LoRA微调7B模型全流程：从数据准备到Loss曲线与推理效果对比

import json
from datasets import Dataset
from transformers import AutoTokenizer

MODEL_PATH = "Qwen/Qwen2.5-7B-Instruct"
tokenizer = AutoTokenizer.from_pretrained(MODEL_PATH, trust_remote_code=True)

def build_sample(item):
    # item: {"instruction":..., "input":..., "output":...}
    user_content = item["instruction"]
    if item.get("input"):
        user_content += "\n" + item["input"]
    messages = [
        {"role": "user", "content": user_content},
        {"role": "assistant", "content": item["output"]},
    ]
    text = tokenizer.apply_chat_template(
        messages, tokenize=False, add_generation_prompt=False
    )
    return {"text": text}

with open("train.jsonl", "r", encoding="utf-8") as f:
    raw = [json.loads(line) for line in f]

dataset = Dataset.from_list([build_sample(x) for x in raw])
split = dataset.train_test_split(test_size=0.05, seed=42)
train_ds, eval_ds = split["train"], split["test"]
print(f"train={len(train_ds)}, eval={len(eval_ds)}")
# 输出: train=3610, eval=190

import torch
from transformers import AutoModelForCausalLM, TrainingArguments
from peft import LoraConfig, get_peft_model
from trl import SFTTrainer

model = AutoModelForCausalLM.from_pretrained(
    MODEL_PATH,
    torch_dtype=torch.bfloat16,
    device_map="auto",
    trust_remote_code=True,
)
model.gradient_checkpointing_enable()
model.enable_input_require_grads()  # 配合gradient checkpointing

lora_config = LoraConfig(
    r=16,
    lora_alpha=32,
    lora_dropout=0.05,
    target_modules=["q_proj","k_proj","v_proj","o_proj",
                    "gate_proj","up_proj","down_proj"],
    bias="none",
    task_type="CAUSAL_LM",
)
model = get_peft_model(model, lora_config)
model.print_trainable_parameters()
# trainable params: 20,185,088 || all params: 7,635,900,416 || trainable%: 0.2643

args = TrainingArguments(
    output_dir="./lora_out",
    per_device_train_batch_size=2,
    gradient_accumulation_steps=8,
    num_train_epochs=2,
    learning_rate=2e-4,
    lr_scheduler_type="cosine",
    warmup_ratio=0.03,
    logging_steps=10,
    eval_strategy="steps",
    eval_steps=100,
    save_strategy="steps",
    save_steps=200,
    save_total_limit=2,
    bf16=True,
    gradient_checkpointing=True,
    optim="paged_adamw_8bit",
    report_to="none",
)

trainer = SFTTrainer(
    model=model,
    args=args,
    train_dataset=train_ds,
    eval_dataset=eval_ds,
    dataset_text_field="text",
    max_seq_length=1024,
    packing=False,
)
trainer.train()
trainer.model.save_pretrained("./lora_adapter")
tokenizer.save_pretrained("./lora_adapter")