config.py

import torch
from typing import List
from dataclasses import dataclass, field


class Config:
    # model 参数 ###########################
    # 文本生成模型,下载地址 https://huggingface.co/Qwen/Qwen1.5-0.5B-Chat
    gpt_model = "E:\\ai_model\\model\\qwen0.5"
    data_path = "E:\\ai_model\\RLHF\\my_code\\RLHF_ORPO/data/train_data.json"
    save_lora_path = "E:\\ai_model\\model\\orpo\\save_lora"
    device = "cuda:0" if torch.cuda.is_available() else "cpu"
    batch_size = 2
    epochs = 30
    lr = 0.001
    # ORPO 参数 ############################
    alpha = 3

@dataclass
class LoraArguments:
    lora_r: int = 2
    lora_alpha: int = 8
    lora_dropout: float = 0
    lora_target_modules: List[str] = field(
        default_factory=lambda: ['k_proj', 'v_proj']
    )
    # lora_target_modules = None
    lora_weight_path: str = ""
    q_lora: bool = False
    load_in_4bit: bool = False
    load_in_8bit: bool = False
    is_reload_trained_params = True  # 是否接着上次训练模型继续训练