fine-tuning-assistant（微调助手）

Fine-Tuning Assistant

The Fine-Tuning Assistant skill guides you through the process of adapting pre-trained models to your specific use case. Fine-tuning can dramatically improve model performance on specialized tasks, teach models your preferred style, and add capabilities that prompting alone cannot achieve.

This skill covers when to fine-tune versus prompt engineer, preparing training data, selecting base models, configuring training parameters, evaluating results, and deploying fine-tuned models. It applies modern techniques including LoRA, QLoRA, and instruction tuning to make fine-tuning practical and cost-effective.

Whether you are fine-tuning GPT models via API, running local training with open-source models, or using platforms like Hugging Face, this skill ensures you approach fine-tuning strategically and effectively.

Core Workflows

Workflow 1: Decide Whether to Fine-Tune

1. Assess the problem:

- Can prompting achieve the goal? - Is the task format or style consistent? - Do you have quality training data? - Is this worth the investment?

1. Compare approaches: Approach When to Use Investment Better prompts First attempt, variable tasks Low Few-shot examples Consistent format, limited data Low RAG Knowledge-intensive, dynamic data Medium Fine-tuning Consistent style, specialized task High
2. Evaluate requirements:

- Minimum 100-1000 quality examples - Clear evaluation criteria - Budget for training and hosting

1. Decision: Fine-tune only if prompting/RAG insufficient

Workflow 2: Prepare Fine-Tuning Dataset

1. Collect training examples:

- Representative of target use case - High quality (no errors in outputs) - Diverse coverage of task variations

1. Format for training: {"messages": [{"role": "system", "content": "You are a helpful assistant..."}, {"role": "user", "content": "User input here"}, {"role": "assistant", "content": "Ideal response here"}]}
2. Quality assurance:

- Review sample of examples manually - Check for consistency in style/format - Remove duplicates and low-quality entries

1. Split train/validation/test sets
2. Validate dataset format

Workflow 3: Execute Fine-Tuning

1. Select base model:

- Consider size vs capability tradeoff - Match model to task complexity - Check licensing for your use case

1. Configure training: # OpenAI fine-tuning training_config = {"model": "gpt-4o-mini-2024-07-18", "training_file": "file-xxx", "hyperparameters": {"n_epochs": 3, "batch_size": "auto", "learning_rate_multiplier": "auto"}} # LoRA fine-tuning (local) lora_config = {"r": 16, # Rank "lora_alpha": 32, "lora_dropout": 0.05, "target_modules": ["q_proj", "v_proj"]}
2. Monitor training:

- Watch loss curves - Check for overfitting - Validate on held-out set

1. Evaluate results:

- Compare to baseline model - Test on diverse inputs - Check for regressions

Quick Reference

Best Practices

• Start with Prompting: Fine-tuning is expensive; exhaust cheaper options first

- Can better prompts achieve 80% of the goal? - Try few-shot examples in the prompt - Consider RAG for knowledge tasks

• Quality Over Quantity: 100 excellent examples beat 10,000 mediocre ones

- Each example should be a gold standard - Better to have humans verify examples - Remove anything you wouldn't want the model to learn

• Match Format to Use Case: Training examples should mirror real usage

- Same prompt structure as production - Realistic input variations - Cover edge cases explicitly

• Don't Over-Train: More epochs isn't always better

- Watch validation loss for overfitting - Start with 1-3 epochs - Early stopping when validation plateaus

• Evaluate Properly: Training loss isn't the goal

- Use held-out test set - Compare to baseline on same tests - Check for capability regressions - Test on edge cases explicitly

• Version Everything: Fine-tuning is iterative

- Version your training data - Track experiment configurations - Document what worked and what didn't

Advanced Techniques

LoRA (Low-Rank Adaptation)

Efficient fine-tuning for large models:

from peft import LoraConfig, get_peft_model

lora_config = LoraConfig(
    r=16,                           # Rank of update matrices
    lora_alpha=32,                  # Scaling factor
    target_modules=["q_proj", "v_proj", "k_proj", "o_proj"],
    lora_dropout=0.05,
    bias="none",
    task_type="CAUSAL_LM"
)

# Apply LoRA to base model
model = get_peft_model(base_model, lora_config)

# Only ~0.1% of parameters are trainable
trainable_params = sum(p.numel() for p in model.parameters() if p.requires_grad)

QLoRA (Quantized LoRA)

Fine-tune large models on consumer hardware:

from transformers import BitsAndBytesConfig

bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=torch.bfloat16,
    bnb_4bit_use_double_quant=True
)

# Load model in 4-bit
model = AutoModelForCausalLM.from_pretrained(
    "meta-llama/Llama-2-7b-hf",
    quantization_config=bnb_config
)

# Apply LoRA on top
model = get_peft_model(model, lora_config)

Instruction Tuning Dataset Creation

Convert raw data to instruction format:

def create_instruction_example(raw_data):
    return {
        "messages": [
            {
                "role": "system",
                "content": "You are a customer service agent for TechCorp..."
            },
            {
                "role": "user",
                "content": f"Customer inquiry: {raw_data['inquiry']}"
            },
            {
                "role": "assistant",
                "content": raw_data['ideal_response']
            }
        ]
    }

# Apply to dataset
instruction_dataset = [create_instruction_example(d) for d in raw_dataset]

Evaluation Framework

Comprehensive assessment of fine-tuned models:

def evaluate_fine_tuned_model(model, test_set, baseline_model=None):
    results = {
        "task_accuracy": [],
        "format_compliance": [],
        "style_match": [],
        "regression_check": []
    }

    for example in test_set:
        output = model.generate(example.input)

        # Task-specific accuracy
        results["task_accuracy"].append(
            check_correctness(output, example.expected)
        )

        # Format compliance
        results["format_compliance"].append(
            matches_expected_format(output)
        )

        # Style matching (for style transfer tasks)
        results["style_match"].append(
            style_similarity(output, example.expected)
        )

        # Regression on general capabilities
        if baseline_model:
            results["regression_check"].append(
                compare_general_capability(model, baseline_model, example)
            )

    return {k: np.mean(v) for k, v in results.items()}

Curriculum Learning

Order training data by difficulty:

def create_curriculum(dataset):
    # Score examples by complexity
    scored = [(score_complexity(ex), ex) for ex in dataset]
    scored.sort(key=lambda x: x[0])

    # Create epochs with increasing difficulty
    n = len(scored)
    curriculum = {
        "epoch_1": [ex for _, ex in scored[:n//3]],           # Easy
        "epoch_2": [ex for _, ex in scored[:2*n//3]],         # Easy + Medium
        "epoch_3": [ex for _, ex in scored],                   # All
    }
    return curriculum

Common Pitfalls to Avoid

• Fine-tuning when better prompting would suffice
• Using low-quality or inconsistent training examples
• Not holding out a proper test set
• Training for too many epochs (overfitting)
• Ignoring capability regressions from fine-tuning
• Not versioning training data and configurations
• Expecting fine-tuning to add factual knowledge (use RAG instead)
• Fine-tuning on data that doesn't match production use