RewardDS is a novel framework that enables fine-tuning LLMs with synthetic data while preserving privacy in sensitive domains like healthcare and finance.

• Uses a reward-driven approach to synthesize high-quality training data with Differential Privacy guarantees
• Identifies and removes flawed synthetic data through an innovative filtering mechanism
• Achieves superior performance compared to existing privacy-preserving methods
• Maintains model utility while providing strong privacy protection

This research is particularly valuable for the medical sector, where organizations can fine-tune domain-specific LLMs without exposing patient data, enabling advanced AI applications while maintaining regulatory compliance and patient trust.

RewardDS: Privacy-Preserving Fine-Tuning for Large Language Models via Reward Driven Data Synthesis