whisper-small-fula

This model is a fine-tuned version of openai/whisper-small on the LAfricaMobile/fulfulde dataset. It achieves the following results on the evaluation set:

• Loss: 0.3445
• Wer: 0.2610

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: 4
• eval_batch_size: 4
• seed: 42
• gradient_accumulation_steps: 8
• total_train_batch_size: 32
• 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_steps: 500
• num_epochs: 60.0
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
0.8644	0.3437	500	0.7433	0.5713
0.6177	0.6874	1000	0.5597	0.4540
0.5026	1.0309	1500	0.4869	0.4128
0.5111	1.3746	2000	0.4482	0.4109
0.4172	1.7183	2500	0.4173	0.3507
0.3695	2.0619	3000	0.3985	0.3167
0.366	2.4056	3500	0.3836	0.2965
0.3372	2.7493	4000	0.3676	0.3009
0.3059	3.0928	4500	0.3578	0.2956
0.305	3.4365	5000	0.3511	0.2822
0.2882	3.7802	5500	0.3445	0.2610
0.2528	4.1237	6000	0.3560	0.3077
0.2757	4.4674	6500	0.3558	0.2969
0.2559	4.8111	7000	0.3556	0.2965
0.2558	5.1547	7500	0.3553	0.2968
0.251	5.4984	8000	0.3546	0.3072

Framework versions

• Transformers 4.50.3
• Pytorch 2.7.0+cu126
• Datasets 3.5.0
• Tokenizers 0.21.1