FedDQC introduces dynamic quality control for federated instruction-tuning of LLMs, enabling organizations to collaboratively train models while keeping sensitive data local.

• Addresses the challenge of filtering low-quality samples in decentralized training environments
• Implements two-phase quality control mechanisms that work within privacy constraints
• Enables privacy-preserving collaboration between organizations with sensitive instruction data
• Demonstrates improved model performance compared to standard federated learning approaches

This research is crucial for security-conscious organizations seeking to leverage collective data resources without exposing proprietary information or user data, while still maintaining high training standards.

Data Quality Control in Federated Instruction-tuning of Large Language Models