Zarnabh/whisper-large-ps

Whisper large Ps - ZFA

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

• Loss: 0.7591
• Wer: 24.1861

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: 2e-05
• train_batch_size: 1
• eval_batch_size: 1
• seed: 42
• gradient_accumulation_steps: 32
• total_train_batch_size: 32
• optimizer: Use OptimizerNames.ADAMW_TORCH_FUSED with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
• lr_scheduler_type: SchedulerType.LINEAR
• lr_scheduler_warmup_steps: 500
• training_steps: 3000
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
14.2763	12.8298	500	0.639	28.7915
0.574	25.6483	1000	0.6891	25.7107
0.0377	38.4668	1500	0.7166	24.8849
0.0033	51.2853	2000	0.7313	24.3132
0.0023	64.1037	2500	0.7548	24.202
0.0019	76.9335	3000	0.7591	24.1861

Framework versions

• Transformers 5.0.0
• Pytorch 2.10.0+cpu
• Datasets 4.0.0
• Tokenizers 0.22.2