{
    "train_pt_args": {
        "stage": "pt",
        "dataset": "wechat-pt",
        "dataset_dir": "./dataset/res_csv/pt",
        "lora_target": "q_proj,v_proj",
        "lora_rank": 2,
        "lora_dropout": 0.1,
        "output_dir": "model_output",
        "overwrite_cache": true,
        "per_device_train_batch_size": 1,
        "gradient_accumulation_steps": 1,
        "lr_scheduler_type": "cosine",
        "logging_steps": 10,
        "save_steps": 1000,
        "learning_rate": 0.001,
        "num_train_epochs": 30,
        "plot_loss": true,
        "fp16": true
    },
    "train_sft_args": {
        "stage": "sft",
        "dataset": "wechat-sft",
        "dataset_dir": "./dataset/res_csv/sft",
        "use_fast_tokenizer": true,
        "lora_target": "q_proj,v_proj",
        "lora_rank": 4,
        "lora_dropout": 0.4,
        "weight_decay": 0.1,
        "overwrite_cache": true, 
        "per_device_train_batch_size": 8,
        "gradient_accumulation_steps": 4,
        "lr_scheduler_type": "cosine",
        "cutoff_len": 512, //后续自动计算
        "logging_steps": 10,
        "save_steps": 100,
        "learning_rate": 1e-4,
        "warmup_ratio": 0.1,
        "num_train_epochs": 3,
        "plot_loss": true,
        "fp16": true,
        "flash_attn": "fa2",
        // "deepspeed": "ds_config.json" //多卡训练
    },
    "infer_args": {
        "repetition_penalty": 1.2,
        "temperature": 0.5,
        "max_length": 50,
        "top_p": 0.65
    },
    "make_dataset_args": {
        // "enable_vision_model": false,//后续实现
        // "include_type": [
        //     "文本"
        // ],
        "single_combine_strategy": "time_window", // 单人组成单句策略
        "qa_match_strategy": "time_window", // 组成qa策略
        "single_combine_time_window": 2, // 单人组成单句时间窗口（分钟）,
        "qa_match_time_window": 5, // 组成qa时间窗口（分钟）,
        "combine_msg_max_length": 256, // 组合后消息最大长度 配合cutoff_len 使用
        "prompt_with_history": false // 是否在prompt中包含历史对话
    },
    "common_args": {
        "model_name_or_path": "./Qwen2.5-7B-Instruct",
        "adapter_name_or_path": "./model_output", //同时做为train_sft_args的output_dir
        "template": "qwen",
        "default_system": "请你扮演一名人类，不要说自己是人工智能",
        "finetuning_type": "lora",
        "trust_remote_code": true
    }
}