LoRA微调7B大模型：从数据准备到推理效果对比的完整记录

import json
from datasets import Dataset

def load_and_format(file_path):
    with open(file_path, 'r', encoding='utf-8') as f:
        raw = json.load(f)

    formatted = []
    for item in raw:
        # 统一成Qwen的chat template
        text = f"user\n{item['instruction']}\nassistant\n{item['output']}"
        formatted.append({"text": text})

    return Dataset.from_list(formatted)

train_ds = load_and_format("data/train.json")
val_ds = load_and_format("data/val.json")
print(f"Train: {len(train_ds)}, Val: {len(val_ds)}")
# 输出：Train: 2560, Val: 320

import torch
from transformers import (
    AutoModelForCausalLM, AutoTokenizer,
    BitsAndBytesConfig, TrainingArguments
)
from peft import LoraConfig, get_peft_model, prepare_model_for_kbit_training
from trl import SFTTrainer

model_name = "Qwen/Qwen2.5-7B-Instruct"

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

tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
tokenizer.pad_token = tokenizer.eos_token

model = AutoModelForCausalLM.from_pretrained(
    model_name,
    quantization_config=bnb_config,
    device_map="auto",
    trust_remote_code=True,
)
model = prepare_model_for_kbit_training(model)

lora_config = LoraConfig(
    r=16,
    lora_alpha=32,
    target_modules=["q_proj","k_proj","v_proj","o_proj",
                    "gate_proj","up_proj","down_proj"],
    lora_dropout=0.05,
    bias="none",
    task_type="CAUSAL_LM",
)

model = get_peft_model(model, lora_config)
model.print_trainable_parameters()
# 输出：trainable params: 51,380,224 || all params: 7,667,111,936 || trainable%: 0.67

training_args = TrainingArguments(
    output_dir="./qwen2.5-7b-lora",
    per_device_train_batch_size=4,
    gradient_accumulation_steps=4,
    num_train_epochs=3,
    learning_rate=2e-4,
    lr_scheduler_type="cosine",
    warmup_ratio=0.03,
    logging_steps=10,
    save_strategy="epoch",
    evaluation_strategy="epoch",
    bf16=True,
    optim="paged_adamw_8bit",
    report_to="none",
    save_total_limit=2,
)

trainer = SFTTrainer(
    model=model,
    args=training_args,
    train_dataset=train_ds,
    eval_dataset=val_ds,
    tokenizer=tokenizer,
    dataset_text_field="text",
    max_seq_length=1024,
    packing=False,
)

trainer.train()
trainer.model.save_pretrained("./qwen2.5-7b-lora-final")