xls-r-300-cv17-upper-sorbian-adap-pl

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: 1.0564
• Wer: 0.7247
• Cer: 0.1754

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: 0.0001
• 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
3.5302	3.9216	100	3.5256	1.0	1.0
3.2181	7.8431	200	3.2314	1.0	1.0
1.5479	11.7647	300	1.6991	0.9797	0.3943
0.3971	15.6863	400	0.9388	0.8582	0.2274
0.2782	19.6078	500	0.9310	0.8291	0.2203
0.1388	23.5294	600	0.9292	0.8	0.2045
0.1438	27.4510	700	0.9533	0.8006	0.2011
0.0815	31.3725	800	0.9446	0.7816	0.1975
0.0873	35.2941	900	0.9855	0.7728	0.1913
0.1213	39.2157	1000	0.9705	0.7652	0.1955
0.0589	43.1373	1100	0.9832	0.7614	0.1876
0.0865	47.0588	1200	1.0001	0.7582	0.1875
0.0762	50.9804	1300	1.0280	0.7538	0.1854
0.0564	54.9020	1400	0.9799	0.7468	0.1820
0.0607	58.8235	1500	1.0192	0.7443	0.1793
0.0729	62.7451	1600	1.0057	0.7424	0.1762
0.0518	66.6667	1700	1.0240	0.7437	0.1765
0.059	70.5882	1800	1.0379	0.7278	0.1759
0.031	74.5098	1900	1.0444	0.7152	0.1718
0.051	78.4314	2000	1.0530	0.7335	0.1773
0.0539	82.3529	2100	1.0402	0.7241	0.1773
0.0399	86.2745	2200	1.0495	0.7177	0.1744
0.06	90.1961	2300	1.0674	0.7222	0.1764
0.0459	94.1176	2400	1.0576	0.7222	0.1747
0.0614	98.0392	2500	1.0564	0.7247	0.1754

Framework versions

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