jetaudio/bert-base-zh-ner

chinese_ner

It achieves the following results on the evaluation set:

• Overall Precision: 84.86%
• Overall Recall: 86.23%
• Overall F1: 85.54%
• Overall Accuracy: 99.14%

Model description

bert-base-chinese finetuned on augmented MSRA dataset (x4) using real-life PER, ORG, LOC, FPER name with context length of 128. Trained in bfloat16. It's has 8 labels:

• "0": "O": Not an entity
• "1": "B-PER": Beginning of a Chinese person's name
• "2": "I-PER": Inside of a Chinese person's name
• "3": "B-ORG": Beginning of an organization's name
• "4": "I-ORG": Inside of an organization's name
• "5": "B-LOC": Beginning of an location's name
• "6": "I-LOC": Inside of an location's name
• "7": "B-FPER": Beginning of a Non-Chinese person's name
• "8": "I-FPER": Inside of a Non-Chinese person's name

Framework versions

• Transformers 4.49.0
• Pytorch 2.6.0+cu126
• Datasets 3.4.0
• Tokenizers 0.21.1