mbert-hate-final-1
This model is a fine-tuned version of bert-base-multilingual-cased on the None dataset. It achieves the following results on the evaluation set:
- Loss: 0.6751
- Accuracy: 0.7272
- Precision: 0.7260
- Recall: 0.7272
- F1: 0.7188
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: 1e-05
- train_batch_size: 32
- eval_batch_size: 32
- seed: 42
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: linear
- num_epochs: 6
Training results
| Training Loss | Epoch | Step | Validation Loss | Accuracy | Precision | Recall | F1 |
|---|---|---|---|---|---|---|---|
| No log | 1.0 | 296 | 0.5531 | 0.6835 | 0.6803 | 0.6835 | 0.6812 |
| 0.5365 | 2.0 | 592 | 0.5304 | 0.7405 | 0.7432 | 0.7405 | 0.7302 |
| 0.5365 | 3.0 | 888 | 0.5526 | 0.7310 | 0.7334 | 0.7310 | 0.7195 |
| 0.4318 | 4.0 | 1184 | 0.6142 | 0.7186 | 0.7153 | 0.7186 | 0.7136 |
| 0.4318 | 5.0 | 1480 | 0.6420 | 0.7243 | 0.7227 | 0.7243 | 0.7162 |
| 0.3507 | 6.0 | 1776 | 0.6751 | 0.7272 | 0.7260 | 0.7272 | 0.7188 |
Framework versions
- Transformers 4.24.0.dev0
- Pytorch 1.11.0+cu102
- Datasets 2.6.1
- Tokenizers 0.13.1