hamsa-tiny-pretrained

This model is a fine-tuned version of nadsoft/Hamsa-tiny on the nadsoft/QASR-Speech-Resource default dataset. It achieves the following results on the evaluation set:

• Loss: 0.3795
• Wer: 28.7264

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: 1e-05
• train_batch_size: 64
• 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: 50000
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
0.6597	0.1	2500	0.6394	48.8384
0.5442	0.2	5000	0.5455	41.8543
0.4954	0.3	7500	0.5018	39.8609
0.474	0.4	10000	0.4770	38.5534
0.4696	0.5	12500	0.4566	36.2515
0.4312	0.6	15000	0.4433	36.8780
0.4208	0.7	17500	0.4308	32.3714
0.4089	0.8	20000	0.4229	33.4109
0.4163	0.9	22500	0.4143	32.5423
0.3831	1.0	25000	0.4077	31.6951
0.3842	1.1	27500	0.4023	33.6316
0.3848	1.2	30000	0.3984	30.1099
0.3774	1.3	32500	0.3948	29.2864
0.3667	1.4	35000	0.3912	29.5166
0.3674	1.5	37500	0.3881	29.6115
0.3721	1.6	40000	0.3851	30.4065
0.3533	1.7	42500	0.3834	27.9693
0.3594	1.8	45000	0.3815	28.8569
0.3628	1.9	47500	0.3802	28.1260
0.3392	2.0	50000	0.3795	28.7264

Framework versions

• Transformers 4.37.0.dev0
• Pytorch 2.1.2+cu121
• Datasets 2.16.2.dev0
• Tokenizers 0.15.0