PEFT Fine-Tuning

LoRA/QLoRA fine-tuning code pro HuggingFace models s rank/alpha guidance a OOM fixes.

od Orchestra-Research · Orchestra-Research/AI-Research-SKILLs

Otestováno · Funguje ★ 8.4/10

PEFT Fine-Tuning — LoRA/QLoRA fine-tuning code pro HuggingFace models s rank/alpha guidance a OOM fixes.

Co umí

Generuje HuggingFace PEFT code pro LoRA a QLoRA fine-tuning of 7B-70B language models, including rank/alpha selection tables, architecture-specific target-module lists, a CUDA OOM troubleshooting. Spouští se na požadavky na fine-tune large models under GPU memory constraints, train adapters instead of full models, or serve multiple task-specific adapters from one base model.

Testovací report

The skill's QLoRA example lists all seven Llama-3.1 linear layers (q/k/v/o_proj plus gate/up/down_proj) as LoRA targets a pairs them s a rank/alpha table a a GPU-memory benchmark table (60GB full FT vs 18GB LoRA vs 6GB QLoRA); a memory-only baseline answer pro the same request produced a plausible but thinner config that only targeted q_proj/v_proj a cited no numbers.

Testováno: 2026-07-14 · Claude Code 2.x (agent harness)

Instalace

git clone https://github.com/Orchestra-Research/AI-Research-SKILLs
cd AI-Research-SKILLs
mkdir -p ~/.claude/skills
cp -r 03-fine-tuning/peft ~/.claude/skills/peft-fine-tuning

Příkazy a ukázkové prompty

  • /peft-fine-tuningLoRA/QLoRA fine-tuning code pro HuggingFace models s rank/alpha guidance a OOM fixes.

Skilly se spouštějí běžnými požadavky — žádné příkazy k zapamatování. Po instalaci ho aktivují prompty jako tyto (anglicky):

  • Fine-tune Llama 3.1 with LoRA using PEFT
  • Set up QLoRA target modules for my model
  • Configure PEFT LoRA for parameter-efficient fine-tuning