PuneethAbhishek/LMA_project

Access requests will be faster if you have a university or research-affiliated email associated with your Hugging Face account

Unfortunate news: Please note that this project is primarily supported by federal grants from the US government. As such, we need to follow certain regulations. Sadly, one is that we cannot share data with researchers from countries designated as "Countries of concern". We hope to soon share this dataset with the many fabulous researchers from these countries, but as of today, we are not legally permitted to do so.

IMPORTANT UPDATE ON 3/19: Due to an issue with git LFS 700k rows in interactions.csv were missing. Please redownload the dataset to use these rows.

Note: Access requests are reviewed Monday. You may need to verify your email or use case, plesae monitor your email.

Coding mistakes (Discovered 9/29, thanks to the person who discovered tihs and emailed us.): There is a potential error on task 5 and 6 when it comes to text matching. We are hoping to fix this and will update the code and numbers accordingly.

Overview of Foundational ASSIST

Foundational ASSIST is a dataset containing all natural text of problems and student answers as recorded by ASSISTments. The problems are from from Illustrative Mathematics 6th - 8th grade math curriculum, a common core aligned curriculum popular in the United States.

The data is in the "Data" folder. The code used to generate results in the original paper by Worden et al. is in the Code folder, and the results (in case you do not wish to reinference) are in the "Results" folder.

The dataset is comprised of three files:

1. Interactions, which contains students attempts of problems,
2. Problems, which includes information relevant to problems, and
3. Skills, which links Problems to skills. The dataset was curated to include 5,000 unique students who have each completed between 211-421 problems in ASSISTments. The dataset includes 1.7 million instances of students solving problems, complete with the answer text, problem text, distractor text, and more.

Interactions File

Interactions consists of 1,722,169 unique instances of students solving problems. The information provided includes a) problem_id, linking to Problems b) hint_count, the number of hints the student requested c) answer_text, the exact text of their first answer d) saw_answer, a boolean indicating whether the student requested to see the correct answer e) discrete_score, 0 the students score f) end_time, the time at which the student put in the correct answer to the problem g) user_xid, a unique identifier for each student.

Note that ASSISTments provides a discrete_score of 1 only if the student gets the problem correct on their first attempt, without requesting any support. If the student requests a hint, requests the answer, or has multiple tries they receive a 0. Accordingly, a student's answer can be the correct answer, but they can receive a 0 if the student requests a hint or sees the answer before entering the answer. This is shown in “table 7 cognitive accuracy when answers are incorrect” in our paper.

Problems File

This file consists of information about 3,395 unique problems. The columns include

1. Problem Set Id, which links problems that follow one another
2. Problem Part, which indicates where in the Problem Set the problem occurs
3. Problem Type, describing the type of problem
4. Answer Type, describing the type of answer (see below table or more information)
5. Problem Body, the problem text with HTML or markup illustrating exactly what the student saw (code to convert to natural language is available on the github)
6. Fill-in Options
7. Fill-in Answers
8. Multiple Choice Options
9. Multiple Choice Answers
10. problem_id

Problem Set example: PSB6N4 consists of three problems. The first is problem_id 151389 (as it has ‘Problem Part’ = 1, the second is problem_id 151533, and the third/last is 151647.

Answer Type	Description	Fill-in Options	Fill-in Answers	Multiple Choice Options	Multiple Choice Answers
Numeric	The student must type in the correct number	The correct answer. If there are multiple correct answers they are separated by a “,”.	The correct answer. If there are multiple correct answers they are separated by a “,”.	n/a	n/a
Drop Down	The student must select the correct option from a drop-down menu. These are similar to multiple choice	All the drop down options, separated by “</p>,”	The correct dropdown option.	n/a	n/a
Algebraic Expression	The student must type in the correct, short, algebraic expression, or a similar equivalent expression. E.g. if the answer is a^2+b^2 then b^2+a^2 would also be correct.	The correct answer. If there are multiple correct answers they are separated by a “,”.	The correct answer. If there are multiple correct answers they are separated by a “,”.	n/a	n/a
Ordering	The student must order some values in some order. Note the initial order of how these are presented to students is randomized.	The correct order of objects, separated by “,”.	The correct order of objects, separated by “,”.	n/a	n/a
Exact Match	The student must type in exactly the correct answer. This could be a number, expression, point, list, etc.	The correct answer. If there are multiple correct answers they are separated by a “,”. Note that some answers, e.g. lists, require the whole text (1,2,3) and there is only a single answer.	The correct answer. If there are multiple correct answers they are separated by a “,”. Note that some answers, e.g. lists, require the whole text (1,2,3) and there is only a single answer.	n/a	n/a
Exact Fraction	The student must type in exactly the correct fraction.	The correct answer. If there are multiple correct answers they are separated by a “,”.	The correct answer. If there are multiple correct answers they are separated by a “,”.	n/a	n/a
Numeric Expression	The student must type in a numeric expression. Note that simplification occurs, e.g. if the answer is 11^3, 1331 is also considered correct.	The correct answer. If there are multiple correct answers they are separated by a “,”.	The correct answer. If there are multiple correct answers they are separated by a “,”.	n/a	n/a
Multiple Choice	The student must select the correct option.	n/a	n/a	A list of options, separated by ‘||’.	The correct option.
Check all that apply	The student must select all correct option(s).	n/a	n/a	A list of options, separated by ‘||’.	The correct option(s), separated by ‘||’.

Skills File

The skills file consists of

1. problem_id, linking to problems in the Problems file
2. skill_id, a unique identifier per skill
3. node_code, which identifies the ASSISTments Skill tag Illustrative Math code for the skill and
4. node_name, a description of the skill. In total there are 224 unique skills.

Data Source

All data across each file are from ASSISTments, where students complete in-class work as well as homework and receive support and feedback from the platform. This work was done in conjunction with Dr. Heffernan’s lab at WPI. To ensure student privacy, our team attempted to remove all Personal Identifiable Information (PII). However, it is possible students could type PII into fill-in problems, which we aimed to detect and remove, but short of a manual review of 1.7 million interaction logs becomes infeasible. Accordingly, we ask that if people using this dataset come across PII to please contact us at etrials@assistments.org so it can be removed.

License and Sharing Agreement

This dataset is licensed under CC-BY-NC-4.0. We require that this dataset is used for research and educational purposes following this Responsible Use Guidelines.

Citation

If you use the FoundationalASSIST dataset in your research, please cite the following paper:

Worden, E., Heffernan, C., Heffernan, N., & Sonkar, S. (2026). FoundationalASSIST: An Educational Dataset for Foundational Knowledge Tracing and Pedagogical Grounding of LLMs. arXiv preprint arXiv:2602.00070.

BibTeX

@inproceedings{worden2026foundationalassist,
  title={FoundationalASSIST: Dataset for Foundational Knowledge Tracing \& Pedagogical Grounding of Large Language Models},
  author={Worden, Eamon and Heffernan, Cristina and Heffernan, Neil and Sonkar, Shashank},
  booktitle={International Conference on Artificial Intelligence in Education},
  pages={489--498},
  year={2026},
  organization={Springer}
}

FAQs

Q: Where is the code that cleans the problem text? A: This repository, in Code/clean_utils.py or cleantext.py, both are similar.

Q: How does discrete_score work? A: Discrete score is 1 if the student solved the problem on their first try without requesting a hint (hint_count) or an explanation/seeing the answer (saw_answer). Note this is recorded by ASSISTments. There are 433 rows where discrete_score = 1 despite saw_answer = True or hint_count > 0. In these instances, the student correctly solved the problem then requested hints/explanation/the answer anyways.

Q: Does the dataset track multiple attempts? (E.g. the student first incorrectly said 3, then incorrectly answered 6, then correctly answered 9) A: No. We recognize it may be valuable to have this data. However, currently, answer_text (in Interactions.csv) is the first answer (right or wrong) the student submitted for the problem. Second/future attempts are not included in this dataset, but stay tuned.

Q: I ran python gptoss120bvllmmcq.py --batch-size 10 --num-students 500 --bin-size 10 --min-history 50 --data-dir ../Data --num-gpus 4(or a similar) line and got a different number of predictions. Why? A: Between running the code and publishing the data we slighlty modified our PII filter. As such, this line: selected_users = np.random.choice(all_users, size=min(num_students, len(all_users)), replace=False) Had a slightly different order, which screwed everything up. If you'd like to run code using the exact same users, please modify your code to get the unique user_ids from Results/inference_data_kt_results.zip, and make selected_users be those unique user_ids. However, note that you may also find more rows (since our prior PII filter was overly strict, and we relaxed it) (~31k if you use the same script).