About
AIED-Unplugged collects handwritten responses from students in Brazilian public schools, most of them in low-income regions. It covers multiple-choice answer sheets, Portuguese essays, and mathematical working, and it exists to support research on automated scoring, feedback generation, and handwriting recognition where the assessment never touched a screen.
The complete collection will hold roughly 100,000 responses and is released at a later date. This competition runs on a preview: a smaller sample of the collection, published ahead of it. The full release bases its prediction formats on the ones declared here, and may refine them and add data this preview does not carry.
The dataset is a 2026 winner of the Tools Competition, organized by The Learning Agency as a program of Renaissance Philanthropy and fiscally sponsored by the Digital Harbor Foundation.
How the preview was annotated
Every label in the preview is a consolidated one. Several people saw each item independently, and what is published is the result of a rule for combining what they said. The rule is different on each track.
Essays
Two annotators scored each essay across the five competences, without seeing each other’s scores. Where the two did not diverge under ENEM rules, the published score is the lower of them. Where they diverged, a third annotator scored the essay, and the published score is that annotator’s.
Taking the lower rather than the mean keeps every published score on the grid a competence is actually scored on: 0, 40, 80, 120, 160, 200. The mean of 120 and 160 is 140, which is not a score a rater can give.
Mathematics
Three annotators labelled each item independently, and the published label is their majority. Where there was no majority, because three annotators reached three different labels, a fourth decided it.
Answer sheets
One annotator transcribed each sheet and a reviewer checked that transcription against the scan. The published transcription is the reviewer’s.
How the sample was drawn
Two properties of the preview sample follow from how the items were drawn. They leave the tasks and the prediction formats unchanged and limit how far a local score generalizes.
On the essay track, the validation split is drawn from a single prompt, and its scores skew on competence 5. A score computed on it measures one theme and generalizes less than the split’s size suggests.
On the mathematics and answer sheet tracks, items are not grouped by the student or document they came from, so the same handwriting can appear in more than one split. A model can learn to recognise a hand it saw in training and score on it. The sample was stratified by label instead, and the grouping keys are kept for the full release to split by source.
This preview is not the final dataset.
The full release may differ from this preview. Annotations may be revised where further review finds them wrong, the prediction formats may change slightly, and the release may carry data this preview does not, such as bounding boxes locating the region a label refers to and finer-grained metadata about each response.
Each such change is made to improve the data. The tasks stay the same, and a pipeline built on the preview carries over with adjustments.
Each track’s test split is published with its targets once the competition closes, on the dataset pages that already host the preview. That completes the preview sample. The full collection follows on its own schedule.
Sponsor
Main sponsor of the AIED Preview Competition.
Institutions
Head researchers and organizers
Rafael Ferreira MelloAssociate Professor and Senior ResearcherCESAR, UFRPE, AiBox Lab
Ig Ibert BittencourtFull ProfessorUFAL, Harvard University
Luiz RodriguesAdjunct ProfessorUTFPR, UFAL
Team
How to cite
Work drawing on the dataset or these benchmarks should cite the dataset.
@dataset{aied_unplugged_2026,
title = {AIED-Unplugged: Handwritten Student Responses from Brazilian Public Schools},
author = {{CESAR} and {UFRPE} and {AiBox Lab}},
year = {2026},
publisher = {Tools Competition},
url = {https://tools-competition.org/winner/aied/}
}License
The preview dataset is published under Creative Commons Attribution 4.0 International (CC BY 4.0). It may be shared and adapted, including commercially, with credit to the source, a link to the licence, and a note of any changes made.
Those terms were chosen for this preview release. They are not a commitment for the full collection, which may be published under different ones.