Adding an additional page lowers accuracy?

Help Document Understanding
,
1

Hi everyone,

My team is working on a modern Document Understanding project to extract signatures by retraining UiPath’s pretrained 1040 document model. Previously, we were using this model on only a single page from our documents, and got very good accuracy. Since retraining the model to look at multi page documents, its accuracy has gone down drastically. I was just wondering if anyone else had experienced something similar, and if so, how you resolved it.

Thanks!

2

Hey @aheath,

This behavior can occur because retraining the pretrained 1040 model with multi-page samples may dilute its page-level feature accuracy. It’s recommended to retrain using a balanced dataset with clear page annotations and ensure consistent labeling across all pages. Additionally, evaluate page-wise extraction performance and consider fine-tuning the model separately for signature detection to restore accuracy.

3

Hi Mir, thanks for your response. Can you clarify what you mean by ‘clear page annotations and consistent labeling’?