从零用LoRA微调Qwen2.5-7B：数据构造、训练参数与loss曲线全记录

{
  "messages": [
    {"role": "system", "content": "你是一个XX领域助手，回答需简洁准确。"},
    {"role": "user", "content": "问题..."},
    {"role": "assistant", "content": "标准答案..."}
  ]
}

from datasets import load_dataset
from transformers import AutoTokenizer

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

def format_and_tokenize(example, max_len=1024):
    text = tokenizer.apply_chat_template(
        example["messages"],
        tokenize=False,
        add_generation_prompt=False
    )
    out = tokenizer(
        text,
        truncation=True,
        max_length=max_len,
        padding=False
    )
    return out

ds = load_dataset("json", data_files={"train": "train.jsonl", "valid": "valid.jsonl"})
ds = ds.map(format_and_tokenize, remove_columns=ds["train"].column_names)

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,
    attn_implementation="flash_attention_2",  # 需要装flash-attn
)
model.gradient_checkpointing_enable()
model.enable_input_require_grads()

lora_config = LoraConfig(
    r=16,
    lora_alpha=32,
    lora_dropout=0.05,
    bias="none",
    task_type="CAUSAL_LM",
    target_modules=["q_proj","k_proj","v_proj","o_proj",
                    "gate_proj","up_proj","down_proj"],
)
model = get_peft_model(model, lora_config)
model.print_trainable_parameters()
# trainable params: 40,370,176 || all params: 7,655,986,688 || trainable%: 0.527

args = TrainingArguments(
    output_dir="./qwen_lora_out",
    per_device_train_batch_size=2,
    gradient_accumulation_steps=8,     # 等效batch=16
    learning_rate=1e-4,
    lr_scheduler_type="cosine",
    warmup_ratio=0.03,
    num_train_epochs=3,
    bf16=True,
    logging_steps=10,
    save_strategy="epoch",
    eval_strategy="epoch",
    gradient_checkpointing=True,
    optim="adamw_torch",
    report_to="tensorboard",
    max_seq_length=1024,
)

trainer = SFTTrainer(
    model=model,
    args=args,
    train_dataset=ds["train"],
    eval_dataset=ds["valid"],
)
trainer.train()

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base = AutoModelForCausalLM.from_pretrained(
    model_path, torch_dtype=torch.bfloat16, device_map="auto"
)
model = PeftModel.from_pretrained(base, "./qwen_lora_out/checkpoint-2250")
model = model.merge_and_unload()  # 合并后可当普通模型用
tokenizer = AutoTokenizer.from_pretrained(model_path)

messages = [
    {"role": "system", "content": "你是一个XX领域助手，回答需简洁准确。"},
    {"role": "user", "content": "XX工艺的临界温度是多少？"},
]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)
out = model.generate(**inputs, max_new_tokens=256,
                     do_sample=False, repetition_penalty=1.1)
print(tokenizer.decode(out[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))