./whisper-large-cit-do0.25-wd0-lr1e-06

This model is a fine-tuned version of openai/whisper-large-v3 on the SF 200 dataset. It achieves the following results on the evaluation set:

• Loss: 0.6753
• Wer: 34.3249

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-06
• train_batch_size: 4
• eval_batch_size: 8
• seed: 42
• distributed_type: multi-GPU
• gradient_accumulation_steps: 4
• total_train_batch_size: 16
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 100
• training_steps: 200
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
1.1266	0.8889	10	1.1143	48.9703
1.0858	1.7778	20	1.0078	39.5881
0.9333	2.6667	30	0.8696	39.1304
0.7546	3.5556	40	0.7925	34.0961
0.7029	4.4444	50	0.7202	34.7826
0.601	5.3333	60	0.6558	32.9519
0.5097	6.2222	70	0.6177	31.5789
0.415	7.1111	80	0.5913	33.4096
0.3794	8.0	90	0.5728	32.9519
0.3026	8.8889	100	0.5693	33.4096
0.2687	9.7778	110	0.5732	33.1808
0.2175	10.6667	120	0.5825	31.8078
0.1864	11.5556	130	0.5942	33.1808
0.155	12.4444	140	0.6060	31.1213
0.1266	13.3333	150	0.6182	33.1808
0.1212	14.2222	160	0.6328	33.8673
0.1043	15.1111	170	0.6499	33.8673
0.0894	16.0	180	0.6616	32.7231
0.084	16.8889	190	0.6724	33.8673
0.0787	17.7778	200	0.6753	34.3249

Framework versions

• Transformers 4.41.1
• Pytorch 1.13.1+cu117
• Datasets 2.19.1
• Tokenizers 0.19.1