Wav2vec2-fula-no0

This model is a fine-tuned version of facebook/wav2vec2-xls-r-300m on the LEONEL-MAIA/FULFULDE-NO0 - DEFAULT dataset. It achieves the following results on the evaluation set:

• Loss: 0.4407
• Wer: 0.5583
• Cer: 0.1549

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: 8
• eval_batch_size: 8
• seed: 42
• gradient_accumulation_steps: 2
• total_train_batch_size: 16
• 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: 1000
• num_epochs: 60.0
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer	Cer
0.5615	0.1970	500	0.6706	0.7329	0.2082
0.3606	0.3939	1000	0.5461	0.6352	0.1786
0.3103	0.5909	1500	0.4960	0.6026	0.1695
0.228	0.7879	2000	0.5087	0.5917	0.1635
0.2502	0.9848	2500	0.4720	0.5835	0.1648
0.2841	1.1816	3000	0.4652	0.6019	0.1701
0.2603	1.3786	3500	0.4726	0.5685	0.1588
0.2538	1.5755	4000	0.4538	0.5667	0.1573
0.2215	1.7725	4500	0.4649	0.5615	0.1581
0.2329	1.9695	5000	0.4407	0.5583	0.1549
0.2111	2.1662	5500	0.4677	0.5529	0.1542
0.2018	2.3632	6000	0.4574	0.5507	0.1531
0.2421	2.5602	6500	0.4417	0.5500	0.1528

Framework versions

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