Hubert-kakeiken-W-closed

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

  • Loss: 0.0273
  • Wer: 0.9988
  • Cer: 1.0164

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: 40.0
  • mixed_precision_training: Native AMP

Training results

Training Loss Epoch Step Validation Loss Wer Cer
28.44 1.0 824 10.8607 1.0 1.1284
9.2184 2.0 1648 7.5106 1.0 1.1284
7.0297 3.0 2472 4.1902 1.0 1.1284
3.6874 4.0 3296 3.1196 1.0 1.1284
2.7259 5.0 4120 2.2997 1.0 1.1284
2.1431 6.0 4944 1.0496 0.9997 1.0314
0.8891 7.0 5768 0.6380 0.9999 1.0174
0.4891 8.0 6592 0.2589 0.9991 1.0305
0.3675 9.0 7416 0.6099 0.9994 1.0932
0.2744 10.0 8240 0.1262 0.9990 1.0239
0.2278 11.0 9064 0.1107 0.9990 1.0234
0.2148 12.0 9888 0.0672 0.9990 1.0261
0.1927 13.0 10712 0.0592 0.9990 1.0219
0.1723 14.0 11536 0.0800 0.9988 1.0228
0.1725 15.0 12360 0.0518 0.9990 1.0211
0.1628 16.0 13184 0.1359 0.9988 1.0142
0.1567 17.0 14008 0.0444 0.9990 1.0196
0.1436 18.0 14832 0.0421 0.9988 1.0193
0.1351 19.0 15656 0.0375 0.9988 1.0173
0.1454 20.0 16480 0.0304 0.9988 1.0186
0.1252 21.0 17304 0.0568 0.9988 1.0237
0.1233 22.0 18128 0.0291 0.9988 1.0176
0.1179 23.0 18952 0.0271 0.9988 1.0167
0.1108 24.0 19776 0.0279 0.9988 1.0178
0.1031 25.0 20600 0.0351 0.9990 1.0183
0.1006 26.0 21424 0.0441 0.9990 1.0173
0.0946 27.0 22248 0.0306 0.9988 1.0170
0.09 28.0 23072 0.0302 0.9988 1.0177
0.0813 29.0 23896 0.0349 0.9988 1.0178
0.0806 30.0 24720 0.0369 0.9988 1.0179
0.0763 31.0 25544 0.0434 0.9990 1.0165
0.075 32.0 26368 0.0308 0.9988 1.0160
0.0708 33.0 27192 0.0308 0.9988 1.0167
0.0668 34.0 28016 0.0298 0.9988 1.0166
0.0639 35.0 28840 0.0272 0.9988 1.0164
0.0628 36.0 29664 0.0265 0.9988 1.0161
0.0628 37.0 30488 0.0267 0.9988 1.0163
0.0586 38.0 31312 0.0263 0.9988 1.0162
0.058 39.0 32136 0.0280 0.9988 1.0164
0.0588 39.9520 32920 0.0269 0.9988 1.0164

Framework versions

  • Transformers 4.48.0
  • Pytorch 2.5.1+cu124
  • Datasets 3.1.0
  • Tokenizers 0.21.0
Downloads last month
68
Safetensors
Model size
94.4M params
Tensor type
F32
·
Inference Providers NEW
This model is not currently available via any of the supported Inference Providers.

Model tree for utakumi/Hubert-kakeiken-W-closed

Finetuned
(47)
this model