Bidwill/whisper-small-sanskrit-lr-scheduler-2

Whisper Small Sanskrit lr scheduler - Bidit Sadhukhan

This model is a fine-tuned version of openai/whisper-small on the load_data dataset. It achieves the following results on the evaluation set:

• Loss: 0.1133
• Wer: 29.2638

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: 2.55e-05
• train_batch_size: 8
• eval_batch_size: 8
• seed: 44
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine
• lr_scheduler_warmup_steps: 500
• training_steps: 11000

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
0.1184	0.12	500	0.1969	46.2909
0.0912	0.23	1000	0.1741	43.2205
0.0808	0.35	1500	0.1676	40.2062
0.0645	0.47	2000	0.1656	39.5282
0.0659	0.58	2500	0.1477	35.9704
0.0826	0.7	3000	0.1515	36.6876
0.0559	0.82	3500	0.1389	36.0769
0.0482	0.93	4000	0.1243	32.7880
0.034	1.05	4500	0.1352	34.4408
0.0275	1.17	5000	0.1301	32.0876
0.0284	1.28	5500	0.1249	30.5188
0.0268	1.4	6000	0.1343	31.7010
0.021	1.52	6500	0.1190	29.6784
0.0245	1.63	7000	0.1133	28.9444
0.0219	1.75	7500	0.1118	27.7734
0.0235	1.87	8000	0.1051	31.2808
0.0217	1.98	8500	0.1125	31.4713
0.0064	2.1	9000	0.1109	27.9751
0.0078	2.21	9500	0.1156	27.6053
0.0099	2.33	10000	0.1136	28.7483
0.01	2.45	10500	0.1133	29.2806
0.009	2.56	11000	0.1133	29.2638

Framework versions

• Transformers 4.35.0.dev0
• Pytorch 2.0.1+cu117
• Datasets 2.14.5
• Tokenizers 0.14.0