Aster Vietnamese NER MiniLM
A 15-label Vietnamese named-entity recognition checkpoint fine-tuned from microsoft/Multilingual-MiniLM-L12-H384.
Labels
ADDRESS, DATETIME, EMAIL, EVENT, IP, LOCATION, MISCELLANEOUS, ORGANIZATION, PERSON, PERSONTYPE, PHONENUMBER, PRODUCT, QUANTITY, SKILL, and URL.
Held-out evaluation
Exact typed entity-span F1:
| Dataset | Split | F1 |
|---|---|---|
| VLSP 2021 | test_update_10t01 |
81.89% |
| WikiANN Vietnamese | test | 86.33% |
| Vietnamese COVID NER | validation | 92.06% |
The WikiANN and COVID figure combined is 90.21% F1. These splits were not used during fine-tuning.
Performance
The model is 470 MB in safetensors format. On an 8-thread AMD Ryzen 7 8840HS CPU, warm inference across 100 VLSP test sentences measured a 19.1 ms median and 33.9 ms p95 per sentence. An NVIDIA L40S measured 5.4-6.1 ms median per sentence.
Training
The model was trained for 10 epochs on the train splits of VLSP 2021, WikiANN Vietnamese, and Vietnamese COVID NER. The head uses the combined PERSON, ORGANIZATION, and LOCATION labels across corpora while retaining VLSP's additional entity types.
Usage
from transformers import AutoModelForTokenClassification, AutoTokenizer
model_id = "Keithsel/Aster-Vietnamese-NER-MiniLM"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForTokenClassification.from_pretrained(model_id).eval()
This is a research checkpoint. Review each training dataset's terms before redistributing or using the weights in a production product.
- Downloads last month
- 11