Hubert-noisy-cv-kakeiken-J_ver4

This model is a fine-tuned version of rinna/japanese-hubert-base on the ORIGINAL_NOISY_KAKEIKEN_W - JA dataset. It achieves the following results on the evaluation set:

• Loss: 0.0362
• Wer: 0.9988
• Cer: 1.0182

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: 3e-05
• train_batch_size: 32
• eval_batch_size: 8
• seed: 42
• gradient_accumulation_steps: 2
• total_train_batch_size: 64
• optimizer: Use adamw_torch with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
• lr_scheduler_type: cosine
• lr_scheduler_warmup_steps: 12500
• num_epochs: 20.0
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer	Cer
28.345	1.0	820	10.8647	1.0	1.1283
9.1788	2.0	1640	7.5464	1.0	1.1284
6.9973	3.0	2460	4.2194	1.0	1.1284
3.6678	4.0	3280	3.0366	1.0	1.1284
2.7018	5.0	4100	2.3639	1.0	1.1284
2.2246	6.0	4920	1.1432	1.0	1.1433
0.8963	7.0	5740	0.5476	0.9997	1.1185
0.4454	8.0	6560	0.2377	0.9991	1.0352
0.3477	9.0	7380	0.1761	0.9990	1.0335
0.2579	10.0	8200	0.1430	0.9990	1.0378
0.2112	11.0	9020	0.1305	0.9990	1.0314
0.1974	12.0	9840	0.0652	0.9990	1.0225
0.1813	13.0	10660	0.1686	0.9990	1.0385
0.179	14.0	11480	0.0616	0.9988	1.0210
0.1692	15.0	12300	0.0616	0.9990	1.0236
0.1632	16.0	13120	0.0531	0.9990	1.0199
0.1485	17.0	13940	0.0423	0.9988	1.0186
0.124	18.0	14760	0.0697	0.9988	1.0180
0.1011	19.0	15580	0.0430	0.9988	1.0186
0.0854	19.9762	16380	0.0377	0.9988	1.0181

Framework versions

• Transformers 4.48.0
• Pytorch 2.5.1+cu124
• Datasets 3.1.0
• Tokenizers 0.21.0