What GearDel is for

Why a catalog, not another download mirror

Hugging Face, AI Hub, GitHub, and NIKL already host the bytes. What is hard is comparing whether a set is for evaluation or training, whether the catalog license is verified, and whether commercial use is even plausible before you click through. GearDel is that comparison layer.

In this snapshot: Hugging Face-family 113, AI Hub-family 96, GitHub-family 73, NIKL 7. Those counts overlap when a row cites more than one host.

How a field gets Verified

Each dataset row has per-field status: source URL, license, language, format, size. Verified means a human or a strict matcher confirmed the value against the original page. Needs review means there is a hint but it must be re-checked. Unknown means the catalog does not claim a value.

• License field: 182 verified, 87 need review, 17 unknown.
• Source URL field: 130 verified, 0 need review, 156 unknown.

We do not invent a license to fill a blank. A missing license stays missing. AI Hub rows keep the provider terms label instead of pretending they are MIT.

Commercial-use marks

Catalog mark “allowed” 161, “conditional” 120, “restricted” 5. Allowed still means you must re-read the original terms. AI Hub and NIKL almost always add signup or a use pledge even when a row sits in the commercial collection.

Who maintains the catalog

One operator maintains the rows. We open the original page and copy source, license label, and size text. We do not download the archive to remeasure it. Verified means the catalog value matched that page. A blank license stays blank. If a host returns 404, the card should leave the catalog instead of sitting as a dead page.

Mistakes and missing datasets: [email protected]

What we will not do

• Host or re-license files. Downloads stay on the original page.
• Guarantee that a catalog size string matches the bytes on disk today.
• Give legal advice. The license guide is a reading aid for this catalog.