sarch7040/praTranv2

M2M101

This model is a fine-tuned version of facebook/m2m100_418M on an unknown dataset. It achieves the following results on the evaluation set:

• Loss: 0.6766
• Bleu: 15.3416
• Meteor: 0.4723
• Gen Len: 28.0271

Model description

This model translates Maharashtri Prakrit (an ancient Indo-Aryan Language) into English. It is fine-tuned on the M2M100 model.

Intended uses & limitations

The use of this model is educational and is intended for students, linguists and casual learners. This model is currently limited and sometimes could give incorrect translation because of small dataset and low resources ness of Maharashtri Prakrit

Training and evaluation data

This model was trained on Custom made dataset of 1474 Prakrit to English sentences. The evaluation of this model was done using BLEU and METEOR score.

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 1e-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: 20
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Bleu	Meteor	Gen Len
No log	1.0	74	5.8261	1.1786	0.1997	36.3831
No log	2.0	148	4.6170	2.648	0.2657	37.6068
No log	3.0	222	3.5128	5.7069	0.3217	32.2169
No log	4.0	296	2.5281	6.3134	0.3547	31.8576
No log	5.0	370	1.7177	8.5036	0.38	29.9729
No log	6.0	444	1.1666	10.1169	0.3925	28.0678
3.5641	7.0	518	0.8702	10.4207	0.4246	31.1051
3.5641	8.0	592	0.7376	12.6153	0.431	28.6339
3.5641	9.0	666	0.6901	13.2966	0.4503	29.2373
3.5641	10.0	740	0.6713	11.9772	0.4396	30.5661
3.5641	11.0	814	0.6651	14.0436	0.4506	30.2
3.5641	12.0	888	0.6678	13.2632	0.4514	31.0542
3.5641	13.0	962	0.6677	14.0924	0.4563	29.278
0.5121	14.0	1036	0.6693	14.746	0.4651	28.4068
0.5121	15.0	1110	0.6698	14.9278	0.4677	28.5153
0.5121	16.0	1184	0.6700	14.7431	0.4674	28.9288
0.5121	17.0	1258	0.6744	15.2934	0.4701	28.8678
0.5121	18.0	1332	0.6741	15.6776	0.4712	28.3492
0.5121	19.0	1406	0.6772	14.942	0.4707	28.9695
0.5121	20.0	1480	0.6766	15.3416	0.4723	28.0271

Framework versions

• Transformers 4.45.1
• Pytorch 2.4.0
• Datasets 3.0.1
• Tokenizers 0.20.0