hubert_new

This model is a fine-tuned version of facebook/hubert-large-ll60k on an unknown dataset. It achieves the following results on the evaluation set:

• Loss: 0.0652
• Wer: 0.0332
• Cer: 0.0321

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: 20
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer	Cer
32.501	1.0	26	30.2361	1.4946	1.4799
18.4587	2.0	52	10.7837	1.0	1.0
7.9821	3.0	78	4.0872	1.0	1.0
4.0414	4.0	104	3.4348	1.0	1.0
3.3351	5.0	130	3.2570	1.0	1.0
3.2641	6.0	156	3.2289	1.0	1.0
3.2492	7.0	182	3.1934	1.0	1.0
3.2815	8.0	208	3.1768	1.0	1.0
3.1516	9.0	234	3.1230	1.0	1.0
3.1305	10.0	260	3.0061	1.0	1.0
2.9981	11.0	286	2.8843	1.0	1.0
2.6928	12.0	312	2.4900	1.0	1.0
2.3977	13.0	338	2.0772	0.9470	0.9594
1.9738	14.0	364	1.5876	0.7353	0.7459
1.2899	15.0	390	0.9695	0.5152	0.5217
0.9936	16.0	416	0.5316	0.2774	0.2768
0.6417	17.0	442	0.2814	0.1217	0.1146
0.4256	18.0	468	0.1658	0.0645	0.0623
0.2974	19.0	494	0.1023	0.0435	0.0419
0.1767	20.0	520	0.0652	0.0332	0.0321

Framework versions

• Transformers 4.41.2
• Pytorch 2.3.0+cu121
• Datasets 2.20.0
• Tokenizers 0.19.1