🎛️

Fine-Tuning LLMs

Full parameter vs PEFT, LoRA, QLoRA, quantization, domain fine-tuning

Fine-Tuning vs Pre-Training vs RAG

Conceptual hierarchy: what each does to model weights

Updating billions of params: memory, compute, GPU cost reality

FP32 → INT4/INT8, size reduction, speed vs accuracy tradeoff

Freezing base weights + training small rank matrices, math intuition

Combining quantization + LoRA for efficient fine-tuning

LoRA, Prefix Tuning, Prompt Tuning, Adapter layers

RLHF, instruction-following, safety layers

Labeled domain data, specialty abbreviations, niche tone