CV Dataset Workbench (Offline)
The same YOLO Dataset Health Checker logic as the browser tool, running as an installed desktop app against a real folder on disk instead of a browser upload. The browser's file APIs are genuinely limited for very large datasets; this native app has none of those limits — it opens a real folder picker, recursively scans the folder, and reads only label files and data.yaml (never image bytes), which is what lets a 200,000-image folder scan in seconds. Same finding codes, same severities, same manifest logic as tuliq.app's browser tool — reused directly, not reimplemented. V0.1 of Tuliq Offline, a new line of native desktop tools for datasets too large, private, or slow for a browser tab. Windows only for now.
A Windows installer (~80MB) — download it, run it, then launch Tuliq Offline — CV Dataset Workbench from the Start Menu. Windows only for now; macOS and Linux builds aren't available yet.
Download for WindowsUnsigned installer — Windows SmartScreen may warn on first run ("Windows protected your PC"). Click "More info" → "Run anyway" to proceed.
What's inside
- A real folder picker — pick any folder on disk, including ones with hundreds of thousands of files
- The same YOLO Dataset Health Checker findings as the browser tool, run natively
- Never reads image file bytes — only label files and data.yaml, so huge photo folders scan fast
About this tool
This is Tuliq Offline's first tool: the same YOLO Dataset Health Checker logic as the browser version, running against a real folder on your machine instead of a browser upload. The browser's File and Directory APIs are genuinely limited for large datasets — practical caps on file counts, no way to skip reading files you don't need, no real progress feedback on a slow scan. This app has none of those limits: it opens a native folder picker, recursively walks the folder on disk, and — critically — only reads the actual content of label files and data.yaml, never image bytes, since only their paths matter for this check. That's what lets a folder of 200,000 images scan in seconds rather than minutes. Everything else is identical to the browser tool: the same manifest-building logic, the same finding codes, the same severities. This is V0.1 of a broader "Tuliq Offline" line — desktop tools for datasets too large, too private, or too slow for a browser tab — built by reusing tuliq.app's existing, already-tested tool logic against a real filesystem rather than rewriting anything from scratch.
What you can do
- Scans a real folder on disk — no browser file-count limits, no upload
- Reads only label files and data.yaml — never image bytes — so 200,000-image folders scan in seconds
- Same checks as the browser's YOLO Dataset Health Checker, run natively
- Nothing is ever sent anywhere — no server, no account, no telemetry
How to use
Step 1
Download and install
Run the Windows installer, then launch it from the Start Menu.
Step 2
Choose a folder
Pick any real folder on disk — your whole dataset, not a browser-uploaded subset.
Step 3
Review the findings
The exact same YOLO dataset checks as the browser tool, grouped by problem type.
Example
The findings look exactly like the browser tool's:
CV.IMAGE_WITHOUT_LABEL
images/train/photo00482.jpg has no matching label file.
Same finding codes and messages as tuliq.app's YOLO Dataset Health Checker — this is the same check, just running against a real folder instead of a browser upload.
How your data is handled
This tool is labeled 🟣 Native desktop app: it's an installed Electron application, not a web page — there's no tuliq.app server involved at any point after you've downloaded it, no account, and no telemetry. It reads files directly from your filesystem using Node.js, entirely on your own machine. The one thing worth being upfront about: the installer itself is currently unsigned (no code-signing certificate yet), so Windows SmartScreen will likely show a warning on first run — this is expected for an unsigned executable, not a sign anything is wrong, but you should always be cautious about running unsigned software and should verify you downloaded it from tuliq.app's own site.
Frequently asked questions
Why does this need to be installed instead of just running in the browser?+
The browser's file-access APIs cap how many files you can reasonably work with and can't selectively skip reading files you don't need — for a dataset of hundreds of thousands of images, that adds up to real slowness and, past a point, outright failure. A native app has direct filesystem access with none of those limits.
Does this upload my dataset anywhere?+
No — there is no server component at all. The app reads files directly from your disk and does everything locally. Nothing about your dataset — file names, image contents, or label contents — ever leaves your machine.
Why does Windows show a warning when I run the installer?+
The installer isn't code-signed yet (that requires a paid certificate), so Windows SmartScreen flags it as being from an "unrecognized publisher." This is expected for a new, unsigned tool — click "More info" then "Run anyway" to proceed, but only do this because you downloaded it directly from tuliq.app.
Is this the same check as the browser's YOLO Dataset Health Checker?+
Yes, exactly — this app reuses that tool's actual logic unchanged, just fed from a real filesystem scan instead of a browser file upload. Finding codes, messages, and severities are identical.
Is macOS or Linux supported?+
Not yet — this first release is Windows-only. macOS and Linux builds are a reasonable next step, but each needs its own packaging and (for macOS) notarization to avoid a Gatekeeper warning even worse than Windows SmartScreen's.