precedentbrute/watching-the-work

Watching the Work: Multimodal Math Tutoring

201,882 images of learners’ handwritten math work, linked to AI tutoring conversations and problem-level scores.

Image files · Paper · Paper PDF · Explore the data · Project · Citation · Contact

Watching the Work follows 117 adults solving introductory calculus problems with an AI tutor. Phone-camera images capture their handwritten work as it develops during problem solving. The image sequences are linked to participants, problems, elapsed times, 4,145 learner–tutor messages, answer attempts, and pre/post assessment scores.

The release includes the images themselves: JPEG image bytes are embedded in the frames configuration, packaged in 68 Parquet shards containing approximately 12.8 GB of image data. The separate frame_metadata configuration provides the image index without downloading the images.

It accompanies Exploring Visual Access and Proactive AI Tutoring for Handwritten Mathematics by Nachiket Kotalwar, Conrad Borchers, Shubham Gandhi, Partha Sarathi Purkayastha, David Yaron, and Carolyn P. Rosé.

Start here

• Load the images: use the frames configuration with Hugging Face Datasets, as shown below, or download the image shards. Each record includes a JPEG image, participant label, problem, and elapsed time.
• Browse without coding: use the trajectory explorer to inspect a participant’s paper timeline alongside the conversation, attempts and assessments.
• Get all problem statements and keys: download problems.json, containing both nine-item assessment forms and the ten assisted problems.
• Analyze the data: load individual tables with Hugging Face Datasets. Images are in a separate configuration, so you can work with the tables without downloading the image collection.
• Understand the measures: read the data dictionary, analysis guide.

from datasets import load_dataset

repo = "precedentbrute/watching-the-work"
scores = load_dataset(repo, "participant_problem_scores", split="train")
chat = load_dataset(repo, "chat", split="train")
frame_metadata = load_dataset(repo, "frame_metadata", split="train")

# Read images incrementally rather than downloading the whole image collection.
frames = load_dataset(repo, "frames", split="train", streaming=True)
example = next(iter(frames))
print(example["participant"], example["problem"], example["t_s"])
example["image"].show()

Every configuration has one split named train for loading compatibility.

What is included

Configuration / file	Records	Contents
problems.json	28 problems	Two nine-item pre/post forms and ten assisted problems, with keys and scoring information
participant_problem_scores	3,276	Scores for 9 pre-test, 9 post-test, and all 10 assisted problems per participant; unobserved assisted scores remain null
participants	117	Condition, test-form assignment, aggregate scores, task summaries, and three self-reported background fields
problem_progress	699	Opened participant–problem trajectories and task events
chat	4,145	Learner messages, requested tutor replies, and proactive tutor messages
problem_attempts	778	Submitted answers and recorded scoring results
assessment_items	2,106	Pre/post item scores and response-status flags
assessment_answers	1,863	Recorded pre/post answers
frame_metadata	201,882	Image keys, elapsed times, content labels, and image quality measures
frames	201,882	The same image metadata with embedded JPEG images

CSV and JSONL exports are in tables/. Parquet files in data/ support the Hugging Face viewer and standard loading tools. Study feedback, post-study survey responses and ratings, survey transcripts, and audio are not released. Three self-reported background fields are included: prior AI use for mathematics, recency of work with derivatives, and preferred working medium.

The paper describes 255,984 recorded images in the study cohort. This release contains 201,882 images retained after privacy screening. Timing gaps and unavailable image references are expected. The export does not include raw model calls, internal wait/intervene decision logs, or the paper’s dialogue annotations. Export counts should not be substituted for the paper’s analysis counts.

Study design

Participants were U.S. adults recruited through Prolific who had previously studied differentiation and passed a brief eligibility check. Each participant completed one session with pre/post assessments, a 35-minute assisted problem-solving activity, and a final survey (not released). Participants worked through a common ordered bank of ten related-rates problems.

Condition	Participants	What the tutor could do
Text-Reactive	40	Respond to learner messages without seeing the paper
Visual-Reactive	37	Respond to learner messages using the paper view
Visual-Proactive	40	Use the paper view and also initiate assistance

Paper was recorded in all three conditions. A recording in Text-Reactive does not mean the tutor could see it. The paper describes recruitment, procedures, assistance policies, analyses, and limitations.

Working with trajectories

The consolidated score table retains the recorded pre/post item scores. For assisted problems, it reports the best recorded submission score, with null when there is no scored submission; opened and submitted distinguish an unreached problem from an opened problem without an answer. It does not assign failure scores to unreached problems.

Join tables using participant. Use (participant, problem) for assisted-problem trajectories and (participant, frame_id) for exact image references. Elapsed times are in seconds from the start of that participant’s session; they are not wall-clock dates. frame_metadata and frames share row_index and image_key.

A message’s referenced image may have been excluded by privacy screening. Preserve that missingness instead of silently treating the nearest retained image as the image the tutor saw. For temporal analyses, distinguish problem solving from worked-solution review using the recorded task events. See the analysis guide for examples.

Research uses and limitations

The linked records support work in learning analytics, educational data mining, human–AI interaction, multimodal tutoring, help seeking, and the timing of assistance. Researchers can examine how written work develops around a tutor message, compare requested and proactive help, or study how interaction patterns relate to task outcomes and immediate test performance.

The sample comprises adults with prior calculus experience who completed a remote study and established a usable camera setup. It should not be treated as representative of schoolchildren or first-time calculus learners. Repeated frames and messages from the same participant are dependent observations. There is no unaided control condition, and immediate pre/post change alone does not isolate the tutor’s effect.

Image content labels and useful flags are machine-generated convenience annotations, not validated labels of learning, correctness, attention, or student behavior. Tutor output can contain mathematical errors. Privacy screening removes parts of trajectories; absence of an image is not evidence of inactivity. No learner-uptake or intervention-quality ground truth is claimed for this release.

Privacy and preparation

Please respect participants’ privacy and do not attempt to identify or contact them. If you notice information that may identify someone, use the private contact below rather than reproducing it in a public discussion.

License

The dataset is released under the MIT License. The companion explorer code is also MIT-licensed; bundled third-party components retain their own notices. The paper has its own publication terms.

Citation

If you use the dataset or explorer in your research, please cite the accompanying paper and state the dataset revision used:

@misc{kotalwar2026visualaccesstutoring,
  title = {Exploring Visual Access and Proactive AI Tutoring for Handwritten Mathematics},
  author = {Kotalwar, Nachiket and Borchers, Conrad and Gandhi, Shubham and
            Purkayastha, Partha Sarathi and Yaron, David and Ros\'{e}, Carolyn P.},
  year = {2026},
  eprint = {2610.26411},
  archivePrefix = {arXiv},
  primaryClass = {cs.CY},
  url = {https://arxiv.org/abs/2610.26411}
}

Questions, corrections, and privacy concerns

For questions, errors, or suggestions, open a dataset discussion or contact Nachiket Kotalwar at nkotalwa@andrew.cmu.edu. Include the dataset revision and, where relevant, the file, participant label, problem, or frame ID.

For a privacy concern, please email directly. Do not post potentially identifying text or images publicly. Send only the record reference and a brief description needed to locate the issue.