wav2vec2-large-xls-r-300m-abkhaz-cv8

This model is a fine-tuned version of facebook/wav2vec2-xls-r-300m on the MOZILLA-FOUNDATION/COMMON_VOICE_8_0 - AB dataset. It achieves the following results on the evaluation set:

• Loss: 0.1614
• Wer: 0.2907

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: 7e-05
• train_batch_size: 32
• eval_batch_size: 16
• seed: 42
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 4000
• num_epochs: 50.0
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
1.2881	4.26	4000	0.3764	0.6461
1.0767	8.53	8000	0.2657	0.5164
0.9841	12.79	12000	0.2330	0.4445
0.9274	17.06	16000	0.2134	0.3929
0.8781	21.32	20000	0.1945	0.3886
0.8381	25.59	24000	0.1840	0.3737
0.8054	29.85	28000	0.1756	0.3523
0.7763	34.12	32000	0.1745	0.3299
0.7474	38.38	36000	0.1677	0.3074
0.7298	42.64	40000	0.1649	0.2963
0.7125	46.91	44000	0.1617	0.2931

Framework versions

• Transformers 4.16.0.dev0
• Pytorch 1.10.1+cu102
• Datasets 1.18.3
• Tokenizers 0.11.0