metadata

license: cc-by-sa-4.0
tags:
  - generated_from_trainer
datasets:
  - te_dx_jp
model-index:
  - name: t5-base-TEDxJP-2front-1body-2rear
    results: []

t5-base-TEDxJP-2front-1body-2rear

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: 0.4473
Wer: 0.1735
Mer: 0.1675
Wil: 0.2549
Wip: 0.7451
Hits: 55674
Substitutions: 6443
Deletions: 2470
Insertions: 2291
Cer: 0.1360

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
0.5915	1.0	1457	0.4825	0.2230	0.2081	0.2976	0.7024	54814	6785	2988	4630	0.1980
0.5402	2.0	2914	0.4359	0.1855	0.1779	0.2664	0.7336	55358	6566	2663	2753	0.1503
0.4635	3.0	4371	0.4277	0.1762	0.1701	0.2584	0.7416	55538	6528	2521	2334	0.1368
0.3608	4.0	5828	0.4271	0.1723	0.1667	0.2529	0.7471	55621	6337	2629	2160	0.1342
0.3825	5.0	7285	0.4276	0.1724	0.1666	0.2527	0.7473	55669	6328	2590	2214	0.1343
0.3814	6.0	8742	0.4278	0.1727	0.1670	0.2537	0.7463	55613	6373	2601	2179	0.1360
0.3412	7.0	10199	0.4344	0.1724	0.1668	0.2539	0.7461	55631	6410	2546	2180	0.1352
0.3175	8.0	11656	0.4392	0.1724	0.1665	0.2534	0.7466	55713	6394	2480	2259	0.1362
0.2871	9.0	13113	0.4456	0.1732	0.1671	0.2538	0.7462	55748	6393	2446	2347	0.1360
0.3317	10.0	14570	0.4473	0.1735	0.1675	0.2549	0.7451	55674	6443	2470	2291	0.1360

Framework versions

Transformers 4.21.2
Pytorch 1.12.1+cu116
Datasets 2.4.0
Tokenizers 0.12.1