LoRA for Efficient LLM Fine-Tuning

Implemented LoRA for GPT-2 by injecting trainable low-rank adapters into attention layers

  • Implemented LoRA for GPT-2 by injecting trainable low-rank adapters into attention layers while freezing base model weights.
  • Achieved ~2% higher accuracy than full fine-tuning while updating <5% of parameters, benchmarking convergence and performance across multiple LoRA ranks.