distilroberta-base-ner-wikiann-conll2003-4-class

This model is a fine-tuned version of distilroberta-base on the wikiann and conll2003 dataset. It consists out of the classes of conll2003.

O (0), B-PER (1), I-PER (2), B-ORG (3), I-ORG (4) B-LOC (5), I-LOC (6) B-MISC (7), I-MISC (8).

eval F1-Score: 95,39 (merged dataset)
test F1-Score: 90,75 (merged dataset)

Model Usage

from transformers import AutoTokenizer, AutoModelForTokenClassification
from transformers import pipeline

tokenizer = AutoTokenizer.from_pretrained("philschmid/distilroberta-base-ner-wikiann-conll2003-4-class")
model = AutoModelForTokenClassification.from_pretrained("philschmid/distilroberta-base-ner-wikiann-conll2003-4-class")

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: 4.9086903597787154e-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: 5.0
• mixed_precision_training: Native AMP

Training results

It achieves the following results on the evaluation set:

• Loss: 0.0705
• Precision: 0.9492
• Recall: 0.9585
• F1: 0.9539
• Accuracy: 0.9882

It achieves the following results on the test set:

• Loss: 0.239
• Precision: 0.8984
• Recall: 0.9168
• F1: 0.9075
• Accuracy: 0.9741

Framework versions

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