Create train.py

train.py

@@ -0,0 +1,58 @@
+import torch
+from transformers import AutoModelForCausalLM, AutoTokenizer, Trainer, TrainingArguments
+from peft import LoraConfig, get_peft_model
+from datasets import load_dataset
+
+# 加载DeepSeek R1模型
+model_name = "DeepSeek/R1"  # 你可以根据实际选择不同的路径
+
+# 加载模型和分词器
+model = AutoModelForCausalLM.from_pretrained(model_name)
+tokenizer = AutoTokenizer.from_pretrained(model_name)
+
+# 配置 LoRA 微调
+lora_config = LoraConfig(
+    r=8,  # LoRA参数
+    lora_alpha=16,
+    lora_dropout=0.1,
+    bias="none",
+)
+
+# 获取LoRA微调模型
+model = get_peft_model(model, lora_config)
+
+# 准备数据集
+data = [
+    {"instruction": "粉丝通跨店版的费用是多少？", "output": "粉丝通跨店版按月付费，500元/月，仅提供增值税普通电子发票。"},
+    {"instruction": "如何充值粉丝通软件的红包？", "output": "商家可以灵活充值红包，每个红包最低0.1元，具体总额根据拉新目标决定。"},
+    {"instruction": "红包的扣费机制是怎样的？", "output": "红包在用户实际使用后才会扣款，未使用到期会自动退回商家公户。"},
+    # 你可以继续添加数据...
+]
+
+# 转换数据为 Hugging Face 数据集格式
+train_data = [{"input_ids": tokenizer.encode(d["instruction"], truncation=True, padding="max_length"), "labels": tokenizer.encode(d["output"], truncation=True, padding="max_length")} for d in data]
+
+train_dataset = load_dataset('json', data_files={'train': train_data})
+
+# 设置训练参数
+training_args = TrainingArguments(
+    output_dir='./results',
+    evaluation_strategy="epoch",
+    learning_rate=5e-5,
+    per_device_train_batch_size=2,
+    per_device_eval_batch_size=2,
+    num_train_epochs=3,
+    weight_decay=0.01,
+    save_steps=10_000,
+    save_total_limit=2,
+)
+
+# 设置 Trainer
+trainer = Trainer(
+    model=model,
+    args=training_args,
+    train_dataset=train_dataset["train"],
+)
+
+# 开始训练
+trainer.train()