dat259-nor-wav2vec2

This model is a fine-tuned version of NbAiLab/nb-wav2vec2-300m-nynorsk on the common_voice_8_0 dataset. It achieves the following results on the evaluation set:

• Loss: 10.9446
• Wer: 1.1259

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: 0.0001
• train_batch_size: 16
• eval_batch_size: 8
• seed: 42
• gradient_accumulation_steps: 2
• total_train_batch_size: 32
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 5
• num_epochs: 10
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
84.8696	1.57	5	91.5942	1.0
62.5471	3.29	10	33.8515	1.0068
20.2215	4.86	15	17.4461	1.0017
15.2892	6.57	20	13.5454	1.0034
12.8086	8.29	25	12.0084	1.0408
11.0168	9.86	30	10.9446	1.1259

Framework versions

• Transformers 4.21.1
• Pytorch 1.10.0+cu113
• Datasets 2.4.0
• Tokenizers 0.12.1