Fine-Tuning LLMs

Back to Large Language Models

Adapting a pretrained LLM to specific tasks or domains by further training on task-specific data. Ranges from full fine-tuning to parameter-efficient methods.

Methods

  • Full Fine-Tuning — update all weights, expensive but most capable
  • LoRA (Low-Rank Adaptation) — add small trainable matrices, freeze original weights
  • QLoRA — LoRA on quantized models, enables fine-tuning on consumer GPUs
  • Adapter Methods — insert small trainable modules between frozen layers

nlp llm fine-tuning lora