xls-r-300-cv17-polish

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: 0.3950
• Wer: 0.2789
• Cer: 0.0606

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
4.0552	1.6	100	4.2577	1.0	1.0
3.2887	3.2	200	3.2578	1.0	1.0
3.1481	4.8	300	3.1634	1.0	1.0
0.742	6.4	400	0.6905	0.6579	0.1603
0.4458	8.0	500	0.4687	0.4969	0.1169
0.2013	9.6	600	0.4327	0.4055	0.0929
0.2196	11.2	700	0.4180	0.4006	0.0903
0.1264	12.8	800	0.4360	0.3943	0.0898
0.1678	14.4	900	0.4157	0.3635	0.0818
0.1306	16.0	1000	0.3980	0.3667	0.0814
0.0471	17.6	1100	0.4206	0.3630	0.0828
0.1018	19.2	1200	0.3908	0.3522	0.0796
0.0637	20.8	1300	0.4277	0.3517	0.0785
0.1134	22.4	1400	0.4209	0.3373	0.0750
0.0709	24.0	1500	0.4255	0.3387	0.0766
0.046	25.6	1600	0.4301	0.3352	0.0746
0.065	27.2	1700	0.4087	0.3278	0.0724
0.0625	28.8	1800	0.4203	0.3454	0.0761
0.0344	30.4	1900	0.4317	0.3203	0.0714
0.0667	32.0	2000	0.4319	0.3258	0.0725
0.0305	33.6	2100	0.4260	0.3216	0.0716
0.04	35.2	2200	0.4172	0.3175	0.0697
0.0454	36.8	2300	0.4182	0.2996	0.0658
0.0273	38.4	2400	0.3966	0.2970	0.0654
0.0463	40.0	2500	0.4111	0.2926	0.0644
0.0321	41.6	2600	0.4094	0.2893	0.0633
0.0197	43.2	2700	0.3953	0.2846	0.0622
0.0306	44.8	2800	0.3980	0.2817	0.0613
0.0459	46.4	2900	0.3937	0.2807	0.0613
0.006	48.0	3000	0.3953	0.2780	0.0604
0.0329	49.6	3100	0.3950	0.2789	0.0606

Framework versions

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