This model is a fine-tuned version of facebook/m2m100_418M on Deshika dataset which is a prallel corpus of Prakrit with their corresponding English Translations It achieves the following results on the evaluation set:

Loss: 1.3269
Bleu: 8.4241
Meteor: 0.3851
Gen Len: 30.7356

Model Description

praTran is a finetuned version of facebook/m2m100_418M which was trained on a downstream task.

Intended uses

This model is intended to use for academic purposes.

Limitation

The models translation is not that good at this current stage to the language being extremely low resource. Impro

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: 16
eval_batch_size: 16
seed: 42
optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
lr_scheduler_type: linear
num_epochs: 5
mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Bleu	Meteor	Gen Len
No log	1.0	74	4.9214	2.8147	0.2385	31.7864
No log	2.0	148	3.1788	5.0267	0.3144	29.7695
No log	3.0	222	2.0374	6.4844	0.3399	30.2237
No log	4.0	296	1.4798	7.4708	0.3768	31.3932
No log	5.0	370	1.3269	8.4241	0.3851	30.7356

Framework versions

Transformers 4.45.1
Pytorch 2.4.0
Datasets 3.0.1
Tokenizers 0.20.0

VIITPune/praTran

praTran