distilroberta-base-ner-conll2003

This model is a fine-tuned version of distilroberta-base on the conll2003 dataset.

eval F1-Score: 95,29 (CoNLL-03)
test F1-Score: 90,74 (CoNLL-03)

eval F1-Score: 95,29 (CoNLL++ / CoNLL-03 corrected)
test F1-Score: 92,23 (CoNLL++ / CoNLL-03 corrected)

Model Usage

from transformers import AutoTokenizer, AutoModelForTokenClassification
from transformers import pipeline

tokenizer = AutoTokenizer.from_pretrained("philschmid/distilroberta-base-ner-conll2003")
model = AutoModelForTokenClassification.from_pretrained("philschmid/distilroberta-base-ner-conll2003")

nlp = pipeline("ner", model=model, tokenizer=tokenizer, grouped_entities=True)
example = "My name is Philipp and live in Germany"

nlp(example)

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 4.9902376275441704e-05
• train_batch_size: 32
• eval_batch_size: 16
• seed: 42
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• num_epochs: 6.0
• mixed_precision_training: Native AMP

Training results

CoNNL2003

It achieves the following results on the evaluation set:

• Loss: 0.0583
• Precision: 0.9493
• Recall: 0.9566
• F1: 0.9529
• Accuracy: 0.9883

It achieves the following results on the test set:

• Loss: 0.2025
• Precision: 0.8999
• Recall: 0.915
• F1: 0.9074
• Accuracy: 0.9741

CoNNL++ / CoNLL2003 corrected

It achieves the following results on the evaluation set:

• Loss: 0.0567
• Precision: 0.9493
• Recall: 0.9566
• F1: 0.9529
• Accuracy: 0.9883

It achieves the following results on the test set:

• Loss: 0.1359
• Precision: 0.92
• Recall: 0.9245
• F1: 0.9223
• Accuracy: 0.9785

Framework versions

• Transformers 4.6.1
• Pytorch 1.8.1+cu101
• Datasets 1.6.2
• Tokenizers 0.10.2