mBERT-base-cased-NER-CONLL (EN-ES)

This model is a fine-tuned version of bert-base-multilingual-cased on the conll2003 and conll2002 datasets. Training was performed separately. It achieves the following results on the evaluation set:

Connll2003:

Loss: 0.0585
Precision: 0.9489
Recall: 0.9541
F1: 0.9515
Accuracy: 0.9880

Conll2002:

Loss: 0.1435
Precision: 0.8621
Recall: 0.8663
F1: 0.8642
Accuracy: 0.9791

Model description

IOB tagging Scheme. PER/LOC/MISC/ORG tags

Intended uses & limitations

More information needed

Training and evaluation data

Conll2002/2003 (ES-EN)

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

learning_rate: 3e-05
train_batch_size: 2
eval_batch_size: 2
seed: 42
optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
lr_scheduler_type: linear
num_epochs: 4

Training results

Conll2003:

Training Loss	Epoch	Step	Validation Loss	Precision	Recall	F1	Accuracy
0.1739	1.0	878	0.0741	0.9246	0.9181	0.9213	0.9823
0.045	2.0	1756	0.0586	0.9469	0.9476	0.9472	0.9870
0.0213	3.0	2634	0.0583	0.9503	0.9510	0.9506	0.9877
0.0113	4.0	3512	0.0585	0.9489	0.9541	0.9515	0.9880

Conll2002:

Training Loss	Epoch	Step	Validation Loss	Precision	Recall	F1	Accuracy
0.0739	1.0	4162	0.1322	0.8430	0.8267	0.8348	0.9741
0.0454	2.0	8324	0.1158	0.8664	0.8614	0.8639	0.9782
0.031	3.0	12486	0.1243	0.8521	0.8660	0.8590	0.9783
0.0136	4.0	16648	0.1435	0.8621	0.8663	0.8642	0.9791

Framework versions

Transformers 4.12.3
Pytorch 1.10.0+cu111
Datasets 1.15.1
Tokenizers 0.10.3

StivenLancheros
/

mBERT-base-cased-NER-CONLL