shpotes/xls-r-eus

This model is a fine-tuned version of facebook/wav2vec2-xls-r-300m on the MOZILLA-FOUNDATION/COMMON_VOICE_8_0 - EU dataset. It achieves the following results on the evaluation set:

• Loss: 0.2278
• Wer: 0.1787

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: 72
• eval_batch_size: 72
• seed: 42
• gradient_accumulation_steps: 2
• total_train_batch_size: 144
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine
• lr_scheduler_warmup_steps: 500
• num_epochs: 100.0
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
0.2548	4.24	500	0.2470	0.3663
0.1435	8.47	1000	0.2000	0.2791
0.1158	12.71	1500	0.2030	0.2652
0.1094	16.95	2000	0.2096	0.2605
0.1004	21.19	2500	0.2150	0.2477
0.0945	25.42	3000	0.2072	0.2369
0.0844	29.66	3500	0.1981	0.2328
0.0877	33.89	4000	0.2041	0.2425
0.0741	38.14	4500	0.2353	0.2421
0.0676	42.37	5000	0.2092	0.2213
0.0623	46.61	5500	0.2217	0.2250
0.0574	50.84	6000	0.2152	0.2179
0.0583	55.08	6500	0.2207	0.2186
0.0488	59.32	7000	0.2225	0.2159
0.0456	63.56	7500	0.2293	0.2031
0.041	67.79	8000	0.2277	0.2013
0.0379	72.03	8500	0.2287	0.1991
0.0381	76.27	9000	0.2233	0.1954
0.0308	80.51	9500	0.2195	0.1835
0.0291	84.74	10000	0.2266	0.1825
0.0266	88.98	10500	0.2285	0.1801
0.0266	93.22	11000	0.2292	0.1801
0.0262	97.46	11500	0.2278	0.1788

Framework versions

• Transformers 4.16.0.dev0
• Pytorch 1.10.1+cu102
• Datasets 1.18.4.dev0
• Tokenizers 0.11.0