aiboxlab/aied-unplugged-preview

AIED-Unplugged: preview competition dataset

Handwritten answers from Brazilian public schools: essays, working-out and bubble sheets, photographed on paper.

This is the preview sample for the AIED Preview Competition, a 2026 winner of the Tools Competition. The full collection of roughly 100,000 responses is released later.

Version 1.0.0. 961 items across three tracks.

Track	Config	Track page	train	validation	test
Automated Essay Scoring	essays	/tracks/aes	74	20	24
Mathematical Exam Diagnostic	math	/tracks/math-diagnostic	471	101	101
Answer Sheet Detection	answer_sheets	/tracks/answer-sheet	117	27	26

Loading

The dataset ships on two channels carrying the same ids, targets and splits, and storing the images differently. This is the Hugging Face one: the image bytes are embedded in data/<track>/<split>.parquet as Image() columns, so there are no image directories to download and load_dataset hands back PIL images.

from datasets import load_dataset

math = load_dataset("aiboxlab/aied-unplugged-preview", "math")
example = math["train"][0]
example["question_image"], example["equation_image"], example["diagnostic"]

schema/                       one JSON per track: fields, taxonomies, metrics
data/<track>/<split>.parquet  rows and image bytes, one file per split
samples/                      one sample submission per track

On Kaggle or from a plain download the same dataset is laid out the other way: the images are files under train/, validation/ and test/, and the Parquet in metadata/ points at them by relative path. Each image here keeps that path inside its column, so a row can be matched across the two channels.

from datasets import Image

paths = math["train"].cast_column("question_image", Image(decode=False))
paths[0]["question_image"]["path"]  # train/math/math-597-01-question.png

The SDK is optional and carries the competition's graders:

pip install aied-unplugged

from aied_unplugged import load_track, evaluate, validate

Source at github.com/aiboxlab/aied-unplugged.

Automated Essay Scoring

You get a photographed redação: one student's handwritten essay, the prompt it answers, and the three textos motivadores printed with it.

Predict five integers, one per ENEM competência, each from 0 to 200 in steps of 40. Ranked by quadratic weighted kappa averaged over the five. Exact-match rate and RMSE are reported alongside, both overall and per competence, so a model that holds up on four competences and fails on the fifth is visible as such.

Field	Competence	What it judges
competence_1	Command of the formal written norm	Spelling, agreement, punctuation, and register, judged against the formal written standard of Brazilian Portuguese.
competence_2	Grasp of the prompt and the essay form	Whether the response addresses the proposed theme and holds to the argumentative-expository form the exam asks for.
competence_3	Selection and organization of argument	Whether information, facts, and opinions are marshalled coherently in defence of a position.
competence_4	Cohesion	Use of linguistic devices that bind the argument across sentences and paragraphs.
competence_5	Intervention proposal	A proposal addressing the problem raised, respecting human rights, with agent, action, means, and purpose.

Mathematical Exam Diagnostic

You get two images: the printed exam question, and the working the student wrote while solving it.

Predict one label naming what happened in the response. Labels are compared exactly as written. Ranked by macro F1, with accuracy reported alongside.

Label	Annotated as	Meaning
Correct answer	Resposta Correta	The working is sound and reaches the right answer.
Multiple answers	Múltiplas respostas	More than one answer is given, with no indication of which one stands.
Correct direct answer	Resposta direta correta	The right answer, written down with no working shown.
Incorrect direct answer	Resposta direta incorreta	A wrong answer written down with no working shown.
Interpretation error	Erro de interpretação	The student solved a different problem from the one the question asks.
Wrong operation	Operação errada	The operation chosen is not the one the question calls for.
Operation setup error	Erro na montagem da operação	The right operation, assembled wrongly: terms misplaced, digits misaligned, or a value carried into the wrong position.
Partial answer	Resposta parcial	Part of what the question asks for is answered and the rest is left.
Incomplete calculation	Cálculo incompleto	The working starts correctly and stops before reaching an answer.
Addition error	Erro de soma	A mistake in adding, in otherwise sound working.
Subtraction error	Erro de subtração	A mistake in subtracting, in otherwise sound working.
Multiplication error	Erro de multiplicação	A mistake in multiplying, in otherwise sound working.
Division error	Erro de divisão	A mistake in dividing, in otherwise sound working.

The preview uses 11 of the 13. Multiplication error and Division error are declared for the full collection.

Answer Sheet Detection

You get a scan of a complete cartão-resposta. Sheets carry between 16 and 26 questions.

Predict what the student marked for every question. The target records the mark itself, whether or not it matches the gabarito. Ranked by cell accuracy, with sheet exact match and macro F1 reported alongside.

Value	Meaning
A	A single clear mark on option A.
B	A single clear mark on option B.
C	A single clear mark on option C.
D	A single clear mark on option D.
Blank	No mark on any option.
Multiple	More than one option marked.
Other	A mark that is present but not readable as any of the above, such as an erasure or a stray annotation.

Images

The images are as they were collected, without deskewing, rotation or colour correction, so expect skew, shadows, folds and the odd sideways page. Essay pages are cropped to the written answer for anonymization.

Splits

train and validation carry their targets. test rows carry their inputs with the target columns empty: those are the items the leaderboard grades.

Submitting

samples/ holds one file per track, covering the graded ids with placeholder predictions.

Track	File	Columns
Essays	sample_submission_aes.csv	essay_id, competence_1 to competence_5
Math	sample_submission_math_diagnostic.csv	equation_id, diagnostic
Answer sheets	sample_submission_answer_sheet.csv	sheet_id, answers (JSON array as a string)

Check a file at /validate before uploading, and read the rules at /rules.

Annotation

Every label is consolidated from several independent annotators. Essays were scored by two annotators with a third resolving disagreements, and the published score is the lower of the two where they agreed. Maths items took the majority of three. Answer sheets were transcribed once and checked by a reviewer.

This preview is not the final dataset

Annotations may be revised and the release may carry data this preview does not, such as bounding boxes. The tasks and the prediction formats are not going to move. The held-back targets are published once the competition closes.

Licence and citation

Creative Commons Attribution 4.0 International (CC-BY-4.0), full text in LICENSE. Share and adapt with credit, a licence link, and a note of any changes.

@misc{aibox_lab_2026,
    author       = { AIBox Lab and Rafael Ferreira Mello and Luiz Rodrigues and Ig Ibert Bittencourt and Moésio Filho and Adolfo Gabriel Domingos Mendonça and Angélica da Silva and Guilherme Pança Franco and Millena Sartori de Oliveira and Alexandre Aparecido da Silva and Gabriel Baruta Kocsis and Kauane dos Santos Vieira },
    title        = { AIED-Unplugged Preview Dataset },
    year         = 2026,
    url          = { https://huggingface.co/datasets/aiboxlab/aied-unplugged-preview },
    doi          = { 10.57967/hf/10774 },
    publisher    = { Hugging Face }
}

Collected by CESAR, UFRPE and AiBox Lab. The Tools Competition is organized by The Learning Agency as a program of Renaissance Philanthropy and fiscally sponsored by the Digital Harbor Foundation.