wav2vec2-large-xls-r-300m-grain

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

• Loss: 0.1510
• Wer: 0.0762

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: 8
• 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
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
4.1496	2.5	400	0.7656	0.8096
0.2914	5.0	800	0.3202	0.3544
0.1152	7.5	1200	0.2666	0.2894
0.0722	10.0	1600	0.2834	0.2458
0.0528	12.5	2000	0.2475	0.2159
0.0423	15.0	2400	0.2430	0.1971
0.0334	17.5	2800	0.2250	0.1925
0.0288	20.0	3200	0.2119	0.1779
0.0253	22.5	3600	0.2226	0.1711
0.0214	25.0	4000	0.2224	0.1685
0.0217	27.5	4400	0.2098	0.1516
0.0182	30.0	4800	0.2153	0.1716
0.0173	32.5	5200	0.1925	0.1451
0.0137	35.0	5600	0.2241	0.1469
0.0118	37.5	6000	0.2013	0.1515
0.0133	40.0	6400	0.1990	0.1332
0.0125	42.5	6800	0.2146	0.1502
0.0103	45.0	7200	0.2191	0.1317
0.0089	47.5	7600	0.1869	0.1246
0.0091	50.0	8000	0.1734	0.1251
0.008	52.5	8400	0.2008	0.1290
0.0071	55.0	8800	0.1828	0.1260
0.0064	57.5	9200	0.1689	0.1081
0.0061	60.0	9600	0.1676	0.1111
0.0051	62.5	10000	0.1707	0.1048
0.0056	65.0	10400	0.1741	0.1131
0.0046	67.5	10800	0.1836	0.1034
0.0036	70.0	11200	0.1655	0.0966
0.0037	72.5	11600	0.1734	0.1047
0.003	75.0	12000	0.1718	0.0975
0.0032	77.5	12400	0.1598	0.0986
0.0023	80.0	12800	0.1640	0.0966
0.0019	82.5	13200	0.1701	0.0862
0.0015	85.0	13600	0.1643	0.0854
0.0016	87.5	14000	0.1470	0.0823
0.0014	90.0	14400	0.1589	0.0838
0.0011	92.5	14800	0.1610	0.0834
0.0013	95.0	15200	0.1457	0.0788
0.001	97.5	15600	0.1537	0.0762
0.001	100.0	16000	0.1510	0.0762

Framework versions

• Transformers 4.42.3
• Pytorch 2.2.0+cu121
• Datasets 2.20.0
• Tokenizers 0.19.1