Makkoen/whisper-large-cit-synth-do0.15-wd0-lr1e-06-1000

./whisper-large-cit-synth-do0.15-wd0-lr1e-06-mask-1000

This model is a fine-tuned version of openai/whisper-large-v3 on the SF 1000 dataset. It achieves the following results on the evaluation set:

• Loss: 0.3877
• Wer: 24.9123

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: 300
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
1.0947	0.3556	20	0.8311	36.7251
0.8896	0.7111	40	0.7202	34.7368
0.8418	1.0667	60	0.6216	32.0078
0.6567	1.4222	80	0.5254	30.7212
0.5491	1.7778	100	0.4690	27.6803
0.5497	2.1333	120	0.4368	26.6667
0.4875	2.4889	140	0.4211	25.7310
0.4721	2.8444	160	0.4124	25.3801
0.46	3.2	180	0.4026	25.3801
0.4342	3.5556	200	0.3960	24.9513
0.4248	3.9111	220	0.3945	24.8733
0.4249	4.2667	240	0.3916	24.9123
0.4192	4.6222	260	0.3899	24.7953
0.3823	4.9778	280	0.3884	24.6004
0.4176	5.3333	300	0.3877	24.9123

Framework versions

• Transformers 4.42.3
• Pytorch 1.13.1+cu117
• Datasets 2.20.0
• Tokenizers 0.19.1