OpenMed-NER-PharmaDetect-SuperClinical-434M (ONNX)

This is an ONNX version of OpenMed/OpenMed-NER-PharmaDetect-SuperClinical-434M. It was automatically converted and uploaded using this Hugging Face Space.

Usage with Transformers.js

See the pipeline documentation for token-classification: https://huggingface.co/docs/transformers.js/api/pipelines#module_pipelines.TokenClassificationPipeline

---

🧬 OpenMed-NER-PharmaDetect-SuperClinical-434M

Specialized model for Chemical Entity Recognition - Chemical entities from the BC5CDR dataset

📋 Model Overview

This model is a state-of-the-art fine-tuned transformer engineered to deliver enterprise-grade accuracy for chemical entity recognition - chemical entities from the bc5cdr dataset. This specialized model excels at identifying and extracting biomedical entities from clinical texts, research papers, and healthcare documents, enabling applications such as drug interaction detection, medication extraction from patient records, adverse event monitoring, literature mining for drug discovery, and biomedical knowledge graph construction with production-ready reliability for clinical and research applications.

🎯 Key Features

• High Precision: Optimized for biomedical entity recognition
• Domain-Specific: Trained on curated BC5CDR_CHEM dataset
• Production-Ready: Validated on clinical benchmarks
• Easy Integration: Compatible with Hugging Face Transformers ecosystem

🏷️ Supported Entity Types

This model can identify and classify the following biomedical entities:

• B-CHEM
• I-CHEM

📊 Dataset

BC5CDR-Chem focuses on chemical entity recognition from the BioCreative V Chemical-Disease Relation extraction task.

The BC5CDR-Chem corpus is part of the BioCreative V Chemical-Disease Relation (CDR) extraction challenge, specifically targeting chemical entity recognition in biomedical texts. This dataset contains 1,500 PubMed abstracts with 4,409 annotated chemical entities, designed to support automated drug discovery and pharmacovigilance applications. The corpus emphasizes chemical compounds, drugs, and therapeutic substances that are relevant for understanding chemical-disease relationships. It serves as a critical resource for developing NER systems that can identify chemical entities for downstream tasks like adverse drug reaction detection and drug repurposing research.

📊 Performance Metrics

Current Model Performance

• F1 Score: 0.96
• Precision: 0.95
• Recall: 0.97
• Accuracy: 0.99

🏆 Comparative Performance on BC5CDR_CHEM Dataset

Rankings based on F1-score performance across all models trained on this dataset.

Figure: OpenMed (Open-Source) vs. Latest SOTA (Closed-Source) performance comparison across biomedical NER datasets.

🚀 Quick Start

Installation

pip install transformers torch

Usage

from transformers import pipeline

# Load the model and tokenizer
# Model: https://huggingface.co/OpenMed/OpenMed-NER-PharmaDetect-SuperClinical-434M
model_name = "OpenMed/OpenMed-NER-PharmaDetect-SuperClinical-434M"

# Create a pipeline
medical_ner_pipeline = pipeline(
    model=model_name,
    aggregation_strategy="simple"
)

# Example usage
text = "Administration of metformin reduced glucose levels significantly."
entities = medical_ner_pipeline(text)

print(entities)

token = entities[0]
print(text[token["start"] : token["end"]])

NOTE: The aggregation_strategy parameter defines how token predictions are grouped into entities. For a detailed explanation, please refer to the Hugging Face documentation.

Here is a summary of the available strategies:

• none: Returns raw token predictions without any aggregation.
• simple: Groups adjacent tokens with the same entity type (e.g., B-LOC followed by I-LOC).
• first: For word-based models, if tokens within a word have different entity tags, the tag of the first token is assigned to the entire word.
• average: For word-based models, this strategy averages the scores of tokens within a word and applies the label with the highest resulting score.
• max: For word-based models, the entity label from the token with the highest score within a word is assigned to the entire word.

Batch Processing

For efficient processing of large datasets, use proper batching with the batch_size parameter:

texts = [
    "Administration of metformin reduced glucose levels significantly.",
    "The study evaluated the efficacy of cisplatin in cancer treatment.",
    "Patients received ibuprofen for inflammation management.",
    "The patient's medication was switched to tamoxifen to prevent breast cancer recurrence.",
    "Lithium carbonate is often prescribed for the management of bipolar disorder.",
]

# Efficient batch processing with optimized batch size
# Adjust batch_size based on your GPU memory (typically 8, 16, 32, or 64)
results = medical_ner_pipeline(texts, batch_size=8)

for i, entities in enumerate(results):
    print(f"Text {i+1} entities:")
    for entity in entities:
        print(f"  - {entity['word']} ({entity['entity_group']}): {entity['score']:.4f}")

Large Dataset Processing

For processing large datasets efficiently:

from transformers.pipelines.pt_utils import KeyDataset
from datasets import Dataset
import pandas as pd

# Load your data
# Load a medical dataset from Hugging Face
from datasets import load_dataset

# Load a public medical dataset (using a subset for testing)
medical_dataset = load_dataset("BI55/MedText", split="train[:100]")  # Load first 100 examples
data = pd.DataFrame({"text": medical_dataset["Completion"]})
dataset = Dataset.from_pandas(data)

# Process with optimal batching for your hardware
batch_size = 16  # Tune this based on your GPU memory
results = []

for out in medical_ner_pipeline(KeyDataset(dataset, "text"), batch_size=batch_size):
    results.extend(out)

print(f"Processed {len(results)} texts with batching")

Performance Optimization

Batch Size Guidelines:

• CPU: Start with batch_size=1-4
• Single GPU: Try batch_size=8-32 depending on GPU memory
• High-end GPU: Can handle batch_size=64 or higher
• Monitor GPU utilization to find the optimal batch size for your hardware

Memory Considerations:

# For limited GPU memory, use smaller batches
medical_ner_pipeline = pipeline(
    model=model_name,
    aggregation_strategy="simple",
    device=0  # Specify GPU device
)

# Process with memory-efficient batching
for batch_start in range(0, len(texts), batch_size):
    batch = texts[batch_start:batch_start + batch_size]
    batch_results = medical_ner_pipeline(batch, batch_size=len(batch))
    results.extend(batch_results)

📚 Dataset Information

• Dataset: BC5CDR_CHEM
• Description: Chemical Entity Recognition - Chemical entities from the BC5CDR dataset

Training Details

• Base Model: deberta-v3-large
• Training Framework: Hugging Face Transformers
• Optimization: AdamW optimizer with learning rate scheduling
• Validation: Cross-validation on held-out test set

🔬 Model Architecture

• Base Architecture: deberta-v3-large
• Task: Token Classification (Named Entity Recognition)
• Labels: Dataset-specific entity types
• Input: Tokenized biomedical text
• Output: BIO-tagged entity predictions

💡 Use Cases

This model is particularly useful for:

• Clinical Text Mining: Extracting entities from medical records
• Biomedical Research: Processing scientific literature
• Drug Discovery: Identifying chemical compounds and drugs
• Healthcare Analytics: Analyzing patient data and outcomes
• Academic Research: Supporting biomedical NLP research

📜 License

Licensed under the Apache License 2.0. See LICENSE for details.

🤝 Contributing

We welcome contributions of all kinds! Whether you have ideas, feature requests, or want to join our mission to advance open-source Healthcare AI, we'd love to hear from you.

Follow OpenMed Org on Hugging Face 🤗 and click "Watch" to stay updated on our latest releases and developments.

Citation

If you use this model in your research or applications, please cite the following paper:

@misc{panahi2025openmedneropensourcedomainadapted,
      title={OpenMed NER: Open-Source, Domain-Adapted State-of-the-Art Transformers for Biomedical NER Across 12 Public Datasets},
      author={Maziyar Panahi},
      year={2025},
      eprint={2508.01630},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2508.01630},
}

Proper citation helps support and acknowledge my work. Thank you!