Fine-Tuning LLMs: When, Why, and How

When to Fine-Tune vs. Use RAG

This is the most common question I get. Here’s my decision framework.

Fine-tune when:

• You need consistent formatting or style
• The task is well-defined and repetitive
• You have high-quality training data
• Latency is critical (no retrieval step)

Use RAG when:

• Knowledge changes frequently
• You need source attribution
• You have limited training data

Step-by-Step Fine-Tuning with Unsloth

Step 1: Install Unsloth

pip install unsloth

Step 2: Load Your Base Model

from unsloth import FastLanguageModel

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name="unsloth/Llama-3.2-3B-Instruct",
    max_seq_length=2048,
    dtype=None,
    load_in_4bit=True,
)

Step 3: Add LoRA Adapters

model = FastLanguageModel.get_peft_model(
    model,
    r=16,
    target_modules=["q_proj", "k_proj", "v_proj", "o_proj"],
    lora_alpha=16,
    lora_dropout=0,
    bias="none",
    use_gradient_checkpointing=True,
)

Step 4: Prepare Your Data

Quality > Quantity. 100 excellent examples often outperform 10,000 mediocre ones.

Step 5: Train

from trl import SFTTrainer

trainer = SFTTrainer(
    model=model,
    tokenizer=tokenizer,
    train_dataset=dataset,
    args=TrainingArguments(
        per_device_train_batch_size=2,
        gradient_accumulation_steps=4,
        learning_rate=2e-4,
        num_train_epochs=3,
    ),
)

Cost Considerations

Conclusion

Fine-tuning is a powerful tool, but it’s not always the right one. Start with prompt engineering, graduate to RAG if you need external knowledge, and fine-tune only when you have a clear, measurable benefit.