fine-tune

name	fine-tune
description	Fine-tune language models with LoRA, QLoRA, or full fine-tuning. Covers unsloth (4x memory reduction), PEFT, trl SFTTrainer, DPO, instruction tuning with chat templates, dataset preparation, and evaluation. Use when fine-tuning any HuggingFace model on custom data.
allowed-tools	Bash(uv run * scripts/prepare_dataset.py *) Bash, Read, Write, Edit, Glob, Grep
argument-hint	base model name or path, dataset path (e.g. "mistralai/Mistral-7B-v0.1 data/train.jsonl")
model	opus
effort	high
disable-model-invocation	true
paths	**/*.jsonl,**/*.json
compatibility	>=1.0
metadata	{"category":"model-training","tags":["fine-tuning","lora","qlora","unsloth","peft","dpo","llm","huggingface","instruction-tuning"],"phase":"train"}

Fine-Tune Skill

Fine-tune language models efficiently with LoRA, QLoRA, or unsloth.

Quick start

# Prepare dataset
uv run ${CLAUDE_SKILL_DIR}/scripts/prepare_dataset.py data/raw.csv --format alpaca --output data/train.jsonl

# Fine-tune with LoRA (unsloth)
python3 -c "
from unsloth import FastLanguageModel
model, tokenizer = FastLanguageModel.from_pretrained('mistralai/Mistral-7B-v0.1', load_in_4bit=True)
model = FastLanguageModel.get_peft_model(model, r=16, lora_alpha=32, target_modules=['q_proj','v_proj'])
# ... training loop
"

Workflow

1. Prepare dataset — convert to alpaca/sharegpt format via scripts/prepare_dataset.py
2. Choose method — LoRA (parameter-efficient), QLoRA (memory-efficient), full (highest quality)
3. Configure LoRA — set r, lora_alpha, target_modules
4. Train — SFTTrainer with packing=True for efficiency
5. Merge — model.merge_and_unload() for deployment
6. Evaluate — perplexity, ROUGE-L, task benchmarks

See references/fine-tune-guide.md for complete LoRA/QLoRA/unsloth documentation.