whisper-non-natives

This model is a fine-tuned version of openai/whisper-large-v2 on the JASMIN-CGN dataset. It achieves the following results on the evaluation set:

• Loss: 0.4342
• Wer: 15.2137

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: 0.0001
• train_batch_size: 48
• eval_batch_size: 32
• seed: 42
• optimizer: Use adamw_torch_fused with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 22
• num_epochs: 3.0
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
1.185	0.2597	20	1.1092	40.1251
0.8088	0.5195	40	0.6543	21.4859
0.668	0.7792	60	0.5297	19.9739
0.6581	1.0390	80	0.4845	16.3618
0.5904	1.2987	100	0.4673	15.6524
0.5183	1.5584	120	0.4547	15.7551
0.515	1.8182	140	0.4466	15.2977
0.4805	2.0779	160	0.4406	15.2697
0.4733	2.3377	180	0.4379	15.0364
0.4846	2.5974	200	0.4353	15.2604
0.4969	2.8571	220	0.4342	15.2137

Framework versions

• PEFT 0.17.1
• Transformers 4.57.6
• Pytorch 2.8.0+cu128
• Datasets 4.5.0
• Tokenizers 0.22.2