wav2vec2-mms-1b-CV17.0-training_set_variations

This model is a fine-tuned version of facebook/mms-1b-all on the common_voice_17_0 dataset. It achieves the following results on the evaluation set:

• Loss: 0.2335
• Wer: 0.3849
• Cer: 0.0627
• Bleu: 0.3848

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: 0.001
• 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_ratio: 0.15
• training_steps: 2000
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer	Cer	Bleu
12.5537	1.5625	50	3.9513	1.0006	0.9854	0.0
2.2034	3.125	100	0.3019	0.4137	0.0683	0.3510
0.226	4.6875	150	0.2305	0.3794	0.0623	0.3981
0.1904	6.25	200	0.2262	0.3776	0.0618	0.3988
0.1798	7.8125	250	0.2275	0.3760	0.0621	0.4040
0.1724	9.375	300	0.2399	0.4021	0.0659	0.3610
0.1791	10.9375	350	0.2310	0.3883	0.0635	0.3797
0.1678	12.5	400	0.2405	0.3961	0.0666	0.3722
0.1527	14.0625	450	0.2335	0.3849	0.0627	0.3848

Framework versions

• Transformers 4.44.2
• Pytorch 2.4.1+cu121
• Datasets 3.0.0
• Tokenizers 0.19.1