训练阶段

from transformers import AutoModelForCausalLM, AutoTokenizer, TrainingArguments, Trainer
from datasets import load_dataset

# 加载预训练模型与分词器
model_name = "meta-llama/Llama-2-7b-hf"
model = AutoModelForCausalLM.from_pretrained(model_name)
tokenizer = AutoTokenizer.from_pretrained(model_name)

# 加载指令数据集
dataset = load_dataset("json", data_files="sft_data.json")

def tokenize_function(examples):
    return tokenizer(examples["text"], truncation=True, padding="max_length", max_length=512)

tokenized_dataset = dataset.map(tokenize_function, batched=True)

# 配置训练参数
training_args = TrainingArguments(
    output_dir="./sft_output",
    num_train_epochs=3,
    per_device_train_batch_size=4,
    learning_rate=2e-5,
    save_steps=500,
    logging_steps=100
)

# 初始化训练器并执行训练
trainer = Trainer(
    model=model,
    args=training_args,
    train_dataset=tokenized_dataset["train"]
)
trainer.train()

使用 Transformers 进行指令微调的最简示例

{
  "model_path": "./sft_output/checkpoint-1500",
  "training_metrics": {
    "train_loss": 0.82,
    "eval_loss": 0.91,
    "epochs_completed": 3,
    "steps_completed": 1500
  },
  "model_config": {
    "base_model": "meta-llama/Llama-2-7b-hf",
    "training_type": "SFT",
    "precision": "FP16",
    "max_length": 512
  },
  "deployment_ready": true
}