DynaMiCS: Fine-Tuning LLMs with Performance Constraints Using Dynamic Mixtures

DynaMiCS optimizes LLM fine-tuning for multi-domain performance with explicit constraints. This new method addresses the challenge of improving performance on target tasks without degrading essential general capabilities. It uses dynamic mixture optimization and domain-specific probing to treat fine-tuning as a constrained optimization problem, offering a more robust approach to model adaptation.

Opening Kapyn…