Hubert-noisy-cv-kakeiken-J_ver3

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: 0.0697
• Wer: 0.9988
• Cer: 1.0209

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: 12500
• num_epochs: 20.0
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer	Cer
No log	1.0	410	22.8613	1.0	1.1283
28.2328	2.0	820	10.8630	1.0	1.1284
11.2649	3.0	1230	9.3514	1.0	1.1284
9.2152	4.0	1640	7.5395	1.0	1.1284
7.0086	5.0	2050	5.6690	1.0	1.1284
7.0086	6.0	2460	4.2144	1.0	1.1283
4.9214	7.0	2870	3.4913	1.0	1.1283
3.6703	8.0	3280	3.0638	1.0	1.1284
3.0929	9.0	3690	2.6988	1.0	1.1283
2.6249	10.0	4100	2.1586	1.0	1.1284
1.9142	11.0	4510	1.3934	1.0	1.1055
1.9142	12.0	4920	0.9993	0.9999	1.1226
1.124	13.0	5330	0.5317	0.9997	1.0892
0.6627	14.0	5740	0.3859	0.9999	1.0674
0.4321	15.0	6150	0.2193	0.9990	1.0247
0.3399	16.0	6560	0.1816	0.9990	1.0275
0.3399	17.0	6970	0.1613	0.9988	1.0289
0.2755	18.0	7380	0.1239	0.9991	1.0290
0.2383	19.0	7790	0.0940	0.9988	1.0251
0.2038	20.0	8200	0.0699	0.9988	1.0209

Framework versions

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