DistilBERT fine-tuned on CoNLL-2003 for NER

distilbert-base-cased fine-tuned for named entity recognition on CoNLL-2003. It tags PER (person), ORG (organisation), LOC (location) and MISC (miscellaneous) entities in English text.

This is the model used by the NER System project, which adds DATE detection with spaCy and a Streamlit app.

Usage

from transformers import pipeline

ner = pipeline(
    "token-classification",
    model="aryanmaurya006/ner-distilbert-conll2003",
    aggregation_strategy="simple",
)
ner("Sundar Pichai met Narendra Modi in Delhi.")

Results (CoNLL-2003 test set, 3,453 sentences)

Entity Precision Recall F1 Support
PER 0.9610 0.9450 0.9529 1,617
ORG 0.8487 0.8814 0.8647 1,661
LOC 0.9180 0.9197 0.9188 1,668
MISC 0.7394 0.7963 0.7668 702
Overall 0.8856 0.9003 0.8929 5,648

Training

  • Base model: distilbert-base-cased
  • Data: CoNLL-2003 (tomaarsen/conll2003), 14,041 training sentences
  • 3 epochs, learning rate 2e-5, batch size 16, weight decay 0.01, seed 42
  • Labels are aligned to the first sub-word of each word; other sub-words are ignored in the loss
  • Trained on CPU (about 55 minutes)

Limitations

  • Trained on English newswire from the 1990s, so accuracy drops on social media, chat-style or lowercase text.
  • Uses a cased model, so capitalisation matters ("Aryan" is easier than "aryan").
  • Does not detect dates, money or other types outside the four CoNLL labels.
Downloads last month
17
Safetensors
Model size
65.2M params
Tensor type
F32
ยท
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for aryanmaurya006/ner-distilbert-conll2003

Finetuned
(347)
this model

Dataset used to train aryanmaurya006/ner-distilbert-conll2003

Evaluation results