bert-finetuned-ner

bert-finetuned-ner

This model is a fine-tuned Persian BERT model based on HooshvareLab/bert-base-parsbert-uncased for Named Entity Recognition (NER). It has been trained to identify entities such as persons, organizations, locations, and products in Persian text.

Model Description

bert-finetuned-ner is designed for token-level classification in Persian. The model uses ParsBERT, a BERT variant pretrained on a large Persian corpus, as the base model and is fine-tuned on a wnut2017-persian dataset. It can predict entity labels for each token in input text, supporting tasks such as text analysis, information extraction, and question answering pipelines.

Intended Uses & Limitations

Intended Uses

• Named Entity Recognition (NER) in Persian text.
• Information extraction for NLP pipelines in Persian language applications.
• Academic research or industrial projects requiring entity tagging.

Limitations

• Performance depends heavily on the coverage and quality of the training data. Entities not represented in the dataset may not be recognized.
• The model may misclassify rare or out-of-vocabulary words.
• For critical applications, manual verification of predictions is recommended.
• Trained on formal text; performance on dialects or colloquial Persian may vary.

Training and Evaluation Data

• Dataset: ('Amir13/wnut2017-persian')
• Entities annotated include: persons, organizations, locations, creative works, and products.
• Tokenization handled using ParsBERT tokenizer (HooshvareLab/bert-base-parsbert-uncased).

Training Procedure

Training Hyperparameters

• Learning rate: 2e-5
• Train batch size: 8
• Evaluation batch size: 8
• Optimizer: AdamW (betas=(0.9, 0.999), epsilon=1e-8)
• Learning rate scheduler: linear
• Number of epochs: 10
• Seed: 42

Training Environment

• Framework: Transformers 4.56.1
• PyTorch 2.8.0+cu126
• Datasets 4.0.0
• Tokenizers 0.22.0

Training Results

• The model achieved training loss of ~0.13 (averaged over all epochs).
• Accuracy, F1, precision, and recall metrics are recommended to be computed on a held-out test set for full evaluation.

How to Use

from transformers import AutoTokenizer, AutoModelForTokenClassification, pipeline

model_path = "path_to_saved_model"

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

ner_pipeline = pipeline(
    "ner",
    model=model,
    tokenizer=tokenizer,
    aggregation_strategy="simple"
)

text = "سلام. در تهران زندگی میکنم."
results = ner_pipeline(text)
print(results)