Hubert-noisy-cv-kakeiken-D

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

• Loss: 0.0053
• Wer: 0.9994
• Cer: 0.0877

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: 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
0.1608	1.0	2732	0.0715	0.9996	0.1000
0.0364	2.0	5464	0.0245	0.9996	0.0928
0.0283	3.0	8196	0.0276	0.9996	0.0924
0.0277	4.0	10928	0.0863	0.9999	0.1027
0.0335	5.0	13660	0.0389	0.9996	0.0953
0.0241	6.0	16392	0.0589	0.9999	0.0996
0.0251	7.0	19124	0.0246	0.9998	0.0919
0.0176	8.0	21856	0.0241	0.9997	0.0916
0.0171	9.0	24588	0.0167	0.9995	0.0907
0.0135	10.0	27320	0.0166	0.9996	0.0911
0.0173	11.0	30052	0.0156	0.9996	0.0899
0.0105	12.0	32784	0.0095	0.9996	0.0892
0.0057	13.0	35516	0.0093	0.9994	0.0890
0.0072	14.0	38248	0.0091	0.9994	0.0886
0.0016	15.0	40980	0.0088	0.9995	0.0884
0.0033	16.0	43712	0.0064	0.9995	0.0880
0.003	17.0	46444	0.0059	0.9995	0.0879
0.0017	18.0	49176	0.0054	0.9994	0.0878
0.0026	19.0	51908	0.0053	0.9994	0.0877
0.0009	19.9929	54620	0.0052	0.9995	0.0877

Framework versions

• Transformers 4.47.0.dev0
• Pytorch 2.5.1+cu124
• Datasets 3.1.0
• Tokenizers 0.20.3