Wav2vec2-fula

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

• Loss: 0.3143
• Wer: 0.5455

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: 5e-05
• train_batch_size: 8
• eval_batch_size: 8
• seed: 42
• gradient_accumulation_steps: 4
• 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: 1000
• num_epochs: 60.0
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
3.1344	0.3437	500	3.0935	1.0
0.7323	0.6874	1000	0.6304	0.7120
0.5416	1.0316	1500	0.4785	0.6491
0.4479	1.3753	2000	0.4202	0.6207
0.4541	1.7190	2500	0.3851	0.6006
0.365	2.0632	3000	0.3701	0.5885
0.3433	2.4069	3500	0.3648	0.5797
0.3561	2.7506	4000	0.3438	0.5716
0.3237	3.0949	4500	0.3647	0.5677
0.322	3.4386	5000	0.3427	0.5638
0.2921	3.7823	5500	0.3345	0.5604
0.3037	4.1265	6000	0.3352	0.5541
0.2695	4.4702	6500	0.3202	0.5515
0.2804	4.8139	7000	0.3353	0.5525
0.2908	5.1581	7500	0.3384	0.5485
0.2646	5.5018	8000	0.3164	0.5462
0.2982	5.8455	8500	0.3143	0.5455
0.2978	6.1897	9000	0.3218	0.5424
0.288	6.5334	9500	0.3152	0.5418
0.2706	6.8771	10000	0.3211	0.5398
0.3008	7.2213	10500	0.3266	0.5398
0.2674	7.5650	11000	0.3185	0.5379

Framework versions

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