Modelpublic

thunderboltc/whisper_sanlish_mnx_lr5e-6

sourceHugging Faceupdated 1mo agoView on Hugging Face

0likes27downloads

Model Card

The following table summarizes the training and validation performance observed during fine-tuning:

Epoch	Training Loss	Validation Loss	WER (%)	CER (%)
1	15.663	3.835	102.93	39.07
2	9.458	2.450	87.59	25.83
3	7.289	1.774	75.28	20.66
4	5.747	1.411	67.10	17.90
5	4.444	1.270	64.48	16.52
6	3.426	1.130	58.53	14.94
7	3.016	1.100	55.30	13.96
8	2.768	1.077	52.47	13.42
9	2.507	1.042	51.66	12.83
10	2.198	0.995	50.86	12.34
11	2.190	0.974	49.55	12.00
12	2.077	0.961	48.23	11.88
13	1.977	0.950	48.84	11.78
14	1.752	0.927	47.23	11.50
15	1.867	0.935	46.01	11.02
16	1.826	0.916	45.21	10.73

At Epoch 16, the model achieved the best validation WER observed so far:

Training is still in progress toward the planned 25 epochs, so these are intermediate validation results and may improve further.