FoundationalASSIST

FoundationalASSIST is the first English-language educational dataset constructed specifically for research with large language models. Derived from authentic student activity on the ASSISTments platform between 2019 and 2024, it comprises 1.7Million interactions from 5,000 unique students across 3,395 mathematics problems from the Illustrative Mathematics curriculum for grades 6–8, each tagged with grade level and aligned to Common Core State Standards.

‍

The dataset removes three constraints on prior work. It preserves the complete natural-language text of every problem rather than an identifier, so models can read and reason about the mathematics itself. It records each student's actual response, including numeric answers, fractions, and symbolic expressions, rather than a first attempt as a binary right/wrong input. It also identifies the specific incorrect answer a student gave, making the underlying misconception recoverable. It also provides substantially more per-student context than any existing natural-language resource, making learning trajectories observable over time.

‍

We explore four research questions: 1) problem-level knowledge tracing, 2) cognitive student modeling, 3) difficulty comparison, and 4) pairwise discrimination comparison. We evaluated four contemporary models and found near-baseline correctness prediction with pronounced optimism bias, moderate competence on relative difficulty, and below-chance discrimination performance across every system tested.

‍

FoundationalASSIST is available on Hugging Face under a CC-BY-NC-4.0 license.

‍

Worden, E., Heffernan, C., Heffernan, N., & Sonkar, S. (2026). FoundationalASSIST: Dataset for Foundational Knowledge Tracing & Pedagogical Grounding of Large Language Models.

Read the full paper here.

‍