llmfinetuningefficiency
LoRA and parameter-efficient finetuning
LoRA, QLoRA, prefix tuning, adapters — finetuning LLMs without touching all the weights.
Уровни глубины
L0Intro~1ч
Reads "low-rank matrices injected into attention" one-liner.
L1Basics~10ч
Uses peft library to LoRA-tune a 7B model on a laptop-grade GPU.
L2Working~15ч
Picks target modules, rank, α; combines with 4-bit QLoRA; merges adapters.
L3Advanced~25ч
DoRA, LoRA+ / rsLoRA; analysis of intrinsic rank of finetuning updates.
L4Research~50ч
Contributes new PEFT methods or theoretical rank analyses.