DB-GPT-Hub 项目发布了pip包，用来降低Text2SQL训练的门槛, 除了通过仓库中提供的脚本的方式进行微调之外，还可以使用我们提供的Python包进行微调。

安装

  1. pip install dbgpt_hub

查看基线

  1. from dbgpt_hub.baseline import show_scores
  2. show_scores()

使用dbgpt_hub包 - 图1

微调

  1. from dbgpt_hub.data_process import preprocess_sft_data
  2. from dbgpt_hub.train import start_sft
  3. from dbgpt_hub.predict import start_predict
  4. from dbgpt_hub.eval import start_evaluate
  6. data_folder = "dbgpt_hub/data"
  7. data_info = [
  8.     {
  9.         "data_source": "spider",
  10.         "train_file": ["train_spider.json", "train_others.json"],
  11.         "dev_file": ["dev.json"],
  12.         "tables_file": "tables.json",
  13.         "db_id_name": "db_id",
  14.         "is_multiple_turn": False,
  15.         "train_output": "spider_train.json",
  16.         "dev_output": "spider_dev.json",
  17.     }
  18. ]
  20. train_args = {
  21.             "model_name_or_path": "codellama/CodeLlama-13b-Instruct-hf",
  22.             "do_train": True,
  23.             "dataset": "example_text2sql_train",
  24.             "max_source_length": 2048,
  25.             "max_target_length": 512,
  26.             "finetuning_type": "lora",
  27.             "lora_target": "q_proj,v_proj",
  28.             "template": "llama2",
  29.             "lora_rank": 64,
  30.             "lora_alpha": 32,
  31.             "output_dir": "dbgpt_hub/output/adapter/CodeLlama-13b-sql-lora",
  32.             "overwrite_cache": True,
  33.             "overwrite_output_dir": True,
  34.             "per_device_train_batch_size": 1,
  35.             "gradient_accumulation_steps": 16,
  36.             "lr_scheduler_type": "cosine_with_restarts",
  37.             "logging_steps": 50,
  38.             "save_steps": 2000,
  39.             "learning_rate": 2e-4,
  40.             "num_train_epochs": 8,
  41.             "plot_loss": True,
  42.             "bf16": True,
  43. }
  45. predict_args = {
  46.             "model_name_or_path": "codellama/CodeLlama-13b-Instruct-hf",
  47.             "template": "llama2",
  48.             "finetuning_type": "lora",
  49.             "checkpoint_dir": "dbgpt_hub/output/adapter/CodeLlama-13b-sql-lora",
  50.             "predict_file_path": "dbgpt_hub/data/eval_data/dev_sql.json",
  51.             "predict_out_dir": "dbgpt_hub/output/",
  52.             "predicted_out_filename": "pred_sql.sql",
  53. }
  55. evaluate_args =  {
  56.             "input": "./dbgpt_hub/output/pred/pred_sql_dev_skeleton.sql",
  57.             "gold": "./dbgpt_hub/data/eval_data/gold.txt",
  58.             "gold_natsql": "./dbgpt_hub/data/eval_data/gold_natsql2sql.txt",
  59.             "db": "./dbgpt_hub/data/spider/database",
  60.             "table": "./dbgpt_hub/data/eval_data/tables.json",
  61.             "table_natsql": "./dbgpt_hub/data/eval_data/tables_for_natsql2sql.json",
  62.             "etype": "exec",
  63.             "plug_value": True,
  64.             "keep_distict": False,
  65.             "progress_bar_for_each_datapoint": False,
  66.             "natsql": False,
  67. }
  69. preprocess_sft_data(
  70.       data_folder = data_folder,
  71.       data_info = data_info
  72. )
  74. start_sft(train_args)
  75. start_predict(predict_args)
  76. start_evaluate(evaluate_args)