Delete fine_tune.py

fine_tune.py

@@ -1,77 +0,0 @@
-import torch
-from datasets import load_dataset
-from peft import LoraConfig, get_peft_model, prepare_model_for_kbit_training
-from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig, TrainingArguments
-from trl import SFTTrainer
-
-# 1. 加载模型和分词器 (CPU优化版本)
-# 使用更小的模型以适配CPU环境
-model_name = "microsoft/DialoGPT-small"  # 更小的模型，适合CPU训练
-
-# CPU环境下不需要量化配置
-model = AutoModelForCausalLM.from_pretrained(
-    model_name,
-    torch_dtype=torch.float32,  # CPU使用float32
-    low_cpu_mem_usage=True,     # 优化CPU内存使用
-)
-model.config.use_cache = False
-
-tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
-tokenizer.pad_token = tokenizer.eos_token # set pad token
-
-# 2. 加载并准备数据集
-def formatting_prompts_func(example):
-    output_texts = []
-    for i in range(len(example['instruction'])):
-        text = f"### Instruction:\n{example['instruction'][i]}\n\n### Input:\n{example['input'][i]}\n\n### Response:\n{example['output'][i]}"
-        output_texts.append(text)
-    return output_texts
-
-dataset = load_dataset("json", data_files="data.json", split="train")
-
-# 3. 配置LoRA参数 (适配DialoGPT)
-lora_config = LoraConfig(
-    r=8, # Rank
-    lora_alpha=32,
-    lora_dropout=0.1,
-    bias="none",
-    task_type="CAUSAL_LM",
-    target_modules=["c_attn", "c_proj"],  # DialoGPT/GPT-2 架构的注意力模块
-)
-
-# 4. 创建PEFT模型 (CPU版本)
-# CPU环境下不需要量化准备
-model = get_peft_model(model, lora_config)
-
-# 5. 配置训练参数 (CPU优化)
-output_dir = "./dialogpt-small-lora"
-training_args = TrainingArguments(
-    output_dir=output_dir,
-    per_device_train_batch_size=1,  # CPU环境使用更小的批次
-    gradient_accumulation_steps=8,   # 增加梯度累积以补偿小批次
-    learning_rate=5e-4,             # 稍微提高学习率
-    logging_steps=5,
-    max_steps=50,                   # 减少训练步数用于演示
-    save_strategy="steps",
-    save_steps=25,
-    dataloader_num_workers=0,       # CPU环境下设为0
-    fp16=False,                     # CPU不支持fp16
-    report_to=None,                 # 禁用wandb等报告
-)
-
-# 6. 创建Trainer并开始训练
-trainer = SFTTrainer(
-    model=model,
-    train_dataset=dataset,
-    args=training_args,
-    peft_config=lora_config,
-    formatting_func=formatting_prompts_func,
-    max_seq_length=512,
-)
-
-trainer.train()
-
-# 7. 保存模型
-print("Saving DialoGPT LoRA adapter...")
-trainer.save_model(output_dir)
-print(f"DialoGPT LoRA adapter saved to {output_dir}")