indic-gpt
This model is a fine-tuned version of gpt2 on an Indian Language(https://ai4bharat.iitm.ac.in/corpora) dataset. Sample Dataset is present on https://huggingface.co/datasets/aashay96/indic-gpt. It achieves the following results on the evaluation set:
- Loss: 1.9482
Model description
Model is trained on multiple Indian Languages - Assamese, bengali, gujarati, Kannada, Malayalam,telugu, tamil, odhiya and punjabi.
Intended uses & limitations
More information needed
Training and evaluation data
TBD - Evaluation on indic_glue
Training procedure
Check the notebook!
Training hyperparameters
The following hyperparameters were used during training:
- learning_rate: 0.0005
- train_batch_size: 128
- eval_batch_size: 128
- seed: 42
- gradient_accumulation_steps: 8
- total_train_batch_size: 1024
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: cosine
- lr_scheduler_warmup_steps: 100
- num_epochs: 1
- mixed_precision_training: Native AMP
Training results
| Training Loss | Epoch | Step | Validation Loss |
|---|---|---|---|
| 3.3653 | 0.3 | 500 | 2.2985 |
| 2.2079 | 0.61 | 1000 | 2.0401 |
| 2.0396 | 0.91 | 1500 | 1.9482 |
Framework versions
- Transformers 4.25.1
- Pytorch 1.13.0+cu116
- Datasets 2.8.0
- Tokenizers 0.13.2