miosipof/asr2_aug_v1

Whisper Medium IT

This model is a fine-tuned version of openai/whisper-medium on the b-brave-balanced-augmented dataset. It achieves the following results on the evaluation set:

• Loss: 0.4682
• Wer: 28.8939
• Cer: 19.7597
• Lr: 0.0000

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: 0.0003
• train_batch_size: 4
• eval_batch_size: 4
• seed: 42
• gradient_accumulation_steps: 2
• total_train_batch_size: 8
• 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_ratio: 0.3
• num_epochs: 12
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer	Cer	Lr
2.6473	1.0	165	1.2219	64.5598	37.6553	0.0001
0.8386	2.0	330	0.7353	59.1422	44.4905	0.0002
0.5917	3.0	495	0.5903	46.0497	29.5775	0.0002
0.4388	4.0	660	0.5117	36.3431	25.7249	0.0003
0.2524	5.0	825	0.5054	35.8916	24.6893	0.0003
0.1335	6.0	990	0.4751	35.4402	23.2809	0.0002
0.0587	7.0	1155	0.4769	30.4740	20.6711	0.0002
0.0388	8.0	1320	0.4668	32.9571	22.0381	0.0001
0.0124	9.0	1485	0.4646	30.4740	20.8368	0.0001
0.0087	10.0	1650	0.4582	28.2167	19.8840	0.0001
0.0022	11.0	1815	0.4655	29.7968	20.5468	0.0000
0.0022	12.0	1980	0.4682	28.8939	19.7597	0.0000

Framework versions

• PEFT 0.14.0
• Transformers 4.49.0
• Pytorch 2.2.0
• Datasets 3.3.2
• Tokenizers 0.21.1