Datasets:

colearnlab
/

misconception-detection

pretty_name: Misconception Detection in K-12 Student Workings
license: other
license_name: colearn-competition-use-only
license_link: >-
  https://www.kaggle.com/competitions/misconception-detection-in-k-12-student-workings/rules
language:
  - id
  - en
task_categories:
  - image-classification
  - image-to-text
tags:
  - education
  - mathematics
  - handwriting
  - misconceptions
  - k-12
  - kaggle-competition
size_categories:
  - 10K<n<100K
extra_gated_heading: Access to the CoLearn competition data
extra_gated_prompt: >-
  This dataset is provided by CoLearn (IQ Edu Pte Ltd) solely for participation
  in the Kaggle competition "Misconception Detection in K-12 Student Workings"
  (https://www.kaggle.com/competitions/misconception-detection-in-k-12-student-workings)
  and for discussion on its Kaggle forums. It is a mirror of the competition
  data, offered because some participants found the Kaggle download slow. By
  requesting access you confirm that you have joined the competition and
  accepted its rules, and you agree that you will not redistribute the data, and
  will not use it for any other purpose, including any commercial purpose,
  without CoLearn's prior written agreement. The same applies to models trained
  on it. The data contains images of children's schoolwork: you must not attempt
  to identify any student. To discuss any other use, contact
  engineering@colearnlab.com.
extra_gated_fields:
  Kaggle username: text
  I have joined the competition and accepted its rules on Kaggle: checkbox
  I will use this data only for the competition and will not redistribute it: checkbox

Misconception Detection in K-12 Student Workings

Mirror of the competition data for the Kaggle competition Misconception Detection in K-12 Student Workings, hosted and sponsored by CoLearn (IQ Edu Pte Ltd, Singapore).

This repository exists only because some participants found the Kaggle download slow. The files are identical to data version 2 on Kaggle (topic names in English). Submissions, the leaderboard, discussion and the rules all remain on Kaggle; nothing here is accepted or scored.

Terms of use

The data may be used only to take part in the competition and to discuss it on the Kaggle forums, under the competition rules. You may not redistribute it or use it for any other purpose, including any commercial purpose, without CoLearn's prior written agreement; the same applies to models trained on it. To discuss other uses, contact engineering@colearnlab.com.

The data contains images of children's schoolwork. Every image was screened for personally identifying information, and images showing a name, face, signature, school identifier or contact detail were removed. No student names, account details or source URLs are included. You must not attempt to identify any student.

The task

Each row is one trace: a single student's handwritten attempt at a single practice question, photographed with a phone. Students are in Indonesian grades 4–12 (ages roughly 9–18). The questions and the handwriting are in Bahasa Indonesia; question texts contain LaTeX for mathematical notation. The 499 questions span 61 curriculum topics: whole-number and integer operations, fractions and decimals, LCM/GCD, powers and surds, algebra, functions, measurement, geometry, statistics and probability.

What you are predicting: root_code, the taxonomy code of the root misconception in the working, or NONE if the working has no major misconception. The root is the misconception the rest of the working follows from, not necessarily the first error written on the page. A root is always a major misconception (one that changes the reasoning or the result). NONE means the working is fully correct or carries only minor slips; about a third of rows are NONE.

Submissions are evaluated on Kaggle with MAP@3: up to three ranked codes per row, scoring 1, 1/2 or 1/3 by the rank of the correct code.

Files

File	Contents
`train.csv`	1,500 labelled rows
`test.csv`	3,498 rows without labels (public/private split hidden)
`sample_submission.csv`	a submission in the correct format, `NONE` for every row
`taxonomy_v1.csv`	the 75 misconception codes that occur in this dataset, with name, description, nature (conceptual / procedural / careless / meta), default severity, typical grade band and section
`unlabelled.csv`	50,000 additional student workings without labels, for semi-supervised use
`images/train/`, `images/test/`, `images/unlabelled/`	one JPEG per row, named by `image_file`
`images/questions/`	119 question diagrams, referenced by `question_image_file`

Columns

In train.csv, test.csv and unlabelled.csv:

trace_id: unique identifier of the trace
question_id: identifier of the practice question; the same question appears in train and test, answered by different students
grade: school grade, 4–12
topic_name: curriculum topic of the question, in English
question_text: the question, with LaTeX for mathematical notation
question_image_file: filename of the question's diagram in images/questions/, or empty
correct_answer: the answer key, as given to the student
image_file: filename of the photo of the student's working

In train.csv only:

root_code: the target; taxonomy code of the root misconception, or NONE
transcription_steps: line-by-line transcription of the handwriting (L1 | ..., L2 | ...); student text verbatim, annotator notes in [brackets]
final_answer_correct: whether the student's final answer matches the key
clean: true if the working has no finding at all, major or minor
misconceptions: all findings as JSON (step, span, category, severity, root flag, explanation in English); auxiliary supervision, only root_code is scored

Test-time inputs. Test rows carry the image and the question context only; there is no transcription. Reading the handwriting is part of the task.

How the labels were made

Every trace was annotated in three steps: the working was transcribed line by line; the question was solved independently of the answer key; each error was recorded as a finding with a taxonomy code and a severity, and one major finding was marked as the root. The annotations follow the protocol and the taxonomy defined by CoLearn's teachers, and the labelled set was reviewed by CoLearn before release. Naming a misconception involves judgement, and some codes have close neighbours on which even trained annotators sometimes differ; expect some label noise in both train and test.

Splits and statistics

Splits are disjoint by student and stratified on root code and grade. Questions are shared across splits.

Split	Rows
Train	1,500
Public leaderboard	1,000
Private leaderboard	2,498

The most common categories (correct work, ANS-01, CAL-01, WOP-01, FRM-01, OTH-00) were capped at about 190 rows each, so the class distribution is flatter than in the raw data.
13 codes have a single training example; every code that is an answer in the test set has at least one training example.
Images are JPEG, typically 1024×1024 px; about 460 MB labelled and 4.7 GB unlabelled.

The unlabelled pool holds 50,000 further workings from the same practice app, with no labels and no transcriptions. No student in it appears in the test set, and nothing in it is a labelled row.

Citation

CoLearn. Misconception Detection in K-12 Student Workings. https://kaggle.com/competitions/misconception-detection-in-k-12-student-workings, 2026. Kaggle.

CoLearn Lab (2026). Teaching open-weight models to spot where students actually go wrong.