metadata
library_name: transformers
tags:
- ner
- biomedical
- disease-recognition
- pubmedbert
- BioMedNLP
datasets:
- rjac/biobert-ner-diseases-dataset
license: mit
language:
- en
metrics:
- precision
- recall
- f1
base_model:
- microsoft/BiomedNLP-BiomedBERT-base-uncased-abstract-fulltext
Model Card for BioMed-NER-English
Fine-tuned BiomedNLP-BiomedBERT model for medical entity recognition, achieving 0.9868 F1-score on disease entity extraction from clinical text.
Model Details
Model Description
- Developed by: Aashish Acharya
- Model type: BiomedNLP-BiomedBERT (Token Classification)
- Language(s): English
- License: MIT
- Finetuned from model: microsoft/BiomedNLP-PubMedBERT-base-uncased-abstract-fulltext
- Source Code: GitHub Link
Model Sources
- Base Model: microsoft/BiomedNLP-PubMedBERT-base-uncased-abstract-fulltext
- Training Dataset: rjac/biobert-ner-diseases-dataset
Uses
Direct Use
This model excels at extracting disease mentions from medical text using BIO tagging scheme:
- B-Disease: Beginning of disease mention
- I-Disease: Continuation of disease mention
- O: Non-disease tokens
Training
Training Data
Dataset: biobert-ner-diseases-dataset Size: 21,225 annotated medical sentences Split: 15,488 training (73%), 5,737 testing (27%) Average sentence length: 24.3 tokens Disease mention frequency: 1.8 per sentence
Training Procedure Training Hyperparameters
- Learning rate: 5e-5
- Batch size: 8
- Epochs: 8
- Optimizer: AdamW with weight decay (0.01)
- Warmup steps: 500
- Early stopping patience: 5
- Loss function: Cross-entropy with label smoothing (0.1)
- Gradient accumulation steps: 4
- Max gradient norm: 1.0
Evaluation
Metrics
Final model performancc
Strict Entity Matching:
Precision: 0.9869
Recall: 0.9868
F1 Score: 0.9868
Partial Entity Matching:
Precision: 0.9527
Recall: 0.9456
F1 Score: 0.9491
Error Analysis
Boundary Errors: 1,154
Type Errors: 0
Environmental Impact
Hardware Type: Google Colab GPU
Hours used: ~2 hours
Cloud Provider: Google Cloud
Carbon Emitted: Not tracked
Technical Specifications Model Architecture
Base model: PubMedBERT
Hidden size: 768
Attention heads: 12
Layers: 12
Parameters: ~110M
Compute Infrastructure
Platform: Google Colab
GPU: Tesla T4/P100
Citation
@misc{acharya2024sapbert,
title={SapBERT-PubMedBERT Fine-tuned on DDXPlus Dataset},
author={Acharya, Aashish},
year={2024},
publisher={Hugging Face Model Hub}
}