bert-base-uncased-ft-news

This model is a fine-tuned version of bert-base-uncased on the news dataset. It achieves the following results on the evaluation set:

• Loss: 0.4300
• Accuracy: 0.9
• F1: 0.8783

and flowing results on the testing set:

• Accuracy: 0.8954
• F1: 0.8784

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
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Accuracy	F1
0.4196	0.37	120	0.3051	0.8875	0.8566
0.3101	0.75	240	0.2979	0.8953	0.8743
0.2693	1.12	360	0.3162	0.9016	0.8831
0.2078	1.5	480	0.3298	0.8984	0.8767
0.1725	1.87	600	0.3801	0.9047	0.8851
0.1369	2.24	720	0.3901	0.8938	0.8677
0.1101	2.62	840	0.4160	0.9016	0.8805
0.1019	2.99	960	0.4300	0.9	0.8783

Framework versions

• Transformers 4.25.1
• Pytorch 1.13.0+cu116
• Datasets 2.8.0
• Tokenizers 0.13.2