Fine-tuning & Adaptation
Techniques for adapting a pretrained model to specific tasks — from full supervised fine-tuning to parameter-efficient methods that train less than 1% of weights.
In This Section
Supervised Fine-Tuning (SFT)
Instruction tuning, chat templates, completion masking, and catastrophic forgetting.
LoRA & QLoRA — Parameter-Efficient Fine-Tuning
Rank decomposition, adapter merging, and fine-tuning 70B models on a single GPU.
Prompt Tuning & Adapter Methods
Soft prompts, prefix tuning, adapter layers, and when to use each PEFT method.