This dataset contains anonymized interaction data from primary and secondary school students in France and Côte d’Ivoire, including minors. It is released strictly for non-commercial research purposes. Users must not attempt to re-identify any individual in the dataset, and must report any personally identifiable information (PII) encountered to the dataset authors. By submitting this form, you confirm that you have read and agree to the CC-BY-NC-SA 4.0 license terms.

Mathematics Dataset Card: AM, Adaptiv World, and Adaptiv College

Overview

This repository contains a cross-country mathematics interaction dataset built from three digital learning platforms used in classrooms in France and Côte d’Ivoire. It spans three educational age bands, from early primary school to middle school and high-school remediation, and covers several mathematical domains, including number sense, arithmetic problem solving, geometry, fractions and proportional reasoning, and data organization and functions.

Adaptiv'Math (AM) is an adaptive digital learning resource for early primary mathematics. It was developed in the context of the French P2IA program and is designed to help young learners build a robust understanding of core mathematical concepts through interactive exercises, structured remediation, and adaptive progression.
Adaptiv College contributes mathematics work on geometry, proportions and fractions, and data organization.
Adaptiv World contributes mathematics work on number sense and the organization and management of data and functions. The data were collected in Côte d’Ivoire.

The broader source platforms contain a larger content universe than the selected modules shared here.

All three sources are part of an AI-driven teaching platform that can personalize learning pathways through the ZPDES logic. Content is organized hierarchically as:

module -> objective -> activity -> exercise

Although unusual, an exercise can belong to several activities or modules. Throughout this card, a source means one of the three platform datasets: am (Adaptiv’Math), adaptiv_world (Adaptiv World), or adaptiv_college (Adaptiv College).

A companion visualization tool is available at: https://miaamdatav2.streamlit.app/. It provides an interactive interface for inspecting the dataset.

Dataset scope

Adaptiv'Math modules kept

M1 — Nombres et calcul
M31 — Résolution de problèmes arithmétiques - Niveau 1
M32 — Résolution de problèmes arithmétiques - Niveau 2
M33 — Résolution de problèmes arithmétiques - Niveau 3

Adaptiv College modules kept

M101 — Espace et Géométrie
M102 — Proportions et fractions
M103 — Organisation et gestion des données

Adaptiv World modules kept

M101 — Réapprentissage du sens des nombres
M103 — Organisation et gestion de données, fonctions

Module labels such as M101 are display codes, not the UUID identifiers stored in module_id. The same code can denote different modules in different sources, as shown above.

Pedagogical and interaction modes

The work_mode column records how the exercise was assigned to the student.

`adaptive-test`

At the beginning of a module, students can complete an adaptive placement test that estimates their level before the main recommendation phase begins. The test chooses subsequent questions dynamically from previous answers, allowing it to identify a suitable module entry point with relatively few attempts.

The goal is to place the student at a suitable entry point in the module rather than starting everyone from the same activity. Adaptive-test rows are present in all three sources.

`playlist`

In playlist mode, a teacher assigns a predefined set of exercises. This is not the adaptive recommendation mode: the sequence is selected by the teacher rather than online by the recommendation system. In the current dataset, playlist rows occur in Adaptiv World. Both interaction tables contain only four playlist attempts, from one student on four exercises.

`zpdes`

zpdes is the main adaptive learning mode. ZPDES stands for “Zone of Proximal Development and Empirical Success”. The system recommends subsequent work from the student's current estimated level and the pedagogical dependency structure, with the aim of selecting tasks that are neither too easy nor too difficult.

The recommendation logic operates over a structured exercise graph organized by modules, objectives, and activities. The aim is to keep students working on tasks that are neither too easy nor too difficult, so that the sequence stays supportive of learning progress.

Counting conventions and identifiers

Several exercise counts are reported because the same exercise can occur in more than one source or pedagogical location:

An attempt is one row in an interaction table. It records a retained response event for an exercise; repeated attempts on the same exercise occupy separate rows.
A source-specific student is identified by (source, user_id).
A source-specific exercise is identified by (source, exercise_id). This is the recommended exercise identifier.
A bare exercise ID is exercise_id without source. Bare IDs are not globally unique across sources.
A hierarchy-membership row represents one placement of an exercise in a module, objective, and activity. The same source-specific exercise can therefore occur in multiple rows of maths_exercises_table.parquet.
An exercise is observed when it has at least one attempt in maths_data.parquet. The exercise catalogue also contains exercises with no retained interaction.

The current release contains:

Scope	Attempts	Source-specific exercises	Bare exercise IDs
Main interaction table	4,715,851	8,831	8,547
Filtered interaction table	4,681,851	8,804	8,520
Complete exercise catalogue	—	9,167	8,763

The exercise catalogue contains 9,938 hierarchy-membership rows for its 9,167 source-specific exercises.

Data files

The Parquet and JSON files below are stored in data/. Screenshot and notebook paths are given relative to the repository root.

`maths_data.parquet`

Main attempt-level dataset.

26,336 source-specific students, identified by (source, user_id)
4,715,851 attempts
8,831 source-specific exercises, identified by (source, exercise_id)
8,547 distinct bare exercise_id values

Columns:

user_id — Student identifier; use together with source.
classroom_id — Classroom identifier; missing for 1,065 main-table attempts.
playlist_or_module_id — Assignment-context identifier. In this release it matches the catalogue module_id on every row, including playlist rows.
objective_id — Identifier of the learning objective associated with this attempt.
activity_id — Identifier of the activity associated with this attempt.
exercise_id — Exercise identifier; use together with source.
created_at — Transformed timestamp without timezone metadata; see the clock explanation below.
data_correct — Boolean correctness flag recorded by the platform.
work_mode — Assignment mode: adaptive-test, playlist, or zpdes.
data_answer — Recorded response stored as a string; its interpretation depends on the gameplay type.
data_duration — Response-duration value in milliseconds.
source — Originating dataset: am, adaptiv_world, or adaptiv_college.
attempt_index — Attempt counter starting at 1 within (source, user_id, exercise_id). It counts attempts on that exercise, not all events for the student.
session_id — Session identifier derived from the login timestamp within (source, user_id), formatted as source::user_id::session_N. It is missing when login time is unavailable (113,287 main-table attempts).

Timestamp interpretation. For a known login time, created_at stores the login date at midnight plus elapsed time since login. For example, a response at 08:01 after a login at 08:00 becomes 00:01 on the login date. If login time is missing, the clock starts at the student’s first observed attempt on that calendar day.

Duration quality. The main table retains 10 negative duration values and some very large values. Treat these as data-quality limitations when analyzing response times.

Current source breakdown:

am: 24,393 students, 4,511,501 attempts
adaptiv_world: 1,076 students, 148,691 attempts
adaptiv_college: 867 students, 55,659 attempts

Current work modes present in the dataset:

adaptive-test
playlist
zpdes

Source-specific work modes:

am: adaptive-test, zpdes
adaptiv_world: adaptive-test, playlist, zpdes
adaptiv_college: adaptive-test, zpdes

`maths_data_filtered.parquet`

Stricter filtered version of the main dataset. This file is produced by preprocessing/pre_processing_maths_data.ipynb, which applies three optional filters in the following order:

remove students whose retained history is adaptive-test only
remove individual attempts whose (source, exercise_id) pair is absent from the exercise table
remove students whose remaining trajectory contains fewer than 5 attempts

The filtered dataset contains:

22,226 source-specific students
4,681,851 attempts
8,804 source-specific exercises
8,520 distinct bare exercise IDs

Compared with maths_data.parquet, this removes:

4,110 students, or 15.61%
34,000 attempts, or 0.72%
27 source-specific exercises and 27 bare exercise IDs

Current source breakdown:

am: 20,812 students, 4,482,271 attempts, 5,422 exercises
adaptiv_world: 857 students, 146,926 attempts, 1,355 exercises
adaptiv_college: 557 students, 52,654 attempts, 2,027 exercises

All interactions in the main table currently have a corresponding exercise in the exercise table. Consequently, the missing-exercise filter removes no attempts in this release; it is retained in the preprocessing pipeline as an integrity safeguard.

`maths_exercises_table.parquet`

Exercise metadata table. It represents pedagogical placements rather than containing exactly one row per exercise.

9,938 hierarchy-membership rows
9,167 unique (source, exercise_id) pairs
8,763 distinct bare exercise IDs
26 gameplay types (interaction formats, such as MULTIPLE_CHOICE or INPUT_FIELD, rather than mathematical topics)

The difference between rows and source-specific exercises is caused by exercise reuse across pedagogical locations:

733 Adaptiv'Math exercises occur in more than one hierarchy location;
three Adaptiv College exercises occur in two activities of the same module.

The difference between 9,167 source-specific exercises and 8,763 bare IDs comes from 404 exercise IDs that occur in both Adaptiv College and Adaptiv World.

Current source breakdown:

am: 6,371 hierarchy rows and 5,603 source-specific exercises
adaptiv_world: 1,527 hierarchy rows and 1,527 source-specific exercises
adaptiv_college: 2,040 hierarchy rows and 2,037 source-specific exercises

Columns:

exercise_id — Exercise identifier within a source.
gameplay_type — Interaction format used by the exercise.
content — Serialized exercise content, including instruction, question, correct answer, feedback, and image references where available. Most values are JSON; 315 AM exercises (389 catalogue rows) use Python dictionary-literal strings.
module_id — UUID identifier of the module containing this placement.
module_name — Human-readable module name.
objective_id — Identifier of the objective containing this placement.
objective_name — Human-readable objective name.
objective_pedagogical_intent — Text describing the intended learning goal of the objective.
activity_id — Identifier of the activity containing this placement.
activity_name — Human-readable activity name.
activity_pedagogical_intent — Text describing the intended learning goal of the activity.
source — Originating dataset.

`maths_dependencies.json`

Simplified pedagogical dependency graph for the selected modules.

9 modules
102 objective placements
553 activity placements
586 activity dependency edges

An objective placement is an objective listed under a module; an activity placement is an activity listed under an objective within a module. Each occurrence is counted, even when its identifier also occurs elsewhere in the hierarchy or in another source. For example, one AM activity_id occurs under M31O4, M32O4, and M33O4, contributing three activity placements. There are 100 distinct bare objective IDs and 511 distinct bare activity IDs. The graph is nested as modules[module_id].objectives[objective_id].activities[activity_id].

For each activity, the dependency file stores:

code — Human-readable activity code.
title — Short and long activity titles.
exercise_ids — Exercises assigned to this activity placement.
prerequisite_activity_ids — Activities listed as prerequisites of this activity.
unlocks_activity_ids — Activities for which this activity is a prerequisite.

Each edge is counted once per module, from prerequisite to dependent activity. The prerequisite and unlock lists encode the same relationships in opposite directions; do not add their counts together.

`screenshots/`

Every source-specific exercise in maths_exercises_table.parquet has a screenshot. Two versions are provided:

screenshots/raw/<source>/<exercise_id>.png — screenshot at its native exported resolution;
screenshots/compressed/<source>/<exercise_id>.png — copy resized to 960 pixels wide while preserving its aspect ratio.

The release contains:

9,167 original screenshots;
9,167 compressed copies;
18,334 PNG files in total.

Of the 9,167 source-specific exercises with screenshots, 8,831 are observed in maths_data.parquet; the remaining 336 belong to the released exercise catalogue but have no retained interaction.

The source directory is part of the screenshot identifier. In particular, 404 bare exercise IDs occur in both Adaptiv College and Adaptiv World and have separate screenshots.

`descriptions.json`

descriptions.json provides textual exercise descriptions using source-specific keys of the form <source>/<exercise_id>, where source is am, adaptiv_college, or adaptiv_world. Use these keys to retrieve the description associated with an exercise in a given source. The file contains 9,167 source-specific descriptions, covering every exercise in the released catalogue. The file also retains 8,744 legacy entries keyed only by exercise_id. These older entries do not provide complete coverage. The file's 17,911 entries reflect these two key formats, rather than 17,911 distinct exercises.

`preprocessing/pre_processing_maths_data.ipynb`

Notebook that reproduces the optional stricter filtering used to generate maths_data_filtered.parquet. It is provided for transparency and reproducibility.

Recommended identifiers and joins

Always retain source when identifying students, exercises, screenshots, or interactions.

To attach the correct pedagogical placement from maths_exercises_table.parquet to an interaction, join using:

(source, objective_id, activity_id, exercise_id)

This combination uniquely identifies a hierarchy-membership row in the current release. Do not join the interaction table to the exercise table using only (source, exercise_id): exercises reused in multiple activities or modules would duplicate interaction rows.

Companion visualization tool

We provide a companion Streamlit application to facilitate dataset inspection and exploratory analysis: https://miaamdatav2.streamlit.app/

It allows users to explore global usage statistics, work-mode distributions, student-level trajectories, gameplay distributions, activity bottlenecks, transition patterns, curriculum structure, and module-level dependency graphs. The tool is intended to support dataset auditing, sanity checks, and hypothesis generation before downstream modeling.

Licensing

This repository contains two types of artefacts governed by different licenses:

Dataset: released under CC-BY-NC-SA 4.0. You are free to share and adapt the data for non-commercial research purposes, provided you give appropriate credit and distribute any derivatives under the same license.
Preprocessing scripts: released under the MIT License.

By accessing this dataset, you agree to the terms of the CC-BY-NC-SA 4.0 license.

Downloads last month: 127

Total file size:

2.08 GB