Edit model card

t5-base-TEDxJP-1body-3context

This model is a fine-tuned version of sonoisa/t5-base-japanese on the te_dx_jp dataset. It achieves the following results on the evaluation set:

  • Loss: 0.4926
  • Wer: 0.1968
  • Mer: 0.1894
  • Wil: 0.2793
  • Wip: 0.7207
  • Hits: 55899
  • Substitutions: 6836
  • Deletions: 3636
  • Insertions: 2590
  • Cer: 0.1733

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: 0.0001
  • train_batch_size: 64
  • eval_batch_size: 8
  • seed: 42
  • optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
  • lr_scheduler_type: linear
  • lr_scheduler_warmup_ratio: 0.1
  • num_epochs: 10

Training results

Training Loss Epoch Step Validation Loss Wer Mer Wil Wip Hits Substitutions Deletions Insertions Cer
0.7082 1.0 746 0.5637 0.2626 0.2430 0.3355 0.6645 54301 7195 4875 5358 0.2552
0.6213 2.0 1492 0.5150 0.2068 0.1994 0.2899 0.7101 55107 6861 4403 2462 0.1866
0.5331 3.0 2238 0.4945 0.2038 0.1958 0.2858 0.7142 55551 6845 3975 2705 0.1816
0.5185 4.0 2984 0.4880 0.2003 0.1929 0.2831 0.7169 55639 6860 3872 2563 0.1779
0.4963 5.0 3730 0.4858 0.1988 0.1912 0.2810 0.7190 55837 6838 3696 2662 0.1772
0.4625 6.0 4476 0.4885 0.1964 0.1894 0.2799 0.7201 55785 6875 3711 2448 0.1720
0.4416 7.0 5222 0.4898 0.1962 0.1890 0.2788 0.7212 55870 6819 3682 2522 0.1726
0.4287 8.0 5968 0.4894 0.1968 0.1894 0.2790 0.7210 55889 6804 3678 2580 0.1743
0.4457 9.0 6714 0.4909 0.1964 0.1891 0.2792 0.7208 55919 6858 3594 2586 0.1739
0.4068 10.0 7460 0.4926 0.1968 0.1894 0.2793 0.7207 55899 6836 3636 2590 0.1733

Framework versions

  • Transformers 4.12.5
  • Pytorch 1.10.0+cu102
  • Datasets 1.15.1
  • Tokenizers 0.10.3
Downloads last month
14
Inference Examples
This model does not have enough activity to be deployed to Inference API (serverless) yet. Increase its social visibility and check back later, or deploy to Inference Endpoints (dedicated) instead.