Kalbert / README.md
Chakita's picture
Update README.md
c7915c9
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