musaabkhan/trocr-ur-v2

TrOCR for Urdu

This model is a fine-tuned version of mohammadalihumayun/trocr-ur-v2r on an unknown dataset. It achieves the following results on the evaluation set:

• Loss: 1.1851
• Cer: 0.5466

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: 5e-05
• train_batch_size: 8
• eval_batch_size: 8
• seed: 42
• optimizer: Use OptimizerNames.ADAMW_TORCH with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
• lr_scheduler_type: linear
• num_epochs: 2
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Cer
1.4219	0.1518	1000	2.5109	0.6301
1.2491	0.3036	2000	2.3298	0.5844
1.636	0.4554	3000	2.1089	0.5860
1.5884	0.6072	4000	1.9481	0.6081
1.4926	0.7590	5000	1.8332	0.5751
1.4351	0.9107	6000	1.6727	0.5776
1.2219	1.0625	7000	1.5796	0.5832
0.9786	1.2143	8000	1.4937	0.5739
0.9635	1.3661	9000	1.4292	0.5414
0.909	1.5179	10000	1.3259	0.5747
0.8943	1.6697	11000	1.2704	0.5794
0.8486	1.8215	12000	1.2211	0.5445
0.8385	1.9733	13000	1.1851	0.5466

Framework versions

• Transformers 4.51.3
• Pytorch 2.5.1+cu124
• Datasets 3.5.0
• Tokenizers 0.21.0