jq/mms-1b-lug-eng

stt

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

• Loss: 0.1619
• Wer Lug: 0.161
• Wer Eng: 0.096
• Wer Mean: 0.129

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: 100
• training_steps: 5000
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer Lug	Wer Eng	Wer Mean
0.1876	0.1	500	0.1711	0.183	0.106	0.144
0.1971	0.2	1000	0.1702	0.172	0.106	0.139
0.1898	0.3	1500	0.1687	0.168	0.108	0.138
0.1903	0.4	2000	0.1686	0.165	0.103	0.134
0.1888	0.5	2500	0.1663	0.165	0.096	0.131
0.1908	1.1	3000	0.1637	0.16	0.095	0.127
0.1792	1.2	3500	0.1642	0.157	0.094	0.125
0.1963	1.3	4000	0.1625	0.158	0.095	0.127
0.184	1.4	4500	0.1623	0.158	0.094	0.126
0.1888	1.5	5000	0.1619	0.161	0.096	0.129

Framework versions

• Transformers 4.38.2
• Pytorch 2.2.1+cu121
• Datasets 2.18.0
• Tokenizers 0.15.2