SketchKT

SketchKT is a de-identified Chinese mathematics dataset for multimodal knowledge-tracing research. It has two data-table directories and two shared scratch-image libraries. Images are stored once only. Temporal KT stores its large image library in tar.gz shards. Extract the shards at the dataset root to create scratch_images/; every existing scratch_image value then works as an ordinary relative path. Progress Verified KT uses ordinary relative image paths without extraction.

Component	Contents
Temporal KT/	10,000 selected student sequence windows, 423,712 deduplicated interactions, questions, knowledge components, and train/validation/test splits. Each sequence contains 80 historical interactions and one target interaction.
Progress Verified KT/	A 1,031-row labeled test set: 841 correct and 190 incorrect. Each row has exactly reference_answer, question_text, scratch_image, student_answer, and is_correct.
scratch_archives/	Shared tar.gz drawings used by Temporal KT, plus an archive manifest.
pvkt_scratch_images/	Shared masked drawings used by Progress Verified KT.

Privacy and responsible use

This release removes source identifiers, absolute dates, source URLs, and demographic/geographic fields. Public image names are newly assigned and every released image is checked for embedded EXIF metadata. Structured personal-data patterns in released text are conservatively redacted. The image pixels are not altered beyond the pre-existing Progress Verified KT answer masking.

No automated check can establish that handwriting has no identifying content in every image. Users must follow DATA_USAGE.md, avoid re-identification, and report suspected personal information to the maintainers rather than redistributing it.

License

This dataset is made available under CC BY-NC 4.0. The additional data-use terms in DATA_USAGE.md, including the prohibition on re-identification, apply to access and use of this release.

See SCHEMA.md for the table fields and image-reference format.