mirror of
https://github.com/xming521/WeClone.git
synced 2026-08-28 18:07:28 +08:00
xming521
a0829d2098
Introduce a ThreadPoolExecutor in the OnlineLLM class to enable concurrent API calls, significantly boosting throughput for LLM operations. Refactor the OlineLLMCleaningStrategy to leverage this new batching capability, allowing multiple data points to be processed in parallel. Increase the default `clean_batch_size` in settings and enhance `n_process`/`batch_size` for PII detection to optimize for the new concurrency. Simplify prompt management by removing a dedicated prompt for online LLM cleaning. Add context manager support to OnlineLLM for reliable resource cleanup. Ensure LLM chat responses explicitly request JSON format.
101 lines
3.5 KiB
JSON
101 lines
3.5 KiB
JSON
{
|
|
"version": "0.3.01",
|
|
"common_args": {
|
|
"model_name_or_path": "./models/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",
|
|
"media_dir": "dataset/media",
|
|
"finetuning_type": "lora",
|
|
"enable_thinking": false,
|
|
"trust_remote_code": true
|
|
},
|
|
"cli_args": {
|
|
"full_log": false,
|
|
"log_level": "INFO"
|
|
},
|
|
"make_dataset_args": {
|
|
//数据处理配置
|
|
"platform": "chat", //chat,telegram
|
|
"language": "zh", // 聊天常用语言: zh(中文) 或 en(英文)
|
|
"telegram_args": {
|
|
"my_id": "user1234567890"
|
|
},
|
|
"include_type": [
|
|
"text"
|
|
],
|
|
"blocked_words": [ // 禁用词
|
|
"例如 姓名",
|
|
"例如 密码",
|
|
"//....."
|
|
],
|
|
"add_time": false,
|
|
"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": 2048, // 组合后消息最大长度 配合cutoff_len 使用
|
|
"messages_max_length": 2048, // messages最长字符数量 配合cutoff_len 使用
|
|
"clean_dataset": {
|
|
"enable_clean": false,
|
|
"clean_strategy": "llm",
|
|
"llm": {
|
|
"accept_score": 2, //可以接受的llm打分阈值,1分最差，5分最好,低于此分数的数据不会用于训练
|
|
"enable_thinking": true
|
|
}
|
|
},
|
|
"online_llm_clear": false,
|
|
"base_url": "https://xxx/v1",
|
|
"llm_api_key": "xxxxx",
|
|
"model_name": "xxx", //建议使用参数较大的模型，例如DeepSeek-V3
|
|
"clean_batch_size": 50,
|
|
"vision_api": {
|
|
"enable": false, // 设置为 true 来开启此功能
|
|
"api_key": "xxx",
|
|
"api_url": "https://xxx/v1", // 兼容OpenAI的API地址
|
|
"model_name": "xxx", // 要使用的多模态模型名称,例如qwen-vl-max
|
|
"max_workers": 5 // 并行调用API的线程数，最多不要超过8
|
|
}
|
|
},
|
|
"train_sft_args": {
|
|
//微调配置
|
|
"stage": "sft",
|
|
"dataset": "chat-sft",
|
|
"dataset_dir": "./dataset/res_csv/sft",
|
|
"use_fast_tokenizer": true,
|
|
"lora_target": "q_proj,v_proj",
|
|
"lora_rank": 8,
|
|
"lora_dropout": 0.25,
|
|
"weight_decay": 0.1,
|
|
"overwrite_cache": true,
|
|
"per_device_train_batch_size": 2,
|
|
"gradient_accumulation_steps": 16,
|
|
"lr_scheduler_type": "cosine",
|
|
"cutoff_len": 2048,
|
|
"logging_steps": 10,
|
|
"save_steps": 100,
|
|
"learning_rate": 1e-4,
|
|
"warmup_ratio": 0.1,
|
|
"num_train_epochs": 2,
|
|
"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": 256,
|
|
"top_p": 0.65
|
|
},
|
|
"vllm_args": {
|
|
"gpu_memory_utilization": 0.9,
|
|
// "data_parallel_size": 2,
|
|
// "quantization": "bitsandbytes",
|
|
// "load_format": "bitsandbytes"
|
|
},
|
|
"test_model_args": {
|
|
"test_data_path": "dataset/eval/test_data-zh.json"
|
|
}
|
|
}
|