macbert训练过程记录

先把demo直接放到服务器上，一个scp命令搞定：

scp -r wangfei@10.30.71.37:/Users/wangfei/Documents/Codex/2026-06-26/jin-t/outputs/macbert_intent_demo /home/ubuntu/zlink/wangf36_temp

文件目录结构就是常规的工程项目，该有的模块文件都在里面，不多说。

接下来切换一下conda环境，别搞混了：

conda activate llama_factory

然后装一下依赖，跑起来之前先把这个搞定：

pip install -r requirements.txt

模型下载，两种方式可选

基座模型怎么弄？有两种方式供选择：

• 提前下载好放服务器上，训练时直接指定本地路径
• 从Hugging Face直接拉取

推荐的做法是提前下载本地，省得网络出幺蛾子。如果服务器上已经有本地MacBERT模型，比如放在：

/home/ubuntu/zlink/base_model/chinese-macbert-base

那就直接这样训练：

python train_intent_macbert.py --model_name /home/ubuntu/zlink/base_model/chinese-macbert-base --data_dir . --output_dir ./macbert-intent-model --epochs 5 --batch_size 16 --learning_rate 2e-5

如果服务器能顺畅访问Hugging Face，也可以直接指定模型名称：

python train_intent_macbert.py --model_name hfl/chinese-macbert-base --data_dir . --output_dir ./macbert-intent-model --epochs 5 --batch_size 16 --learning_rate 2e-5

我采用的是先下载到本地再上传到服务器的方式，稳当。

训练报错，直击问题

跑起来之后，控制台里直接报了个PermissionError，错误信息指向/tmp/ssl_keys.log。怎么回事呢？

(llama_factory) ubuntu@VM-12-16-ubuntu:~/zlink/wangf36_temp/macbert_intent_demo$ python train_intent_macbert.py --model_name /home/ubuntu/zlink/base_model/chinese-macbert-base --data_dir . --output_dir ./macbert-intent-model --epochs 5 --batch_size 16 --learning_rate 2e-5
Traceback (most recent call last):
  File "/home/ubuntu/zlink/wangf36_temp/macbert_intent_demo/train_intent_macbert.py", line 6, in 
    from datasets import load_dataset
  ...
    PermissionError: [Errno 13] Permission denied: '/tmp/ssl_keys.log'

其实是因为SSLKEYLOGFILE这个环境变量作祟，导致Python在创建SSL上下文时试图往/tmp/ssl_keys.log写文件，但权限不够。

解决办法，直接取消

解决方法很简单，把这个环境变量unset掉就行：

(llama_factory) ubuntu@VM-12-16-ubuntu:~/zlink/wangf36_temp/macbert_intent_demo$ unset SSLKEYLOGFILE
python train_intent_macbert.py --model_name /home/ubuntu/zlink/base_model/chinese-macbert-base --data_dir . --output_dir ./macbert-intent-model --epochs 5 --batch_size 16 --learning_rate 2e-5
Generating train split: 120 examples [00:00, 76445.39 examples/s]
Generating validation split: 30 examples [00:00, 41432.04 examples/s]
Generating test split: 18 examples [00:00, 31868.92 examples/s]
Map: 100%|...| 120/120 [00:00<00:00, 16187.97 examples/s]
Map: 100%|...| 30/30 [00:00<00:00, 10606.85 examples/s]
Map: 100%|...| 18/18 [00:00<00:00, 7173.84 examples/s]
Some weights of BertForSequenceClassification were not initialized from the model checkpoint at /home/ubuntu/zlink/base_model/chinese-macbert-base and are newly initialized: ['classifier.bias', 'classifier.weight']
You should probably TRAIN this model on a down-stream task to be able to use it for predictions and inference.
/home/ubuntu/zlink/wangf36_temp/macbert_intent_demo/train_intent_macbert.py:121: FutureWarning: `tokenizer` is deprecated and will be removed in version 5.0.0 for `Trainer.__init__`. Use `processing_class` instead.
  trainer = Trainer(
0%| 0/10 [00:00

(llama_factory) ubuntu@VM-12-16-ubuntu:~/zlink/wangf36_temp/macbert_intent_demo$ python predict_intent.py --model_dir ./macbert-intent-model/best --text "明天上海会下雨吗"
{"text": "明天上海会下雨吗","intent": "unknown","raw_intent": "query_weather","confidence": 0.4739,"threshold": 0.6,"scores": {"query_weather": 0.4739,"book_ticket": 0.1637,"open_app": 0.087,"search_knowledge": 0.0848,"chat": 0.1355,"unknown": 0.0551}}