), and safe (60k) content across 9 safety-critical domains.
Generation Methodology
Subtopic Generation: Created 797 specific subtopics across 9 safety categories using systematic domain analysis.
Configuration Sampling: Generated 7.3M+ unique prompt configurations by combining:
