Fine-Tuning Large Language Models
Adapt pre-trained LLMs.
Full Fine-Tuning
Update all parameters. Needs GPU memory. LoRA: Low-Rank Adaptation.
Parameter Efficient
LoRA: add adapters. Prefix tuning: add prompt tokens. Adapter fusion.
Prompt Engineering
Few-shot examples. Chain-of-thought. ReAct: reasoning + acting.
Key Takeaways
- Full fine-tuning updates all params
- LoRA efficient adapter approach
- Prompt engineering before fine-tuning