# bert-small-buddhist-nonbuddhist-sanskrit

BERT model trained on a lemmatized corpus containing Buddhist and non-Buddhist Sanskrit texts. 

## Model description

The model has the bert architecture and was pretrained from scratch as a masked language model 
on the lemmatized Sanskrit corpus. Due to lack of resources and to prevent overfitting, the model is smaller than bert-base, 
i.e. the number of attention heads and hidden layers have been reduced to 8 and the context has been reduced to 128 tokens. Vocabulary size is 10000 tokens.

## How to use it

```
model = AutoModelForMaskedLM.from_pretrained("Matej/bert-small-buddhist-nonbuddhist-sanskrit")
tokenizer = AutoTokenizer.from_pretrained("Matej/bert-small-buddhist-nonbuddhist-sanskrit", use_fast=True)
```

## Intended uses & limitations

MIT license, no limitations

## Training and evaluation data

See the paper 'Embeddings models for Buddhist Sanskrit' for details on the corpora and the evaluation procedure.

### Training hyperparameters

The following hyperparameters were used during training:
- learning_rate: 5e-05
- train_batch_size: 32
- eval_batch_size: 4
- seed: 42
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: linear
- num_epochs: 200

### Framework versions

- Transformers 4.20.0
- Pytorch 1.9.0
- Datasets 2.3.2
- Tokenizers 0.12.1