---
id: wiki-2026-0508-fine-tuning
title: Fine-tuning
category: 10_Wiki/Topics
status: verified
canonical_id: self
aliases: [fine-tuning, FT, LoRA, QLoRA, full fine-tune, instruction tuning, continual]
duplicate_of: none
source_trust_level: A
confidence_score: 0.98
verification_status: applied
tags: [machine-learning, fine-tuning, lora, qlora, transfer-learning, peft, instruction-tuning]
raw_sources: []
last_reinforced: 2026-05-10
github_commit: pending
tech_stack:
  language: Python
  framework: HuggingFace transformers / peft / TRL / Unsloth / Axolotl
---

# Fine-tuning

## 매 한 줄
> **"매 pretrained model 의 task / domain 의 의 adapt"**. 매 modern: 매 LoRA / QLoRA (PEFT) — 매 fraction parameter 의 update. 매 instruction tuning, RLHF, DPO. 매 alternative: prompt engineering, RAG. 매 cost: 매 GPU + data + eval.

## 매 핵심

### 매 spectrum
- **Full fine-tune**: 매 모든 weight.
- **PEFT** (parameter-efficient):
  - LoRA: 매 rank decomposition.
  - QLoRA: 매 4-bit quant + LoRA.
  - Adapter: 매 inserted layer.
  - IA³, Prefix-tuning, Prompt-tuning.
- **Instruction tuning**: 매 (prompt, response).
- **DPO / SimPO**: 매 preference.
- **RLHF**: 매 PPO + reward.

### 매 vs alternatives
- **Prompt engineering**: 매 cheapest, 매 limited.
- **Few-shot**: 매 ICL, 매 token cost.
- **RAG**: 매 fresh knowledge.
- **Fine-tune**: 매 style / format / domain skill.

### 매 응용
1. **Domain adapt**: 매 medical, legal.
2. **Style**: 매 brand voice.
3. **Format**: 매 JSON, structured.
4. **Tool use**: 매 function calling.
5. **Multilingual**: 매 low-resource.
6. **Safety**: 매 refuse harmful.

### 매 modern (2024+)
- **Unsloth**: 매 2x faster, lower VRAM.
- **Axolotl**: 매 YAML config.
- **TRL**: 매 SFT + DPO.
- **MLX-LM** (Apple): 매 on-device.

## 💻 패턴

### Full fine-tune (HF Trainer)
```python
from transformers import AutoModelForCausalLM, AutoTokenizer, Trainer, TrainingArguments

model = AutoModelForCausalLM.from_pretrained('meta-llama/Llama-3-8B')
tokenizer = AutoTokenizer.from_pretrained('meta-llama/Llama-3-8B')

args = TrainingArguments(
    output_dir='out', num_train_epochs=3,
    per_device_train_batch_size=4, gradient_accumulation_steps=4,
    learning_rate=1e-5, warmup_steps=100, fp16=True,
    save_steps=500, evaluation_strategy='steps', eval_steps=500,
)
Trainer(model=model, args=args, train_dataset=ds, eval_dataset=eval_ds).train()
```

### LoRA (peft)
```python
from peft import LoraConfig, get_peft_model

config = LoraConfig(
    r=16, lora_alpha=32, lora_dropout=0.1,
    target_modules=['q_proj', 'k_proj', 'v_proj', 'o_proj'],
    bias='none', task_type='CAUSAL_LM',
)
model = get_peft_model(model, config)
model.print_trainable_parameters()  # 매 ~0.1% of full
```

### QLoRA (4-bit)
```python
from transformers import BitsAndBytesConfig
bnb = BitsAndBytesConfig(
    load_in_4bit=True, bnb_4bit_quant_type='nf4',
    bnb_4bit_compute_dtype=torch.bfloat16, bnb_4bit_use_double_quant=True,
)
model = AutoModelForCausalLM.from_pretrained('llama-3-8b', quantization_config=bnb)
model = get_peft_model(model, lora_config)
```

### TRL SFT
```python
from trl import SFTTrainer

def format_instruction(ex):
    return f"### Instruction\n{ex['instruction']}\n### Response\n{ex['response']}"

trainer = SFTTrainer(
    model=model, args=args, train_dataset=ds,
    formatting_func=format_instruction, max_seq_length=2048,
)
trainer.train()
```

### DPO (preference)
```python
from trl import DPOTrainer

# 매 dataset: prompt, chosen, rejected
dpo = DPOTrainer(
    model=model, ref_model=ref_model,
    args=args, beta=0.1,
    train_dataset=preference_ds,
    tokenizer=tokenizer,
)
dpo.train()
```

### Unsloth (faster)
```python
from unsloth import FastLanguageModel

model, tokenizer = FastLanguageModel.from_pretrained(
    'unsloth/llama-3-8b-Instruct-bnb-4bit',
    max_seq_length=4096, load_in_4bit=True,
)
model = FastLanguageModel.get_peft_model(model, r=16, lora_alpha=32)
# 매 2x faster + lower memory
```

### Axolotl YAML
```yaml
base_model: meta-llama/Llama-3-8B
load_in_4bit: true
adapter: qlora
lora_r: 16
lora_alpha: 32
datasets:
  - path: my_data.jsonl
    type: alpaca
val_set_size: 0.05
sequence_len: 2048
gradient_accumulation_steps: 4
micro_batch_size: 4
num_epochs: 3
learning_rate: 2e-4
optimizer: adamw_bnb_8bit
```

### Merge LoRA back
```python
from peft import PeftModel
base = AutoModelForCausalLM.from_pretrained('llama-3-8b')
lora = PeftModel.from_pretrained(base, 'lora_checkpoint')
merged = lora.merge_and_unload()
merged.save_pretrained('llama3-finetuned')
```

### Eval (held-out)
```python
def eval_finetuned(model, tokenizer, eval_ds):
    correct = 0
    for ex in eval_ds:
        out = model.generate(**tokenizer(ex['prompt'], return_tensors='pt'), max_new_tokens=128)
        pred = tokenizer.decode(out[0], skip_special_tokens=True)
        if grade(pred, ex['answer']): correct += 1
    return correct / len(eval_ds)
```

### Prevent catastrophic forgetting
```python
# 매 mix in original-task data
def mixed_training(specific_data, general_data, ratio=0.2):
    mixed = list(specific_data) + random.sample(list(general_data), int(len(specific_data) * ratio))
    random.shuffle(mixed)
    return mixed
```

### Function calling fine-tune
```python
def format_function_call(ex):
    return {
        'prompt': f"User: {ex['user']}\n",
        'response': f"<tool_call>{json.dumps(ex['call'])}</tool_call>\n",
    }
```

### LR schedule (cosine + warmup)
```python
args = TrainingArguments(
    learning_rate=2e-4, warmup_ratio=0.03, lr_scheduler_type='cosine',
    num_train_epochs=3,
)
```

### Data quality (LIMA-style)
```python
# 매 LIMA 2023: 매 1000 high-quality > 매 100k noisy
def filter_quality(dataset, criteria):
    return [d for d in dataset if all(c(d) for c in criteria)]
```

### MLX-LM (Apple)
```bash
pip install mlx-lm
python -m mlx_lm.lora --model llama-3-8b --train --data data.jsonl --batch-size 1 --lora-layers 16
```

### Quantize-aware training
```python
# 매 GPTQ / AWQ for inference
from auto_gptq import AutoGPTQForCausalLM
quantized = AutoGPTQForCausalLM.from_pretrained(
    merged_model, quantize_config=BaseQuantizeConfig(bits=4, group_size=128),
)
```

## 매 결정 기준
| 상황 | Approach |
|---|---|
| Style / format | Prompt eng (try first) |
| Domain knowledge | RAG (try first) |
| Custom skill | LoRA fine-tune |
| Limited GPU | QLoRA (4-bit) |
| Production speed | Unsloth |
| Preference align | DPO |
| Best quality cost-aware | LoRA r=64 + DPO |

**기본값**: 매 prompt → RAG → LoRA. 매 GPU 부족 = QLoRA. 매 alignment = DPO. 매 quality = LIMA-style data.

## 🔗 Graph
- 부모: [[Machine-Learning]]
- 변형: [[LoRA]] · [[QLoRA]] · [[DPO]] · [[RLHF]] · [[Fine-tuning|Instruction-Tuning]]
- 응용: [[PEFT]] · [[Axolotl]]
- Adjacent: [[Catastrophic-Forgetting]] · [[Foundation-Models]] · [[RAG]] · [[Prompt_Engineering|Prompt-Engineering]]

## 🤖 LLM 활용
**언제**: 매 specific style / skill. 매 domain expert. 매 alignment.
**언제 X**: 매 quick fix (prompt). 매 fresh data (RAG).

## ❌ 안티패턴
- **Fine-tune for facts**: 매 RAG 의 use.
- **Tiny dataset full FT**: 매 catastrophic forget.
- **No eval baseline**: 매 regress 의 invisible.
- **Skip LoRA → full FT**: 매 overkill.
- **No quant inference**: 매 cost.

## 🧪 검증 / 중복
- Verified (Hu LoRA 2021, Dettmers QLoRA 2023, Rafailov DPO 2023, LIMA 2023).
- 신뢰도 A.

## 🕓 Changelog
| 날짜 | 변경 |
|---|---|
| 2026-04-20 | Auto-reinforced |
| 2026-05-08 | Phase 1 |
| 2026-05-10 | Manual cleanup — full / LoRA / QLoRA / DPO / Unsloth / Axolotl / merge code |