torgo_tiny_finetune_F01

This model is a fine-tuned version of openai/whisper-tiny on the None dataset. It achieves the following results on the evaluation set:

• Loss: 0.2909
• Wer: 24.6180

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: 0.0001
• train_batch_size: 16
• eval_batch_size: 1
• seed: 42
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 1000
• num_epochs: 20

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
0.6321	0.83	500	0.3242	51.0187
0.1046	1.66	1000	0.4511	55.3480
0.099	2.49	1500	0.3225	31.5789
0.0656	3.32	2000	0.3007	50.6791
0.0506	4.15	2500	0.2984	27.6740
0.0383	4.98	3000	0.2853	23.6842
0.0296	5.8	3500	0.3449	32.3430
0.0198	6.63	4000	0.2730	26.6553
0.0192	7.46	4500	0.3049	49.2360
0.0136	8.29	5000	0.3279	25.8065
0.0121	9.12	5500	0.3082	23.8540
0.0101	9.95	6000	0.2722	25.5518
0.0065	10.78	6500	0.3414	32.0883
0.0062	11.61	7000	0.3140	22.9202
0.0053	12.44	7500	0.2601	24.7029
0.002	13.27	8000	0.2978	33.8710
0.0021	14.1	8500	0.2798	31.1545
0.0011	14.93	9000	0.3137	25.1273
0.0006	15.75	9500	0.2926	22.2411
0.0003	16.58	10000	0.2891	23.4295
0.0001	17.41	10500	0.2930	25.2122
0.0001	18.24	11000	0.2906	24.7878
0.0001	19.07	11500	0.2906	24.6180
0.0	19.9	12000	0.2909	24.6180

Framework versions

• Transformers 4.32.0
• Pytorch 2.1.0+cu121
• Datasets 2.14.7
• Tokenizers 0.13.3