vrclc/W2V2-BERT-withLM-Malayalam-Studio

W2V2-BERT-withLM-Studio-Malayalam

This model is a fine-tuned version of facebook/w2v-bert-2.0 on IMASC, OpenSLR Malayalam Train split, Festvox Malayalamdataset. It achieves the following results on the evaluation set:

• Loss: 0.1587
• Wer: 0.1157

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: 5e-05
• train_batch_size: 16
• eval_batch_size: 8
• seed: 42
• gradient_accumulation_steps: 2
• total_train_batch_size: 32
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 500
• num_epochs: 10
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
1.0335	0.4932	600	0.3654	0.4387
0.1531	0.9864	1200	0.2373	0.3332
0.1074	1.4797	1800	0.2069	0.2953
0.0928	1.9729	2400	0.2146	0.2814
0.0734	2.4661	3000	0.1947	0.2433
0.0678	2.9593	3600	0.1938	0.2406
0.0522	3.4525	4200	0.1566	0.2053
0.0493	3.9457	4800	0.1649	0.1988
0.0366	4.4390	5400	0.1417	0.1834
0.0372	4.9322	6000	0.1542	0.1749
0.028	5.4254	6600	0.1476	0.1620
0.0263	5.9186	7200	0.1388	0.1622
0.0195	6.4118	7800	0.1384	0.1495
0.0185	6.9051	8400	0.1351	0.1383
0.0136	7.3983	9000	0.1404	0.1344
0.0119	7.8915	9600	0.1253	0.1276
0.0087	8.3847	10200	0.1443	0.1284
0.0066	8.8779	10800	0.1475	0.1252
0.0049	9.3711	11400	0.1577	0.1227
0.0038	9.8644	12000	0.1587	0.1157

Framework versions

• Transformers 4.42.2
• Pytorch 2.1.1+cu121
• Datasets 2.20.0
• Tokenizers 0.19.1