Skip to main content
In DevelopmentResearch

AdapterLLRD

Memory-free multilingual continual learning with PEFT + LLRD.

🎯The Problem

Catastrophic forgetting when fine-tuning sequentially by language/domain. Privacy constraints prevent storing prior data.

💡The Solution

Train only PEFT modules (LoRA/adapters) at each hop. Apply LLRD across the base model to stabilize low layers while allowing adaptation.

Key Highlights

  • 50+ hop sequences without replay
  • Privacy-compatible (no buffers)
  • BWT/FWT/AULC metrics