Makkoen/whisper-large-cit-do0-wd0-lr5

whisper-large-cit-do0-wd0

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: 0.6895
• Wer: 34.0961

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
1.1267	0.8889	10	1.1143	48.9703
1.0863	1.7778	20	1.0078	40.7323
0.9336	2.6667	30	0.8691	38.9016
0.7543	3.5556	40	0.7925	34.0961
0.7023	4.4444	50	0.7212	35.0114
0.6007	5.3333	60	0.6558	32.9519
0.5085	6.2222	70	0.6167	31.3501
0.4119	7.1111	80	0.5898	33.1808
0.3749	8.0	90	0.5723	32.9519
0.2971	8.8889	100	0.5698	33.1808
0.2621	9.7778	110	0.5747	32.7231
0.2108	10.6667	120	0.5854	31.8078
0.1793	11.5556	130	0.5977	32.4943
0.1488	12.4444	140	0.6118	31.3501
0.1199	13.3333	150	0.6255	33.4096
0.1135	14.2222	160	0.6416	34.7826
0.097	15.1111	170	0.6606	34.5538
0.0823	16.0	180	0.6738	33.4096
0.0767	16.8889	190	0.6860	33.4096
0.0713	17.7778	200	0.6895	34.0961

Framework versions

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