xlsr-mk-adap-ru

This model is a fine-tuned version of facebook/wav2vec2-xls-r-300m on the common_voice_17_0 dataset. It achieves the following results on the evaluation set:

• Loss: 0.5237
• Wer: 0.3740
• Cer: 0.0926

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: 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: 500
• num_epochs: 50
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer	Cer
3.0092	1.8868	100	2.9963	1.0	1.0
1.8956	3.7736	200	1.7089	0.9936	0.4894
0.5532	5.6604	300	0.5738	0.5799	0.1469
0.3567	7.5472	400	0.5090	0.5316	0.1336
0.374	9.4340	500	0.4753	0.4835	0.1219
0.2431	11.3208	600	0.5123	0.4868	0.1225
0.2573	13.2075	700	0.5502	0.4870	0.1240
0.1353	15.0943	800	0.5542	0.4836	0.1362
0.1978	16.9811	900	0.5199	0.4606	0.1205
0.1433	18.8679	1000	0.4968	0.4410	0.1131
0.0685	20.7547	1100	0.5464	0.4352	0.1125
0.151	22.6415	1200	0.5290	0.4329	0.1099
0.0845	24.5283	1300	0.5226	0.4159	0.1059
0.109	26.4151	1400	0.5479	0.4209	0.1085
0.0647	28.3019	1500	0.5292	0.4139	0.1058
0.0763	30.1887	1600	0.5182	0.4003	0.1022
0.091	32.0755	1700	0.5272	0.4033	0.1054
0.0707	33.9623	1800	0.5087	0.3956	0.1030
0.083	35.8491	1900	0.5098	0.3928	0.1017
0.1112	37.7358	2000	0.5140	0.3842	0.0937
0.112	39.6226	2100	0.5382	0.3961	0.0987
0.0472	41.5094	2200	0.5132	0.3805	0.0956
0.1136	43.3962	2300	0.5124	0.3831	0.0958
0.0547	45.2830	2400	0.5136	0.3827	0.0956
0.0844	47.1698	2500	0.5091	0.3713	0.0929
0.0643	49.0566	2600	0.5237	0.3740	0.0926

Framework versions

• Transformers 4.42.0.dev0
• Pytorch 2.3.1+cu121
• Datasets 2.19.2
• Tokenizers 0.19.1