Building Responsible, Emotion-Driven Synthetic Dialogue Datasets

We present a persona-grounded framework for generating task-oriented dialogues that prioritizes safety, diversity, and quality. We formalize a structured persona schema and employ stratified sampling to ensure balanced coverage across demographics, regions, goals, and emotional states. Our scalable pipeline leverages GPT-4 for generation, implements FAISS-backed semantic deduplication, and enforces multi-stage safety validation. This approach produced 50,470 dialogues with 12.4\% redundancy removed and 99.7\% safety compliance. We evaluate performance across persona dimensions, measuring coverage, consistency, quality, and task success rates, and benchmark our results against MultiWOZ 2.2. Our findings demonstrate that persona conditioning combined with balanced sampling yields dialogues that are more coherent, controllable, and safer than standard baseline approaches.