w2v-bert-2.0-hausa_naijavoices_1h

This model is a fine-tuned version of facebook/w2v-bert-2.0 on the CLEAR-GLOBAL/NAIJAVOICES_1H - NA dataset. It achieves the following results on the evaluation set:

• Loss: 0.5645
• Wer: 0.4720
• Cer: 0.2334

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: 3e-05
• train_batch_size: 160
• eval_batch_size: 160
• seed: 42
• gradient_accumulation_steps: 2
• total_train_batch_size: 320
• optimizer: Use OptimizerNames.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_ratio: 0.1
• num_epochs: 25000.0
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer	Cer
5.1544	25.0	100	4.9154	1.4117	1.1735
3.0183	50.0	200	3.0473	0.9804	0.9046
2.6561	75.0	300	2.7221	0.9997	0.8925
1.9571	100.0	400	1.8462	1.0862	0.5879
1.0563	125.0	500	0.9029	0.7354	0.3151
0.5795	150.0	600	0.6469	0.5355	0.2493
0.5822	175.0	700	0.5837	0.4906	0.2379
0.408	200.0	800	0.5645	0.4719	0.2334
0.2582	225.0	900	0.5650	0.4604	0.2307
0.2256	250.0	1000	0.5710	0.4618	0.2310
0.1369	275.0	1100	0.5925	0.4547	0.2290
0.1305	300.0	1200	0.6295	0.4502	0.2277
0.1187	325.0	1300	0.6640	0.4531	0.2284

Framework versions

• Transformers 4.48.1
• Pytorch 2.7.1+cu126
• Datasets 3.6.0
• Tokenizers 0.21.1