Make the boundaries visible.
This public educational site pairs the current Dataset Production research edition with original interactive demonstrations. Chapters 1–15 and the labs teach general dataset-production principles. Chapter 16 and the application section explore a proposed Tuldok interface.
Evidence and version
The book is the delivered 2 October 2026 revised research edition, 16 chapters and four appendices, with a 80-page PDF. Primary sources and pinned repository references are in its evidence notes. The source audit inspected code and documentation. It did not train a model, run Tuldok, test a Pumas integration, or validate a live personal dataset.
Current Tuldok observations, branch-specific findings, and proposed features are distinguished in the book. The mockup is not the actual application. A link to a source is not a blanket license for its data.
Original teaching fixtures
The lab illustrations, tiny pixel mask, synthetic captions, source families, judge scores, and lineage records are original procedural examples created for this site. They contain no real participant records, photographs, voices, or private datasets. No remote model or image generator is called by the labs.
The fixture-rights checkbox documents the scope of the exercise. It does not confer permission to use unrelated media, verify a real license, or establish a participant’s consent. Public visibility of this repository and book does not itself grant a redistribution license. No distribution license has been selected here; source-specific terms remain applicable to referenced material.
What the downloads mean
- The geometry PNG is a real 20 × 12, 8-bit single-channel grayscale mask with only 0 and 255 values
- The geometry JSON shows point-center and half-open-box conventions
- The release JSON freezes the exact local teaching state, review history, and structural checks
- A release checksum identifies its downloaded JSON bytes. It does not establish correctness or trainer compatibility
- The release JSON is not a full training bundle. A production export requires real media, a task-specific adapter, and validation against the actual consumer
Privacy and local state
There are no analytics trackers, third-party fonts, backend APIs, accounts, or model calls in this site’s code. Lab state stays in localStorage in this browser. Downloaded files are produced locally. Use “Reset all labs” in a workspace to clear edits, decisions, and frozen snapshots from this site’s browser storage. Clearing browser storage also removes them. Standard GitHub Pages hosting may receive normal web-request information.
Do not treat this site as a secure store for confidential datasets. The provided labs do not ask you to upload personal or sensitive media.
Accessibility and reproducibility
Every workspace has keyboard-operable controls and visible focus indicators. The canvas also offers numeric coordinate controls; visual status is accompanied by text. The book is readable without JavaScript and the PDF and Markdown editions are available for download. Browser labs require a modern browser with JavaScript; the PNG exporter uses CompressionStream and release hashing uses Web Crypto.
Build and test instructions are in the public repository. The site is static HTML, CSS, and JavaScript with no frontend framework or runtime dependencies.