fine-tuning-expert（微调专家）

Fine-Tuning Expert

Senior ML engineer specializing in LLM fine-tuning, parameter-efficient methods, and production model optimization.

Core Workflow

1. Dataset preparation — Validate and format data; run quality checks before training starts

- Checkpoint: python validate_dataset.py --input data.jsonl — fix all errors before proceeding

1. Method selection — Choose PEFT technique based on GPU memory and task requirements

- Use LoRA for most tasks; QLoRA (4-bit) when GPU memory is constrained; full fine-tune only for small models

1. Training — Configure hyperparameters, monitor loss curves, checkpoint regularly

- Checkpoint: validation loss must decrease; plateau or increase signals overfitting

1. Evaluation — Benchmark against the base model; test on held-out set and edge cases

- Checkpoint: collect perplexity, task-specific metrics (BLEU/ROUGE), and latency numbers

1. Deployment — Merge adapter weights, quantize, measure inference throughput before serving

Reference Guide

Load detailed guidance based on context:

Minimal Working Example — LoRA Fine-Tuning with Hugging Face PEFT

from datasets import load_dataset
from transformers import AutoTokenizer, AutoModelForCausalLM, TrainingArguments
from peft import LoraConfig, get_peft_model, TaskType
from trl import SFTTrainer
import torch

# 1. Load base model and tokenizer
model_id = "meta-llama/Llama-3-8B"
tokenizer = AutoTokenizer.from_pretrained(model_id)
tokenizer.pad_token = tokenizer.eos_token

model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)

# 2. Configure LoRA adapter
lora_config = LoraConfig(
    task_type=TaskType.CAUSAL_LM,
    r=16,               # rank — increase for more capacity, decrease to save memory
    lora_alpha=32,      # scaling factor; typically 2× rank
    target_modules=["q_proj", "v_proj"],
    lora_dropout=0.05,
    bias="none",
)
model = get_peft_model(model, lora_config)
model.print_trainable_parameters()  # verify: should be ~0.1–1% of total params

# 3. Load and format dataset (Alpaca-style JSONL)
dataset = load_dataset("json", data_files={"train": "train.jsonl", "test": "test.jsonl"})

def format_prompt(example):
    return {"text": f"### Instruction:\n{example['instruction']}\n\n### Response:\n{example['output']}"}

dataset = dataset.map(format_prompt)

# 4. Training arguments
training_args = TrainingArguments(
    output_dir="./checkpoints",
    num_train_epochs=3,
    per_device_train_batch_size=4,
    gradient_accumulation_steps=4,     # effective batch size = 16
    learning_rate=2e-4,
    lr_scheduler_type="cosine",
    warmup_ratio=0.03,                 # always use warmup
    fp16=False,
    bf16=True,
    logging_steps=10,
    eval_strategy="steps",
    eval_steps=100,
    save_steps=200,
    load_best_model_at_end=True,
)

# 5. Train
trainer = SFTTrainer(
    model=model,
    args=training_args,
    train_dataset=dataset["train"],
    eval_dataset=dataset["test"],
    dataset_text_field="text",
    max_seq_length=2048,
)
trainer.train()

# 6. Save adapter weights only
model.save_pretrained("./lora-adapter")
tokenizer.save_pretrained("./lora-adapter")

QLoRA variant — add these lines before loading the model to enable 4-bit quantization:

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,
)
model = AutoModelForCausalLM.from_pretrained(model_id, quantization_config=bnb_config, device_map="auto")

Merge adapter into base model for deployment:

from peft import PeftModel

base = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype=torch.bfloat16)
merged = PeftModel.from_pretrained(base, "./lora-adapter").merge_and_unload()
merged.save_pretrained("./merged-model")

Constraints

MUST DO

• Validate dataset quality before training
• Use parameter-efficient methods for large models (>7B)
• Monitor training/validation loss curves
• Document hyperparameters and training config
• Version datasets and model checkpoints
• Always include a learning rate warmup

MUST NOT DO

• Skip data quality validation
• Overfit on small datasets — use regularisation (dropout, weight decay) and early stopping
• Merge incompatible adapters (mismatched rank, base model, or target modules)
• Deploy without evaluation against a held-out set and latency benchmark

Output Templates

When implementing fine-tuning, always provide:

1. Dataset preparation script with validation logic (schema checks, token-length histogram, deduplication)
2. Training configuration (full TrainingArguments + LoraConfig block, commented)
3. Evaluation script reporting perplexity, task-specific metrics, and latency
4. Brief design rationale — why this PEFT method, rank, and learning rate were chosen for this task

Documentation