wh-ft-lre4-adm-ga1ba16-st15k-pat3

This model is a fine-tuned version of openai/whisper-large-v3 on an unknown dataset. It achieves the following results on the evaluation set:

• Loss: 1.1800
• Wer: 0.5459
• Cer: 0.4128

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: 8
• seed: 42
• optimizer: Use adamw_torch with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 750
• training_steps: 15000
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer	Cer
1.7117	0.1994	1000	1.8245	2.2154	1.7040
1.4651	0.3989	2000	1.7432	1.9903	1.8929
1.5113	0.5983	3000	1.6188	1.0851	0.8643
1.2538	0.7978	4000	1.5483	1.3806	1.1527
1.2508	0.9972	5000	1.4236	1.2864	1.0345
0.9083	1.1966	6000	1.4354	1.5115	1.2498
0.9828	1.3961	7000	1.3658	0.9501	0.7283
0.8083	1.5955	8000	1.3190	0.9102	0.7358
0.8031	1.7950	9000	1.2867	1.2175	0.9841
0.8002	1.9944	10000	1.2368	1.0241	0.8324
0.4293	2.1939	11000	1.2690	0.7188	0.5591
0.3809	2.3933	12000	1.2431	0.6259	0.4710
0.3762	2.5927	13000	1.2194	0.5935	0.4456
0.4077	2.7922	14000	1.1852	0.5891	0.4419
0.3022	2.9916	15000	1.1800	0.5459	0.4128

Framework versions

• Transformers 4.52.3
• Pytorch 2.7.0+cu118
• Datasets 3.5.1
• Tokenizers 0.21.1