jssaluja/wav2vec2-large-mms-1b-all-rajinder_singh-without-separator-epochs-10-test-datasets-10

This model is a fine-tuned version of facebook/mms-1b-all on the jssaluja/rajinder_singh dataset. It achieves the following results on the evaluation set:

• Loss: 0.3321
• Wer: 0.3618
• Wil: 0.5605
• Mer: 0.3544
• Cer: 0.0961

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: 0.001
• train_batch_size: 8
• eval_batch_size: 8
• seed: 42
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 75
• num_epochs: 10
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer	Wil	Mer	Cer
0.3757	1.0	7548	0.3935	0.4691	0.6659	0.4422	0.1178
0.2114	2.0	15096	0.3443	0.4027	0.6111	0.3941	0.1032
0.1858	3.0	22644	0.3323	0.3945	0.6017	0.3878	0.1010
0.1687	4.0	30192	0.3211	0.3886	0.5925	0.3806	0.0982
0.1553	5.0	37740	0.3307	0.3819	0.5856	0.3744	0.0992
0.1455	6.0	45288	0.3278	0.3826	0.5888	0.3773	0.0984
0.1376	7.0	52836	0.3235	0.3666	0.5689	0.3597	0.0954
0.1304	8.0	60384	0.3309	0.3638	0.5636	0.3558	0.0956
0.1239	9.0	67932	0.3290	0.3647	0.5654	0.3587	0.0960
0.1185	10.0	75480	0.3321	0.3618	0.5605	0.3544	0.0961

Framework versions

• Transformers 4.42.3
• Pytorch 2.2.1+cu121
• Datasets 2.20.0
• Tokenizers 0.19.1