metadata

license: cc-by-sa-4.0
tags:
  - generated_from_trainer
datasets:
  - te_dx_jp
model-index:
  - name: t5-base-TEDxJP-8front-1body-8rear
    results: []

t5-base-TEDxJP-8front-1body-8rear

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: 0.4357
Wer: 0.1706
Mer: 0.1649
Wil: 0.2512
Wip: 0.7488
Hits: 55798
Substitutions: 6352
Deletions: 2437
Insertions: 2230
Cer: 0.1344

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: 30
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
0.585	1.0	1457	0.4589	0.2109	0.1988	0.2866	0.7134	54910	6576	3101	3947	0.1850
0.5098	2.0	2914	0.4190	0.1814	0.1744	0.2618	0.7382	55486	6473	2628	2617	0.1435
0.4648	3.0	4371	0.4108	0.1728	0.1669	0.2526	0.7474	55696	6298	2593	2267	0.1366
0.4075	4.0	5828	0.4071	0.1728	0.1670	0.2534	0.7466	55649	6356	2582	2220	0.1364
0.3904	5.0	7285	0.4118	0.1709	0.1652	0.2514	0.7486	55753	6334	2500	2203	0.1343
0.343	6.0	8742	0.4131	0.1701	0.1647	0.2509	0.7491	55741	6334	2512	2142	0.1338
0.2981	7.0	10199	0.4211	0.1701	0.1645	0.2503	0.7497	55788	6302	2497	2187	0.1345
0.2663	8.0	11656	0.4291	0.1698	0.1642	0.2503	0.7497	55851	6338	2398	2234	0.1339
0.2938	9.0	13113	0.4317	0.1699	0.1642	0.2503	0.7497	55833	6329	2425	2218	0.1340
0.2692	10.0	14570	0.4357	0.1706	0.1649	0.2512	0.7488	55798	6352	2437	2230	0.1344

Framework versions

Transformers 4.21.2
Pytorch 1.12.1+cu116
Datasets 2.4.0
Tokenizers 0.12.1