fine-tuning-customization（微调定制）

Fine-Tuning & Customization

Customize LLMs for specific domains using parameter-efficient fine-tuning and alignment techniques.

Unsloth : 7x longer context RL, FP8 RL on consumer GPUs, rsLoRA support. TRL: OpenEnv integration, vLLM server mode, transformers 5.0.0+ compatible.

Decision Framework: Fine-Tune or Not?

Fine-tune ONLY when:

1. Prompt engineering tried and insufficient
2. RAG doesn't capture domain nuances
3. Specific output format consistently required
4. Persona/style must be deeply embedded
5. You have ~1000+ high-quality examples

LoRA vs QLoRA (Unsloth)

Quick Reference: LoRA Training

from unsloth import FastLanguageModel
from trl import SFTTrainer

# Load with 4-bit quantization (QLoRA)
model, tokenizer = FastLanguageModel.from_pretrained(
    model_name="unsloth/Meta-Llama-3.1-8B",
    max_seq_length=2048,
    load_in_4bit=True,
)

# Add LoRA adapters
model = FastLanguageModel.get_peft_model(
    model,
    r=16,              # Rank (16-64 typical)
    lora_alpha=32,     # Scaling (2x r)
    lora_dropout=0.05,
    target_modules=[
        "q_proj", "k_proj", "v_proj", "o_proj",  # Attention
        "gate_proj", "up_proj", "down_proj",      # MLP (QLoRA paper)
    ],
)

# Train
trainer = SFTTrainer(
    model=model,
    tokenizer=tokenizer,
    train_dataset=dataset,
    max_seq_length=2048,
)
trainer.train()

DPO Alignment

from trl import DPOTrainer, DPOConfig

config = DPOConfig(
    learning_rate=5e-6,  # Lower for alignment
    beta=0.1,            # KL penalty coefficient
    per_device_train_batch_size=4,
    num_train_epochs=1,
)

# Preference dataset: {prompt, chosen, rejected}
trainer = DPOTrainer(
    model=model,
    ref_model=ref_model,  # Frozen reference
    args=config,
    train_dataset=preference_dataset,
    tokenizer=tokenizer,
)
trainer.train()

Synthetic Data Generation

async def generate_synthetic(topic: str, n: int = 100) -> list[dict]:
    """Generate training examples using teacher model."""
    examples = []
    for _ in range(n):
        response = await client.chat.completions.create(
            model="gpt-5.2",  # Teacher
            messages=[{
                "role": "system",
                "content": f"Generate a training example about {topic}. "
                          "Include instruction and response."
            }],
            response_format={"type": "json_object"}
        )
        examples.append(json.loads(response.choices[0].message.content))
    return examples

Key Hyperparameters

Anti-Patterns (FORBIDDEN)

# NEVER fine-tune without trying alternatives first
model.fine_tune(data)  # Try prompt engineering & RAG first!

# NEVER use low-quality training data
data = scrape_random_web()  # Garbage in, garbage out

# NEVER skip evaluation
trainer.train()
deploy(model)  # Always evaluate before deploy!

# ALWAYS use separate eval set
train, eval = split(data, test_size=0.1)
trainer = SFTTrainer(..., eval_dataset=eval)

Detailed Documentation

Related Skills

• llm-evaluation - Evaluate fine-tuned models
• embeddings - When to use embeddings instead
• rag-retrieval - When RAG is better than fine-tuning
• langfuse-observability - Track training experiments

Capability Details

lora-qlora

Keywords: LoRA, QLoRA, PEFT, parameter efficient, adapter, low-rank Solves:

• Fine-tune large models on consumer hardware
• Configure LoRA hyperparameters
• Choose target modules for adapters

dpo-alignment

Keywords: DPO, RLHF, preference, alignment, human feedback, preference data Solves:

• Align models to human preferences
• Create preference datasets
• Configure DPO training

synthetic-data

Keywords: synthetic data, data generation, teacher model, distillation Solves:

• Generate training data with LLMs
• Implement teacher-student training
• Scale training data quality

when-to-finetune

Keywords: should I fine-tune, fine-tune decision, customize model Solves:

• Decide when fine-tuning is appropriate
• Evaluate alternatives to fine-tuning
• Assess data requirements