miosipof/asr2_aug_IT_v4_full

Whisper Medium IT

This model is a fine-tuned version of miosipof/asr2_aug_IT_v4_merged on the b-brave-balanced-augmented dataset. It achieves the following results on the evaluation set:

• Loss: 0.0005
• Wer: 0.0
• Cer: 0.0
• 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: 1e-05
• train_batch_size: 32
• eval_batch_size: 16
• seed: 42
• gradient_accumulation_steps: 4
• total_train_batch_size: 128
• optimizer: Use adamw_torch_4bit 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: 10
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer	Cer	Lr
1.1582	1.0	68	0.1790	24.6508	16.4515	0.0000
0.9748	2.0	136	0.1054	15.6943	10.5674	0.0000
0.5207	3.0	204	0.0574	10.3533	7.6854	0.0000
0.3516	4.0	272	0.0296	19.9671	18.2828	0.0000
0.2355	5.0	340	0.0113	2.3007	1.6361	0.0000
0.1018	6.0	408	0.0056	0.3287	0.2252	0.0000
0.0639	7.0	476	0.0037	0.1643	0.0600	0.0000
0.0443	8.0	544	0.0033	0.4108	0.2252	0.0000
0.015	9.0	612	0.0008	0.0	0.0	0.0000
0.0095	9.8625	670	0.0005	0.0	0.0	0.0000

Framework versions

• Transformers 4.47.1
• Pytorch 2.6.0+cu124
• Datasets 3.5.0
• Tokenizers 0.21.1