ModernBERT for Named Entity Recognition (NER)

A fine-tuned ModernBERT-base model for Named Entity Recognition, trained on the CoNLL-2003 dataset to identify persons, organizations, locations, and miscellaneous entities in English text.

Model Overview

This model achieves strong performance on NER tasks:

• F1 Score: 84.55%
• Precision: 83.49%
• Recall: 85.63%
• Accuracy: 97.52%

Quick Start

from transformers import pipeline

ner     = pipeline(task="token-classification", model="IsmaelMousa/modernbert-ner-conll2003", aggregation_strategy="max")

text    = "Hi, I'm Ismael Mousa from Palestine working for NVIDIA inc."

results = ner(text)

for entity in results: print(f"{entity['word']} => {entity['entity_group']}")

Output:

Ismael Mousa => PER
Palestine    => LOC
NVIDIA       => ORG

Entity Types

The model recognizes four entity categories:

• PER: Person names
• ORG: Organizations
• LOC: Locations
• MISC: Miscellaneous entities

Training Details

• Base Model: ModernBERT-base (149.6M parameters)
• Dataset: CoNLL-2003
• Training Examples: 14,041
• Epochs: 10
• Learning Rate: 1e-6

Performance

Evaluated on CoNLL-2003 validation set (3,250 examples):

Metric	Score
F1	0.8455
Precision	0.8349
Recall	0.8563
Accuracy	0.9752

Implementation

The training process involves:

1. Tokenizing text with word-level alignment
2. Mapping labels to handle subword tokenization
3. Training with cross-entropy loss on token classification
4. Evaluating with seqeval metrics

See the training notebook for complete implementation details.

References

• ModernBERT Paper: Smarter, Better, Faster, Longer: A Modern Bidirectional Encoder for Fast, Memory Efficient, and Long Context Finetuning and Inference
• Base Model: answerdotai/ModernBERT-base
• Dataset: CoNLL-2003

License

Apache 2.0