xlsr-128upper-sorbian

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: 0.7625
• Wer: 0.5044
• Cer: 0.1106

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: 100
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer	Cer
3.8489	3.9216	100	4.0479	1.0	1.0
3.1996	7.8431	200	3.2124	0.9804	0.9850
2.3527	11.7647	300	2.4026	1.0	0.6858
0.4424	15.6863	400	0.7917	0.7418	0.1910
0.2617	19.6078	500	0.7624	0.6804	0.1696
0.1421	23.5294	600	0.7839	0.6582	0.1579
0.097	27.4510	700	0.8322	0.6316	0.1495
0.0459	31.3725	800	0.8119	0.6171	0.1446
0.0668	35.2941	900	0.8534	0.6418	0.1535
0.0627	39.2157	1000	0.8256	0.6019	0.1397
0.0454	43.1373	1100	0.7747	0.5994	0.1363
0.04	47.0588	1200	0.8046	0.5810	0.1321
0.0563	50.9804	1300	0.7910	0.5797	0.1325
0.039	54.9020	1400	0.7370	0.5595	0.1265
0.0254	58.8235	1500	0.7395	0.5418	0.1188
0.0211	62.7451	1600	0.7582	0.5430	0.1209
0.0218	66.6667	1700	0.7123	0.5051	0.1121
0.0206	70.5882	1800	0.7912	0.5297	0.1165
0.0155	74.5098	1900	0.7671	0.5367	0.1183
0.0242	78.4314	2000	0.7926	0.5418	0.1170
0.0081	82.3529	2100	0.7817	0.5373	0.1221
0.0087	86.2745	2200	0.7989	0.5285	0.1165
0.0088	90.1961	2300	0.7523	0.5165	0.1141
0.0173	94.1176	2400	0.7646	0.5038	0.1108
0.0217	98.0392	2500	0.7625	0.5044	0.1106

Framework versions

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