w2v-bert-2.0-br

This model is a fine-tuned version of facebook/w2v-bert-2.0 on the None dataset. It achieves the following results on the evaluation set:

• Loss: 0.6660
• Wer: 42.4942
• Cer: 13.6525

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: 6e-05
• train_batch_size: 4
• eval_batch_size: 8
• seed: 42
• gradient_accumulation_steps: 2
• total_train_batch_size: 8
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_ratio: 0.08
• lr_scheduler_warmup_steps: 500
• training_steps: 8001
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer	Cer
1.0369	0.58	500	1.2289	85.3288	32.8021
0.7211	1.16	1000	0.9727	70.1973	24.6147
0.5669	1.75	1500	0.8496	64.6176	21.7978
0.4229	2.33	2000	0.7448	57.2663	19.3988
0.4352	2.91	2500	0.6749	52.9790	17.4075
0.3392	3.49	3000	0.6703	50.9678	16.8375
0.2508	4.07	3500	0.6143	49.6249	16.2547
0.2303	4.65	4000	0.7121	48.4648	15.8534
0.1776	5.24	4500	0.6667	47.0777	15.2910
0.1645	5.82	5000	0.6715	46.1825	14.8910
0.1304	6.4	5500	0.7212	44.2784	14.5139
0.1157	6.98	6000	0.6678	44.2721	14.3043
0.0924	7.56	6500	0.6935	43.1310	13.9171
0.0517	8.14	7000	0.6746	42.8851	13.7599
0.0667	8.73	7500	0.6327	42.9733	13.8136
0.0483	9.31	8000	0.6660	42.4942	13.6525

Framework versions

• Transformers 4.39.1
• Pytorch 2.0.1+cu117
• Datasets 2.18.0
• Tokenizers 0.15.2