cbc-528a/BamiBERT-ViMedNER

Base model

This model is fine-tuned from Qualcomm-AI-Research/BamiBERT for Vietnamese biomedical Named Entity Recognition (NER).

Dataset

This model was fine-tuned on the ViMedNER dataset.

Repository: https://github.com/tdtrinh11/ViMedNer

Training Hyperparameters

The model was fine-tuned from Qualcomm-AI-Research/BamiBERT for Vietnamese biomedical Named Entity Recognition (NER) using the following training configuration.

Hyperparameter	Value
Max sequence length	128
Batch size	32
Learning rate	5e-5
Number of epochs	30
Warmup ratio	0.1
Weight decay	0.01
Gradient accumulation steps	1
Optimizer	AdamW
Learning rate scheduler	Linear with warmup
Task	Token Classification (NER)
Framework	Hugging Face Transformers

Evaluation

The model was evaluated on the ViMedNER test set using the seqeval evaluation metrics for named entity recognition.

Overall Performance

Metric	Score
Precision (Micro)	0.6605
Recall (Micro)	0.7133
F1 Score (Micro)	0.6859

Per-Entity Performance

Entity Type	Precision	Recall	F1 Score	Support
ten_benh	0.77	0.84	0.81	1805
trieu_chung_benh	0.60	0.63	0.62	703
bien_phap_dieu_tri	0.57	0.61	0.59	631
bien_phap_chan_doan	0.64	0.67	0.66	303
nguyen_nhan_benh	0.31	0.36	0.34	276

Aggregate Metrics

Average	Precision	Recall	F1 Score
Micro	0.66	0.71	0.69
Macro	0.58	0.62	0.60
Weighted	0.66	0.71	0.69

Note: Evaluation was performed using the seqeval library on the official ViMedNER test split. The model achieves an overall micro F1-score of 68.6% on the ViMedNER test set. Performance is strongest on the ten_benh (Disease) entity type with an F1-score of 0.81, while nguyen_nhan_benh (Cause) remains the most challenging category due to its relatively small number of training examples and semantic diversity.

Quick Start

Install dependencies

pip install transformers torch

Load the Model

from transformers import AutoTokenizer
from transformers import AutoModelForTokenClassification

MODEL_NAME = "cbc-528a/BamiBERT-ViMedNER"

tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)
model = AutoModelForTokenClassification.from_pretrained(MODEL_NAME)

Example Inference

import torch
from transformers import AutoTokenizer
from transformers import AutoModelForTokenClassification

MODEL_NAME = "your-org/BamiBERT-ViMedNER"

tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)
model = AutoModelForTokenClassification.from_pretrained(MODEL_NAME)

sentence = "Bệnh nhân bị ung thư gan và được điều trị bằng Sorafenib."

words = sentence.split()

inputs = tokenizer(
    words,
    is_split_into_words=True,
    return_tensors="pt"
)

with torch.no_grad():
    outputs = model(**inputs)

predictions = outputs.logits.argmax(dim=-1)[0]

word_ids = inputs.word_ids()

id2label = model.config.id2label

previous_word = None

for pred, word_idx in zip(predictions, word_ids):

    if word_idx is None or word_idx == previous_word:
        continue

    previous_word = word_idx

    print(
        words[word_idx],
        id2label[pred.item()]
    )

Sample Output

Bệnh        O
nhân        O
bị          O
ung         B-ten_benh
thư         I-ten_benh
gan         I-ten_benh
và          O
được        O
điều        O
trị         O
bằng        O
Sorafenib   B-bien_phap_dieu_tri
.           O

License

This model is a fine-tuned derivative of Qualcomm-AI-Research/BamiBERT.

Accordingly, use of this model is subject to:

• BSD 3-Clause Clear License
• Qualcomm Responsible AI License

Please refer to the original BamiBERT repository for the complete license terms.

Acknowledgements & Citation

This model is fine-tuned based on the BamiBERT architecture. I would like to express my gratitude to the original authors for their foundational work.

BibTeX:

@article{BamiBERT,
  title    = {{BamiBERT: A New BERT-based Language Model for Vietnamese}},
  author   = {Dat Quoc Nguyen and Thinh Pham and Chi Tran and Linh The Nguyen},
  journal  = {arXiv preprint},
  volume   = {arXiv:2607.02259},
  year     = {2026}
}