Ground Truth
To begin evaluating your model's performance, you must first configure your Ground Truth.
To get started, upload your first Ground Truth document. These documents are essential for evaluating your model's performance.
Each validated document you upload serves as a benchmark, which is used to compare your model's output and calculate a performance score. You can then use this score to optimize your model and achieve a higher performance.
Configure Ground Truth
To get started, go to the Performance tab and select the Ground Truth tab.
You need to upload and correctly label at least 3 documents to get started.
For the best results, try to use a diverse dataset that includes different scenarios for all your fields. The more varied your documents are, the more accurate your skill has the potential to become.
Keep in mind that each version of your skill has its own Ground Truth. Documents you add to a version will not be available in older ones, but they will be carried forward into newer versions once they have been reviewed.
Add documents
Simply press Add documents to upload multiple files at once. You can upload them in any of the following formats:
- .png, .jpg, .jpeg
- .tif, .tiff
Your files will be displayed in a list and will show a Syncing status until they are ready for review.
Once the files are ready, their status will update to For review, indicating they are ready for you to review them.
Review a document
To start your review, just press Review on a document line.
This will take you to a new view, where you'll see a preview of the document on the left and all of the labels extracted by the skill on the right. You can navigate between the pages of the current document, as well as see how many documents you have left to review and move between them.
From here, you also have the option to delete any documents that aren't relevant for your Ground Truth.
You can, and should, edit the extracted values. These corrections are crucial because they create a verified dataset that the model can use as a benchmark.
The number of corrections you have to make also helps you evaluate your skill's overall performance. The ultimate goal is to reach a point where no corrections are needed and the skill extracts everything perfectly.
If your skill fails to extract a value for a specific label that is present in the document, you can manually add it.
Just press Add label, select the missing label, and input the value yourself. Be aware that values added manually will not have any visual evidence, such as a bounding box, inside the document.
Next steps
Once you have configured your Ground Truth and reviewed at least 3 documents, you can start measuring your model's performance.