Hubert-noisy-cv-kakeiken-J_ver2

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: 0.0512
• Wer: 0.9988
• Cer: 1.0195

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: 64
• eval_batch_size: 8
• seed: 42
• gradient_accumulation_steps: 2
• total_train_batch_size: 128
• 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: 500
• num_epochs: 20.0
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer	Cer
No log	1.0	410	5.7169	1.0	1.1283
12.6371	2.0	820	3.2106	1.0	1.1284
3.3223	3.0	1230	2.8038	1.0	1.1284
2.8381	4.0	1640	2.0643	1.0	1.1284
1.6443	5.0	2050	0.6131	0.9994	1.0887
1.6443	6.0	2460	0.3250	0.9988	1.0469
0.6006	7.0	2870	0.2140	0.9988	1.0342
0.3876	8.0	3280	0.1747	0.9988	1.0352
0.2949	9.0	3690	0.1427	0.9990	1.0270
0.2541	10.0	4100	0.1015	0.9988	1.0233
0.2166	11.0	4510	0.0879	0.9988	1.0214
0.2166	12.0	4920	0.0746	0.9988	1.0212
0.1832	13.0	5330	0.0633	0.9990	1.0207
0.169	14.0	5740	0.0568	0.9988	1.0199
0.1544	15.0	6150	0.0555	0.9988	1.0187
0.1499	16.0	6560	0.0548	0.9988	1.0198
0.1499	17.0	6970	0.0542	0.9988	1.0199
0.1359	18.0	7380	0.0522	0.9988	1.0193
0.1352	19.0	7790	0.0512	0.9988	1.0197
0.1291	20.0	8200	0.0512	0.9988	1.0195

Framework versions

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