KispiOcr Predict Datasets Review labels Models Jobs API docs

Datasets

NameImagesLabelsLabelled images Images without labelsLabels without images

Upload

Folders doc_<DOCID>/ with page_<PAGEINDEX>.png and its labels page_<PAGEINDEX>.json for each page, e.g. doc_20260917105020860/page_001.png + page_001.json. Field-level labels ("boxes": null) are split into boxes on the image. Pages already in the dataset are replaced.

Shows the dataset's labelled boxes grouped by label, to find and fix wrong labels. The first time, preparing the boxes takes about a minute per 150 pages.

Label

Field

Pages (disagreements)

Selected box

Open a dataset.

Legend

7 model prediction (top left)

7 model disagrees with the label

7 label (bottom right)

7 label changed in this session

Keys

0–9, A–Z set the label

Space empty box

? leave the box out of training

Enter accept the model's prediction

←→↑↓ move

Esc close the page view

Each change is saved to the page's label file at once; the next training uses it.

A page where most boxes disagree usually has misaligned boxes: the boxes are from other fields. Don't relabel those; fix or delete the page's labels instead.

blue: this field with its labels · orange: the selected box · grey: other labelled boxes · red: left out (?)
page

Models

NameModifiedSizeMetrics

Upload

Train

more

Jobs

JobKindStatusParametersCreated Finished