Clean Set

1. What it does

Makes clean data to train on saves bloat.

2. Users and problem

Lots of people train their own models now: indie devs, small startups, and people at home who want their own AI. Clean Set is for them, though bigger labs could use it too. It's built for people with less SSD space or VRAM.

3. Proposed solution

It scans the files, flags duplicates and junk, and asks you how you want to handle them. It makes a clean copy, and your originals only change if you say yes.

4. User flow

1. Select the dataset or folder.
2. Start the scan.
3. Read the report.
4. Decide what to do: make a clean copy, move or delete the flagged files, or leave them.

5. What is different

It finds duplicate data inside files, not just duplicate files. And it all runs on your own NAS, so your data never gets uploaded anywhere.

6. NAS capabilities

• Scan: ug-filesystem find, stat, sha256sum; ug-filemgr scan-dedupe.
• Judge: ug-document qa, metadata, page-count.
• Report: ug-knowledge save-text.
• Copy: ug-filesystem mkdir, cpmv, cat, echo.
• Tidy: file-workflow batch-move, batch-del.

7. Technical approach

We tell the NAS how it works, and UGREEN's AI helper does the work. Changes wait for your yes.

8. Current status

It's ready, and it passes UGREEN's checker with no errors. The next step is testing on a real NAS, which will show how fast it is and how well it catches near-copies.

9. Validation plan

Testing shows whether it really cleans the data and why that matters. Clean data shouldn't fill the SSD, so there's room left to save the next training run. It should also train evenly, so a phrase said 10 times in one file doesn't train deeper than something said once when both are worth the same. And the extra training you add for weak spots, like rules on what not to do, should carry the weight it's meant to instead of getting drowned out. NAS tests: Core 1, exact copies. Core 2, repeated paragraphs plus junk files. Edge, a broken PDF and an existing output folder.

10. Limits and risks

It's AI, whether it's the NAS's or anyone else's. It should do the job well, but it's limited by how good the AI it runs with is. It's told never to invent anything: if it isn't in the files it's given, it isn't allowed. Big sets may be slow, and PDF and Word files are kept or left out whole. I tested the steps with a code stand-in that Grok wrote, not with UGREEN's AI, on a sample book set. It handled 10 of 10 planted problems and changed no originals, and made 0 wrong drops.

Third-party components and licenses

• None. Clean Set uses only UGREEN NAS CLI capabilities through UGREEN expert agents.
• No outside libraries, APIs, services, or data. No data flows off the NAS.

Built during the hackathon

• Every file in this Skill was written new during the hackathon, since I joined on Oct 25, 2026.

AI tool disclosure

Made by @Nefarious_Trey with Grok.

Grok helped with grammar and with listing the NAS tools. The idea and the answers are mine.

• Official tool: checked and updated with the official UGOS Skill Creation Tool (uliya-skill-kit) from the Hackathon Starter Kit, run with Grok as the AI assistant, following its steps. One fix came from it (Step 3 now has file-agent make the output folder first). Its checker, check-skill.mjs, gave 0 errors, 0 warnings. A static pass does not prove it runs on a NAS.

Built With

• grok
• skills
• ugos
• ugreen