{"id":437,"date":"2026-03-06T13:23:42","date_gmt":"2026-03-06T05:23:42","guid":{"rendered":"https:\/\/www.liaoxinghui.com\/?p=437"},"modified":"2026-03-06T13:23:42","modified_gmt":"2026-03-06T05:23:42","slug":"fine-tune-bert-huggingface-customer-service-chatbot","status":"publish","type":"post","link":"https:\/\/www.liaoxinghui.com\/?p=437","title":{"rendered":"\u4f7f\u7528Hugging Face\u5fae\u8c03BERT\u6a21\u578b\u6784\u5efa\u667a\u80fd\u5ba2\u670d\u804a\u5929\u673a\u5668\u4eba"},"content":{"rendered":"

\u4f7f\u7528Hugging Face\u5fae\u8c03BERT\u6a21\u578b\u6784\u5efa\u667a\u80fd\u5ba2\u670d\u804a\u5929\u673a\u5668\u4eba<\/h1>\n

1. \u4e1a\u52a1\u573a\u666f\u4e0e\u76ee\u6807<\/h2>\n

\u7535\u5546\u5e73\u53f0\u5ba2\u670d\u7cfb\u7edf\u6bcf\u5929\u9762\u4e34\u5927\u91cf\u91cd\u590d\u6027\u95ee\u9898\uff0c\u5982\u8ba2\u5355\u67e5\u8be2\u6216\u9000\u8d27\u653f\u7b56\u54a8\u8be2\uff0c\u4eba\u5de5\u5904\u7406\u6210\u672c\u9ad8\u4e14\u54cd\u5e94\u6162\u3002\u672c\u9879\u76ee\u7684\u76ee\u6807\u662f\u4f7f\u7528Hugging Face\u5fae\u8c03BERT\u6a21\u578b\uff0c\u6784\u5efa\u4e00\u4e2a\u667a\u80fd\u804a\u5929\u673a\u5668\u4eba\uff0c\u81ea\u52a8\u8bc6\u522b\u7528\u6237\u610f\u56fe\u5e76\u56de\u7b54\u5e38\u89c1\u95ee\u9898\uff0c\u4efb\u52a1\u7c7b\u578b\u4e3a\u6587\u672c\u5206\u7c7b\uff08\u610f\u56fe\u8bc6\u522b\uff09\uff0c\u65e8\u5728\u63d0\u5347\u54cd\u5e94\u901f\u5ea6\u3001\u964d\u4f4e\u8fd0\u8425\u6210\u672c\uff0c\u5e76\u5b9e\u73b024\/7\u5728\u7ebf\u670d\u52a1\u3002<\/p>\n

2. \u73af\u5883\u51c6\u5907<\/h2>\n

\u672c\u9879\u76ee\u4f7f\u7528uv\uff08\u4e00\u4e2a\u5feb\u901f\u7684Python\u5305\u7ba1\u7406\u5668\uff09\u6765\u7ba1\u7406\u73af\u5883\uff0c\u786e\u4fdd\u4f9d\u8d56\u9694\u79bb\u548c\u5feb\u901f\u5b89\u88c5\u3002\u9996\u5148\uff0c\u786e\u4fdd\u7cfb\u7edf\u5df2\u5b89\u88c5uv\uff0c\u7136\u540e\u521b\u5efa\u9879\u76ee\u5e76\u6dfb\u52a0\u4f9d\u8d56\u3002<\/p>\n

\u793a\u4f8b\u547d\u4ee4\uff1a<\/p>\n

uv init customer-service-bot\ncd customer-service-bot\nuv add transformers datasets torch fastapi uvicorn scikit-learn<\/code><\/pre>\n
\u4f9d\u8d56\u5305\u62ec\uff1atransformers\uff08\u7528\u4e8e\u52a0\u8f7d\u548c\u5fae\u8c03BERT\u6a21\u578b\uff09\u3001datasets\uff08\u6570\u636e\u7ba1\u7406\uff09\u3001torch\uff08\u6df1\u5ea6\u5b66\u4e60\u6846\u67b6\uff09\u3001fastapi\u548cuvicorn\uff08\u6784\u5efa\u548c\u8fd0\u884cAPI\uff09\u3001scikit-learn\uff08\u6307\u6807\u8ba1\u7b97\uff09\u3002<\/p>\n
3. \u6570\u636e\u8bf4\u660e<\/h2>\n
\u7531\u4e8e\u771f\u5b9e\u5ba2\u670d\u6570\u636e\u53ef\u80fd\u6d89\u53ca\u9690\u79c1\uff0c\u6211\u4eec\u6a21\u62df\u4e00\u4e2aQA\u6570\u636e\u96c6\u3002\u6bcf\u4e2a\u6837\u672c\u5305\u62ec\u7528\u6237\u95ee\u9898\u6587\u672c\u548c\u5bf9\u5e94\u7684\u610f\u56fe\u6807\u7b7e\uff08\u4f8b\u59820\u4ee3\u8868\u8ba2\u5355\u72b6\u6001\u67e5\u8be2\uff0c1\u4ee3\u8868\u9000\u8d27\u653f\u7b56\uff09\u3002\u751f\u6210\u903b\u8f91\uff1a\u57fa\u4e8e\u6a21\u677f\u521b\u5efa1000\u4e2a\u6837\u672c\uff0c\u5982”\u5982\u4f55\u67e5\u8be2\u8ba2\u5355\uff1f”\u6807\u7b7e\u4e3a0\uff0c”\u9000\u8d27\u6d41\u7a0b\u662f\u4ec0\u4e48\uff1f”\u6807\u7b7e\u4e3a1\u3002\u5728\u5b9e\u9645\u4e1a\u52a1\u4e2d\uff0c\u5e94\u4f7f\u7528\u5386\u53f2\u5ba2\u670d\u5bf9\u8bdd\u6570\u636e\uff0c\u5e76\u786e\u4fdd\u6570\u636e\u53e3\u5f84\u4e00\u81f4\u3002<\/p>\n
4. \u8bad\u7ec3\/\u5b9e\u73b0\u6b65\u9aa4<\/h2>\n
\u5b8c\u6574\u4ee3\u7801\uff0c\u6db5\u76d6\u6570\u636e\u52a0\u8f7d\u3001BERT\u6a21\u578b\u5fae\u8c03\u3001\u8bc4\u4f30\u548c\u4fdd\u5b58\uff0c\u53ef\u76f4\u63a5\u590d\u5236\u8fd0\u884c\u3002<\/p>\n
import torch\nfrom transformers import BertTokenizer, BertForSequenceClassification, Trainer, TrainingArguments\nfrom datasets import Dataset\nfrom sklearn.model_selection import train_test_split\nimport numpy as np\n\n# \u6a21\u62df\u6570\u636e\u751f\u6210\nquestions = ["\u5982\u4f55\u67e5\u8be2\u8ba2\u5355\u72b6\u6001\uff1f", "\u9000\u8d27\u653f\u7b56\u8be6\u60c5", "\u8ba2\u5355\u53d6\u6d88\u6d41\u7a0b", "\u8fd0\u8d39\u8ba1\u7b97\u65b9\u5f0f"] * 250  # 1000\u4e2a\u6837\u672c\nlabels = [0, 1, 2, 3] * 250  # \u610f\u56fe\u6807\u7b7e\uff1a0-\u8ba2\u5355\u67e5\u8be2,1-\u9000\u8d27\u653f\u7b56,2-\u8ba2\u5355\u53d6\u6d88,3-\u5176\u4ed6\nlabels = np.array(labels)\n\n# \u5206\u5272\u6570\u636e\u96c6\nX_train, X_test, y_train, y_test = train_test_split(questions, labels, test_size=0.2, random_state=42)\ntrain_dataset = Dataset.from_dict({"text": X_train, "label": y_train})\ntest_dataset = Dataset.from_dict({"text": X_test, "label": y_test})\n\n# \u52a0\u8f7dBERT\u6a21\u578b\u548ctokenizer\nmodel_name = "bert-base-uncased"\ntokenizer = BertTokenizer.from_pretrained(model_name)\nmodel = BertForSequenceClassification.from_pretrained(model_name, num_labels=4)  # 4\u4e2a\u610f\u56fe\u7c7b\u522b\n\n# Tokenize\u6570\u636e\ndef tokenize_function(examples):\n    return tokenizer(examples["text"], padding="max_length", truncation=True, max_length=128)\n\ntrain_dataset = train_dataset.map(tokenize_function, batched=True)\ntest_dataset = test_dataset.map(tokenize_function, batched=True)\n\n# \u8bbe\u7f6e\u8bad\u7ec3\u53c2\u6570\ntraining_args = TrainingArguments(\n    output_dir=".\/results",\n    num_train_epochs=3,\n    per_device_train_batch_size=8,\n    per_device_eval_batch_size=8,\n    evaluation_strategy="epoch",\n    logging_dir=".\/logs",\n    save_strategy="epoch",\n    load_best_model_at_end=True,\n)\n\n# \u521d\u59cb\u5316Trainer\u5e76\u8bad\u7ec3\ntrainer = Trainer(\n    model=model,\n    args=training_args,\n    train_dataset=train_dataset,\n    eval_dataset=test_dataset,\n)\ntrainer.train()\n\n# \u4fdd\u5b58\u5fae\u8c03\u6a21\u578b\nmodel.save_pretrained(".\/fine-tuned-bert-model")\ntokenizer.save_pretrained(".\/fine-tuned-bert-model")\n\n# \u8bc4\u4f30\u6a21\u578b\npredictions = trainer.predict(test_dataset)\npreds = np.argmax(predictions.predictions, axis=-1)\nacc = np.mean(preds == y_test)\nfrom sklearn.metrics import f1_score\nf1 = f1_score(y_test, preds, average="weighted")\nprint(f"\u51c6\u786e\u7387: {acc:.4f}, F1\u5206\u6570: {f1:.4f}")<\/code><\/pre>\n
5. \u8c03\u7528\u65b9\u5f0f<\/h2>\n
\u63d0\u4f9b\u79bb\u7ebf\u6279\u91cf\u9884\u6d4b\u548c\u5355\u6761FastAPI\u8c03\u7528\u793a\u4f8b\u3002<\/p>\n
\u79bb\u7ebf\u6279\u91cf\u8c03\u7528<\/strong>\uff1a<\/p>\n
from transformers import BertTokenizer, BertForSequenceClassification\nimport torch\n\nmodel_path = ".\/fine-tuned-bert-model"\ntokenizer = BertTokenizer.from_pretrained(model_path)\nmodel = BertForSequenceClassification.from_pretrained(model_path)\n\ndef predict_batch(questions_list):\n    inputs = tokenizer(questions_list, padding=True, truncation=True, return_tensors="pt", max_length=128)\n    with torch.no_grad():\n        outputs = model(**inputs)\n        predictions = torch.argmax(outputs.logits, dim=-1)\n    return predictions.tolist()\n\n# \u793a\u4f8b\uff1a\u6279\u91cf\u9884\u6d4b\nbatch_questions = ["\u6211\u7684\u8ba2\u5355\u5230\u54ea\u91cc\u4e86\uff1f", "\u9000\u8d27\u9700\u8981\u4ec0\u4e48\u6761\u4ef6\uff1f"]\nresults = predict_batch(batch_questions)\nprint(f"\u9884\u6d4b\u610f\u56fe\u6807\u7b7e: {results}")  # \u8f93\u51fa\u5982 [0, 1]<\/code><\/pre>\n
\u5355\u6761\u793a\u4f8b\u901a\u8fc7FastAPI\u63a5\u53e3<\/strong>\uff1a<\/p>\n
from fastapi import FastAPI\nfrom pydantic import BaseModel\nfrom transformers import BertTokenizer, BertForSequenceClassification\nimport torch\n\napp = FastAPI()\nmodel_path = ".\/fine-tuned-bert-model"\ntokenizer = BertTokenizer.from_pretrained(model_path)\nmodel = BertForSequenceClassification.from_pretrained(model_path)\n\nclass UserQuery(BaseModel):\n    question: str\n\n@app.post("\/predict_intent")\nasync def predict_intent(query: UserQuery):\n    inputs = tokenizer(query.question, return_tensors="pt", padding=True, truncation=True, max_length=128)\n    with torch.no_grad():\n        outputs = model(**inputs)\n        intent_id = torch.argmax(outputs.logits, dim=-1).item()\n    intent_map = {0: "\u8ba2\u5355\u67e5\u8be2", 1: "\u9000\u8d27\u653f\u7b56", 2: "\u8ba2\u5355\u53d6\u6d88", 3: "\u5176\u4ed6"}\n    return {"question": query.question, "predicted_intent": intent_map[intent_id]}\n\n# \u8fd0\u884c\u547d\u4ee4\uff1auvicorn main:app --reload\uff0c\u7136\u540e\u8bbf\u95eehttp:\/\/localhost:8000\/docs\u6d4b\u8bd5<\/code><\/pre>\n
6. \u6307\u6807\u8bf4\u660e<\/h2>\n
    \n
  • \u51c6\u786e\u7387<\/strong>\uff1a\u6a21\u578b\u9884\u6d4b\u6b63\u786e\u7684\u6837\u672c\u6570\u5360\u603b\u6837\u672c\u6570\u7684\u767e\u5206\u6bd4\u3002\u4f8b\u5982\uff0c\u5728100\u4e2a\u95ee\u9898\u4e2d\uff0c\u6a21\u578b\u6b63\u786e\u5206\u7c7b90\u4e2a\uff0c\u51c6\u786e\u7387\u4e3a90%\u3002\u5b83\u7b80\u5355\u8861\u91cf\u6574\u4f53\u6b63\u786e\u6027\u3002<\/li>\n
  • F1\u5206\u6570<\/strong>\uff1a\u7ed3\u5408\u7cbe\u786e\u7387\u548c\u53ec\u56de\u7387\u7684\u7efc\u5408\u6307\u6807\u3002\u7cbe\u786e\u7387\u6307\u9884\u6d4b\u4e3a\u6b63\u7684\u6837\u672c\u4e2d\u5b9e\u9645\u4e3a\u6b63\u7684\u6bd4\u4f8b\uff1b\u53ec\u56de\u7387\u6307\u5b9e\u9645\u4e3a\u6b63\u7684\u6837\u672c\u4e2d\u88ab\u6b63\u786e\u9884\u6d4b\u7684\u6bd4\u4f8b\u3002F1\u5206\u6570\u662f\u4e24\u8005\u7684\u8c03\u548c\u5e73\u5747\uff0c\u8303\u56f40-1\uff0c\u8d8a\u9ad8\u8d8a\u597d\uff0c\u9002\u7528\u4e8e\u7c7b\u522b\u4e0d\u5e73\u8861\u7684\u6570\u636e\u3002<\/li>\n
  • \u635f\u5931<\/strong>\uff1a\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u6a21\u578b\u9884\u6d4b\u503c\u4e0e\u771f\u5b9e\u503c\u4e4b\u95f4\u7684\u8bef\u5dee\u5ea6\u91cf\uff0c\u5e38\u7528\u4ea4\u53c9\u71b5\u635f\u5931\uff0c\u503c\u8d8a\u5c0f\u8868\u793a\u6a21\u578b\u62df\u5408\u8d8a\u597d\u3002<\/li>\n
  • \u610f\u56fe\u8bc6\u522b<\/strong>\uff1a\u672c\u9879\u76ee\u4efb\u52a1\u7c7b\u578b\uff0c\u6307\u5c06\u7528\u6237\u6587\u672c\u95ee\u9898\u5206\u7c7b\u5230\u9884\u5b9a\u4e49\u7684\u610f\u56fe\u7c7b\u522b\uff08\u5982\u8ba2\u5355\u67e5\u8be2\u3001\u9000\u8d27\u653f\u7b56\uff09\u3002<\/li>\n
  • \u5fae\u8c03<\/strong>\uff1a\u5728\u9884\u8bad\u7ec3\u6a21\u578b\uff08\u5982BERT\uff09\u57fa\u7840\u4e0a\uff0c\u7528\u7279\u5b9a\u6570\u636e\u8bad\u7ec3\u4ee5\u9002\u914d\u65b0\u4efb\u52a1\uff0c\u63d0\u5347\u6027\u80fd\u3002<\/li>\n<\/ul>\n
    7. \u4e0a\u7ebf\u540e\u8bc4\u4f30<\/h2>\n
    \u90e8\u7f72\u540e\u9700\u76d1\u63a7\u8fd0\u884c\u6307\u6807\uff1a<\/p>\n
      \n
    • \u54cd\u5e94\u5ef6\u8fdf<\/strong>\uff1aAPI\u4ece\u63a5\u6536\u8bf7\u6c42\u5230\u8fd4\u56de\u7ed3\u679c\u7684\u65f6\u95f4\uff0c\u76ee\u6807\u5e94\u4f4e\u4e8e300\u6beb\u79d2\u4ee5\u907f\u514d\u7528\u6237\u4f53\u9a8c\u4e0b\u964d\u3002<\/li>\n
    • \u541e\u5410\u91cf<\/strong>\uff1a\u6bcf\u79d2\u80fd\u5904\u7406\u7684\u8bf7\u6c42\u6570\uff0c\u8bc4\u4f30\u7cfb\u7edf\u5e76\u53d1\u80fd\u529b\u3002<\/li>\n
    • \u51c6\u786e\u7387\u548cF1\u5206\u6570<\/strong>\uff1a\u5b9a\u671f\u7528\u65b0\u6570\u636e\u8bc4\u4f30\u6a21\u578b\u6027\u80fd\uff0c\u786e\u4fdd\u4e0d\u9000\u5316\u3002\n\u8bc4\u4f30\u65b9\u6cd5\uff1a\u901a\u8fc7\u65e5\u5fd7\u5206\u6790\u5ef6\u8fdf\u548c\u541e\u5410\u91cf\uff1b\u4f7f\u7528A\/B\u6d4b\u8bd5\u6bd4\u8f83\u673a\u5668\u4eba\u56de\u7b54\u4e0e\u4eba\u5de5\u56de\u7b54\u7684\u7528\u6237\u6ee1\u610f\u5ea6\uff1b\u8bbe\u7f6e\u9608\u503c\uff08\u5982\u51c6\u786e\u7387>85%\uff09\u89e6\u53d1\u91cd\u65b0\u8bad\u7ec3\u3002<\/li>\n<\/ul>\n
      8. \u5e38\u89c1\u5751\u4e0e\u6392\u67e5<\/h2>\n
        \n
      • \u8fc7\u62df\u5408<\/strong>\uff1a\u8bad\u7ec3\u6570\u636e\u4e0d\u8db3\u65f6\uff0c\u6a21\u578b\u5728\u8bad\u7ec3\u96c6\u8868\u73b0\u597d\u4f46\u6d4b\u8bd5\u96c6\u5dee\u3002\u6392\u67e5\uff1a\u68c0\u67e5\u8bad\u7ec3\u548c\u6d4b\u8bd5\u96c6\u5206\u5e03\uff1b\u589e\u52a0\u6570\u636e\u91cf\u6216\u4f7f\u7528\u6570\u636e\u589e\u5f3a\uff1b\u6dfb\u52a0Dropout\u6216\u6b63\u5219\u5316\u3002<\/li>\n
      • \u54cd\u5e94\u5ef6\u8fdf\u9ad8<\/strong>\uff1a\u5f71\u54cd\u7528\u6237\u4f53\u9a8c\u3002\u6392\u67e5\uff1a\u786e\u4fdd\u4f7f\u7528GPU\u63a8\u7406\uff1b\u4f18\u5316\u6a21\u578b\u5927\u5c0f\uff08\u5982\u4f7f\u7528\u84b8\u998f\u7248BERT\uff09\uff1b\u7f13\u5b58\u5e38\u89c1\u95ee\u9898\u56de\u7b54\u3002<\/li>\n
      • \u6570\u636e\u8d28\u91cf\u5dee<\/strong>\uff1a\u810f\u6570\u636e\u5bfc\u81f4\u6a21\u578b\u6027\u80fd\u4f4e\u3002\u6392\u67e5\uff1a\u6e05\u6d17\u6570\u636e\uff08\u53bb\u91cd\u3001\u7ea0\u6b63\u62fc\u5199\uff09\uff1b\u7edf\u4e00\u6807\u6ce8\u6807\u51c6\uff1b\u8fdb\u884c\u6570\u636e\u63a2\u7d22\u5206\u6790\u3002<\/li>\n
      • \u4f9d\u8d56\u7248\u672c\u51b2\u7a81<\/strong>\uff1auv\u5b89\u88c5\u5931\u8d25\u6216\u8fd0\u884c\u65f6\u9519\u8bef\u3002\u6392\u67e5\uff1a\u68c0\u67e5Python\u548c\u5305\u7248\u672c\u517c\u5bb9\u6027\uff1b\u4f7f\u7528\u865a\u62df\u73af\u5883\u9694\u79bb\u3002<\/li>\n<\/ul>","protected":false},"excerpt":{"rendered":"
        \u672c\u6559\u7a0b\u8be6\u7ec6\u4ecb\u7ecd\u5982\u4f55\u4f7f\u7528Hugging Face\u5e93\u5fae\u8c03BERT\u6a21\u578b\uff0c\u6784\u5efa\u7535\u5546\u5ba2\u670d\u667a\u80fd\u804a\u5929\u673a\u5668\u4eba\uff0c\u5305\u62ec\u73af\u5883\u51c6\u5907\u3001\u6570\u636e\u6a21\u62df\u3001\u5b8c\u6574\u4ee3\u7801\u3001\u6307\u6807\u89e3\u91ca\u548c\u90e8\u7f72\u793a\u4f8b\uff0c\u5e2e\u52a9\u5f00\u53d1\u8005\u5feb\u901f\u5b9e\u73b0\u81ea\u52a8\u5316\u5ba2\u670d\u89e3\u51b3\u65b9\u6848\u3002<\/p>","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[94,98],"tags":[137],"class_list":["post-437","post","type-post","status-publish","format-standard","hentry","category-ai","tag-137"],"views":154,"_links":{"self":[{"href":"https:\/\/www.liaoxinghui.com\/index.php?rest_route=\/wp\/v2\/posts\/437","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.liaoxinghui.com\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.liaoxinghui.com\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.liaoxinghui.com\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.liaoxinghui.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=437"}],"version-history":[{"count":1,"href":"https:\/\/www.liaoxinghui.com\/index.php?rest_route=\/wp\/v2\/posts\/437\/revisions"}],"predecessor-version":[{"id":446,"href":"https:\/\/www.liaoxinghui.com\/index.php?rest_route=\/wp\/v2\/posts\/437\/revisions\/446"}],"wp:attachment":[{"href":"https:\/\/www.liaoxinghui.com\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=437"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.liaoxinghui.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=437"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.liaoxinghui.com\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=437"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}