MIAAM Dataset

MIAAM is a large-scale mathematics learning dataset built from real interactions collected on the Adaptiv'Math and MIA Seconde adaptive learning platforms.

7.24Mstudent attempts
45,848learners
7,845exercises
365activities

The dataset links student attempts to exercise content and pedagogical context. Released resources include:

  • attempt-level interaction logs with anonymized identifiers, exercise identifiers, submitted answers, correctness, timing, work mode, and attempt order;
  • exercise-level metadata including gameplay type, structured content, expected answer, feedback, curriculum mapping, and pedagogical intents;
  • screenshots of rendered exercise interfaces for visual and multimodal analysis;
  • objective-level and activity-level prerequisite graphs used to structure adaptive progression;
  • filtered experiment-ready tables, preprocessing scripts, baseline evaluation code, and visualization tools.
MIAAM on Hugging Face Dataset repository with documentation and access conditions. Visualization application Interactive exploration of statistics, trajectories, curriculum graphs, bottlenecks, and activity-level patterns. Evaluation code Experiment code for reproducing baseline evaluations and knowledge-tracing comparisons.

The companion visualization application can generate several complementary views of learning trajectories.

Classroom-level progression Sankey diagram generated by the MIAAM visualization application
Example classroom-level progression Sankey diagram from the companion visualization tool. Each node represents an activity reached by students, and link widths indicate the number of unique students following that part of the progression path. The figure shows only the first five newly reached activities, making visible the diversity of early learning trajectories within the selected classroom. Revisits to already seen activities are ignored, so the visualization emphasizes progression breadth rather than back-and-forth navigation.
Module-level activity matrix generated by the MIAAM visualization application
Example module-level activity matrix from the companion visualization tool for Adaptiv'Math Level 1 of the Arithmetic Problem Solving module. Rows correspond to objectives and columns to activity positions within objectives. Each populated cell represents one objective-activity pair and shows the selected metric; in this example, the metric is the number of recorded attempts on exercises in that activity. Other available metrics include success rate, exercise-balanced success, repeat-attempt rate, first-attempt success, number of unique playlist exercises, and mean exercise Elo.
Student-level radar chart generated by the MIAAM visualization application
Example student-level radar chart from the companion visualization tool for the MIA Seconde module Relearning Number Sense. Each spoke corresponds to one objective in the module. The orange trace shows coverage, defined as the percentage of exercises in that objective attempted at least once by the student, relative to the full objective catalog. The green trace shows the student's success rate, defined as the percentage of correct answers across all attempts made within that objective. Objectives never reached by the student remain visible with 0% coverage, while the success-rate trace is left undefined for untouched objectives.

Physics Dataset

The Physics dataset contains expert-designed physics and chemistry activities prepared as reference material for synthetic exercise generation. It is based on open educational resources and adapted to EvidenceB-style interactive activities, with pedagogical metadata and structure suitable for downstream generation and validation.

This dataset supports the generation of new exercises from expert references while preserving target objectives, activity formats, and pedagogical constraints.

Physics on Hugging Face Expert reference activities and documentation for generation experiments. Access may be gated depending on resource status.

Evaluation Goals

GAIMHE uses these resources to evaluate several families of models and methods.

Knowledge tracing

Predict future student performance from interaction histories.

Multimodal modeling

Use exercise text, screenshots, answers, and pedagogical metadata.

Efficient student models

Keep evaluation methods practical for large-scale adaptive learning.

Generation validation

Assess exercises, hints, distractors, feedback, and visualization tools.

Data Protection and Access

Educational data is treated as high-stakes data. Public datasets are anonymized and minimized before release: student, classroom, and session identifiers are anonymized; temporal information is transformed to preserve order and durations without exposing direct timestamps; names, school identifiers, demographic free text, and geographic identifiers are not shared.

Some resources are distributed through gated access on Hugging Face so users acknowledge the license and intended use conditions before downloading the data. This allows GAIMHE to support open research while preserving safeguards around sensitive educational traces.