metadata
license: mit
tags:
- generated_from_trainer
model-index:
- name: Kalbert
results: []
Kalbert
This model is a fine-tuned version of ai4bharat/indic-bert on a kannada news dataset. It achieves the following results on the evaluation set:
- Loss: 1.5324
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: 4
- eval_batch_size: 4
- seed: 42
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: linear
- num_epochs: 10
- mixed_precision_training: Native AMP
Training results
| Training Loss | Epoch | Step | Validation Loss |
|---|---|---|---|
| 1.5835 | 1.0 | 3953 | 1.7985 |
| 1.6098 | 2.0 | 7906 | 1.7434 |
| 1.5266 | 3.0 | 11859 | 1.6934 |
| 1.5179 | 4.0 | 15812 | 1.6665 |
| 1.5459 | 5.0 | 19765 | 1.6135 |
| 1.5511 | 6.0 | 23718 | 1.6002 |
| 1.5209 | 7.0 | 27671 | 1.5657 |
| 1.5413 | 8.0 | 31624 | 1.5578 |
| 1.4828 | 9.0 | 35577 | 1.5465 |
| 1.4651 | 10.0 | 39530 | 1.5451 |
Framework versions
- Transformers 4.25.1
- Pytorch 1.13.0+cu116
- Datasets 2.8.0
- Tokenizers 0.13.2