Shaza2004/MacBERT-Chinese-de-id

MacBERT Chinese de-identification with GlobalPointer

Encoder: hfl/chinese-macbert-base. The repository contains the best dev-F1 checkpoint, tokenizer, label mapping, and training metadata.

The checkpoint is a PyTorch state dict for the custom GlobalPointer class defined in the training notebook. To load it, define that class and run model.load_state_dict(torch.load('model.pt', map_location='cpu', weights_only=True)).

Best dev metrics: { "epoch": 12, "train_loss": 0.023357975111436096, "dev": { "loss": 0.032338257023958705, "precision": 0.9403027706369609, "recall": 0.9841554559043348, "f1": 0.9617294770669004 } }