Hubert-noisy-cv-kakeiken-F

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: 6.2254
• Wer: 1.0
• Cer: 1.2345

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
0.1936	1.0	2732	2.4481	1.0	1.4689
0.0292	2.0	5464	4.7625	1.0	1.5319
0.0318	3.0	8196	4.2177	1.0	1.5485
0.0349	4.0	10928	5.2287	1.0	1.3069
0.0311	5.0	13660	7.3641	1.0	1.7554
0.0615	6.0	16392	5.4349	1.0	1.2858
0.0209	7.0	19124	6.7846	1.0	1.6721
0.021	8.0	21856	5.4731	1.0	1.3740
0.0192	9.0	24588	4.2649	1.0	1.3813
0.013	10.0	27320	5.8057	1.0	1.4152
0.0118	11.0	30052	6.2095	1.0	1.2175
0.007	12.0	32784	6.1471	1.0	1.3806
0.0073	13.0	35516	6.6287	1.0	1.3807
0.0064	14.0	38248	5.9500	1.0	1.2425
0.002	15.0	40980	7.6846	1.0	1.4058
0.0048	16.0	43712	6.0112	1.0	1.1907
0.0023	17.0	46444	5.8869	1.0	1.1885
0.0031	18.0	49176	5.9530	1.0	1.2386
0.0021	19.0	51908	6.2662	1.0	1.2392
0.0012	19.9929	54620	6.2391	1.0	1.2330

Framework versions

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