trocr-small-printedkorean-deleteunusedchar

This model is a fine-tuned version of team-lucid/trocr-small-korean on an unknown dataset. It achieves the following results on the evaluation set:

• Loss: 2.1253
• Cer: 0.3884
• Wer: 0.4405
• Accuracy: 27.0903

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: 4e-05
• train_batch_size: 128
• eval_batch_size: 192
• seed: 42
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• num_epochs: 10
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Cer	Wer	Accuracy
3.5404	0.43	1000	3.5688	0.5737	0.6446	17.7258
3.365	0.85	2000	3.4212	0.5664	0.6546	18.0602
3.266	1.28	3000	3.3538	0.5637	0.6815	18.7291
3.2922	1.71	4000	3.3051	0.5613	0.7161	18.0602
3.1287	2.13	5000	3.2649	0.5577	0.7164	18.7291
3.0906	2.56	6000	3.2016	0.5572	0.6139	18.3946
3.0015	2.99	7000	3.0851	0.5530	0.6018	18.3946
2.7418	3.42	8000	2.8927	0.5264	0.6052	19.3980
2.5107	3.84	9000	2.6466	0.4821	0.5881	19.3980
2.3091	4.27	10000	2.4865	0.4545	0.5651	20.4013
2.1395	4.7	11000	2.3749	0.4336	0.5323	21.7391
2.0453	5.12	12000	2.3070	0.4222	0.5157	21.7391
2.0421	5.55	13000	2.2667	0.4143	0.5001	24.0803
2.0085	5.98	14000	2.2201	0.4060	0.4809	24.0803
1.8766	6.4	15000	2.1988	0.4026	0.4766	23.7458
1.8987	6.83	16000	2.1788	0.3969	0.4658	24.4147
1.8921	7.26	17000	2.1702	0.3955	0.4581	24.4147
1.8253	7.69	18000	2.1554	0.3938	0.4513	26.7559
1.8281	8.11	19000	2.1481	0.3925	0.4534	26.0870
1.8751	8.54	20000	2.1419	0.3905	0.4500	26.0870
1.8226	8.97	21000	2.1317	0.3905	0.4439	26.7559
1.813	9.39	22000	2.1285	0.3895	0.4407	26.0870
1.7472	9.82	23000	2.1253	0.3884	0.4405	27.0903

Framework versions

• Transformers 4.28.0
• Pytorch 1.13.1+cu116
• Datasets 2.14.4
• Tokenizers 0.13.3