kapilkd13/xls-r-300m-hi-prod

This model is a fine-tuned version of facebook/wav2vec2-xls-r-300m on the MOZILLA-FOUNDATION/COMMON_VOICE_7_0 - HI dataset. It achieves the following results on the evaluation set:

• Loss: 0.7805
• Wer: 0.4340

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: 0.0003
• train_batch_size: 16
• 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: 500
• training_steps: 8000
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
No log	1.36	400	1.9130	0.9244
5.0013	2.71	800	0.7789	0.5944
0.6544	4.07	1200	0.7298	0.5852
0.4021	5.42	1600	0.6978	0.5667
0.3003	6.78	2000	0.6764	0.5382
0.3003	8.14	2400	0.7249	0.5463
0.2345	9.49	2800	0.7280	0.5124
0.1993	10.85	3200	0.7289	0.4690
0.1617	12.2	3600	0.7431	0.4733
0.1432	13.56	4000	0.7448	0.4733
0.1432	14.92	4400	0.7746	0.4485
0.1172	16.27	4800	0.7589	0.4742
0.1035	17.63	5200	0.7539	0.4353
0.0956	18.98	5600	0.7648	0.4495
0.0845	20.34	6000	0.7877	0.4719
0.0845	21.69	6400	0.7884	0.4434
0.0761	23.05	6800	0.7796	0.4386
0.0634	24.41	7200	0.7729	0.4306
0.0571	25.76	7600	0.7826	0.4298
0.0508	27.12	8000	0.7805	0.4340

Framework versions

• Transformers 4.16.2
• Pytorch 1.10.1+cu102
• Datasets 1.18.3
• Tokenizers 0.11.0