metadata

license: cc-by-sa-4.0
tags:
  - generated_from_trainer
datasets:
  - te_dx_jp
model-index:
  - name: t5-base-TEDxJP-4front-1body-4rear
    results: []

t5-base-TEDxJP-4front-1body-4rear

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: 0.4419
Wer: 0.1696
Mer: 0.1638
Wil: 0.2494
Wip: 0.7506
Hits: 55898
Substitutions: 6290
Deletions: 2399
Insertions: 2262
Cer: 0.1328

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: 32
eval_batch_size: 32
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
0.5923	1.0	1457	0.4757	0.2180	0.2043	0.2941	0.7059	54841	6796	2950	4332	0.1931
0.5311	2.0	2914	0.4278	0.1817	0.1745	0.2611	0.7389	55507	6397	2683	2656	0.1467
0.4511	3.0	4371	0.4187	0.1715	0.1659	0.2529	0.7471	55668	6395	2524	2157	0.1322
0.3544	4.0	5828	0.4179	0.1700	0.1644	0.2498	0.7502	55795	6272	2520	2185	0.1319
0.3649	5.0	7285	0.4210	0.1700	0.1645	0.2506	0.7494	55799	6334	2454	2195	0.1319
0.3622	6.0	8742	0.4223	0.1706	0.1649	0.2503	0.7497	55823	6278	2486	2256	0.1358
0.3286	7.0	10199	0.4258	0.1692	0.1638	0.2492	0.7508	55807	6273	2507	2149	0.1325
0.3069	8.0	11656	0.4307	0.1697	0.1640	0.2493	0.7507	55861	6265	2461	2232	0.1329
0.2776	9.0	13113	0.4403	0.1697	0.1640	0.2497	0.7503	55883	6304	2400	2257	0.1328
0.3175	10.0	14570	0.4419	0.1696	0.1638	0.2494	0.7506	55898	6290	2399	2262	0.1328

Framework versions

Transformers 4.21.2
Pytorch 1.12.1+cu116
Datasets 2.4.0
Tokenizers 0.12.1