Token Classification
Transformers
PyTorch
TensorBoard
bert
Generated from Trainer
Eval Results (legacy)
# Load model directly
from transformers import AutoTokenizer, AutoModelForTokenClassification
tokenizer = AutoTokenizer.from_pretrained("Wende/bert-finetuned-ner")
model = AutoModelForTokenClassification.from_pretrained("Wende/bert-finetuned-ner")Quick Links
bert-finetuned-ner
This model is a fine-tuned version of bert-base-cased on the conll2003 dataset. It achieves the following results on the evaluation set:
- Loss: 0.0575
- Precision: 0.9322
- Recall: 0.9505
- F1: 0.9413
- Accuracy: 0.9860
Model description
More information needed
Intended uses & limitations
More information needed
Training and evaluation data
More information needed
Training procedure
Training hyperparameters
The following hyperparameters were used during training:
- learning_rate: 2e-05
- train_batch_size: 16
- eval_batch_size: 16
- seed: 42
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: linear
- num_epochs: 3
Training results
| Training Loss | Epoch | Step | Validation Loss | Precision | Recall | F1 | Accuracy |
|---|---|---|---|---|---|---|---|
| 0.2219 | 1.0 | 878 | 0.0716 | 0.9076 | 0.9288 | 0.9181 | 0.9808 |
| 0.0453 | 2.0 | 1756 | 0.0597 | 0.9297 | 0.9477 | 0.9386 | 0.9852 |
| 0.0239 | 3.0 | 2634 | 0.0575 | 0.9322 | 0.9505 | 0.9413 | 0.9860 |
Framework versions
- Transformers 4.12.3
- Pytorch 1.8.2+cu111
- Datasets 1.15.1
- Tokenizers 0.10.3
- Downloads last month
- 6
Dataset used to train Wende/bert-finetuned-ner
Evaluation results
- Precision on conll2003self-reported0.932
- Recall on conll2003self-reported0.951
- F1 on conll2003self-reported0.941
- Accuracy on conll2003self-reported0.986
# Use a pipeline as a high-level helper from transformers import pipeline pipe = pipeline("token-classification", model="Wende/bert-finetuned-ner")