wav2vec2-large-xls-r-300m-greek

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

• Loss: 0.6592
• Wer: 0.4564

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
3.0928	4.42	500	3.0804	1.0073
1.4505	8.85	1000	0.9038	0.7330
1.2207	13.27	1500	0.7375	0.6045
1.0695	17.7	2000	0.7119	0.5441
1.0104	22.12	2500	0.6069	0.5296
0.9299	26.55	3000	0.6168	0.5206
0.8588	30.97	3500	0.6382	0.5171
0.7942	35.4	4000	0.6048	0.4988
0.7808	39.82	4500	0.6730	0.5084
0.743	44.25	5000	0.6749	0.5012
0.6652	48.67	5500	0.6491	0.4735
0.6386	53.1	6000	0.6928	0.4954
0.5945	57.52	6500	0.6359	0.4798
0.5561	61.95	7000	0.6409	0.4799
0.5464	66.37	7500	0.6452	0.4691
0.5119	70.8	8000	0.6376	0.4657
0.474	75.22	8500	0.6541	0.4700
0.45	79.65	9000	0.6374	0.4571
0.4315	84.07	9500	0.6568	0.4625
0.3967	88.5	10000	0.6636	0.4605
0.3937	92.92	10500	0.6537	0.4597
0.3788	97.35	11000	0.6614	0.4589

Framework versions

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