wav2vec2-large-xls-r-300m-georgian

This model is a fine-tuned version of facebook/wav2vec2-xls-r-300m on the MOZILLA-FOUNDATION/COMMON_VOICE_7_0 - KA dataset. It achieves the following results on the evaluation set:

• Loss: 0.3666
• Wer: 0.4211

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: 0.0003
• train_batch_size: 32
• eval_batch_size: 32
• seed: 42
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 500
• num_epochs: 100.0
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
1.8805	5.95	500	0.7547	0.8438
1.2123	11.9	1000	0.4732	0.6542
1.0822	17.86	1500	0.4027	0.5778
0.9938	23.81	2000	0.3847	0.5524
0.9383	29.76	2500	0.3845	0.5204
0.8932	35.71	3000	0.3833	0.5297
0.8495	41.67	3500	0.3759	0.5036
0.8201	47.62	4000	0.3616	0.4859
0.7794	53.57	4500	0.3874	0.4938
0.735	59.52	5000	0.3748	0.4782
0.7082	65.48	5500	0.3615	0.4675
0.669	71.43	6000	0.3797	0.4601
0.6457	77.38	6500	0.3812	0.4515
0.6098	83.33	7000	0.3660	0.4343
0.5874	89.29	7500	0.3640	0.4257
0.5627	95.24	8000	0.3661	0.4239

Framework versions

• Transformers 4.16.0.dev0
• Pytorch 1.10.1+cu102
• Datasets 1.17.1.dev0
• Tokenizers 0.11.0