Where the data comes from
How much of this registry is real, where each record came from, and what attribution each source requires. If you are going to cite a number from CVIN anywhere, read this page first.
1 · Two scopes that must never be added together
Device-level (8,479 records, 8,016 real) — one listing equals one physical machine, identified by a host and card fingerprint. Identity grade L0.
Model-level (1,140 records, 1,140 real) — a platform's quote for a model, grade M. It does not point to any specific device and cannot be used for collateral, underwriting or pool admission.
Adding the two together produces a number that means nothing. The machine-readable endpoint GET /v1/market/models reports device-level counts only, and says so in its caliber.not_additive field.
2 · Sources
| Source | Scope | Licence | Attribution as recorded |
|---|---|---|---|
| vast.ai 4,971 records |
device | Platform terms — no redistribution granted | Public market API snapshot; host + card fingerprint. |
| Clore.ai 3,045 records |
device | Platform terms — no redistribution granted | Public market API; whole-machine server id. |
| rentgpu.link 1,052 records |
model | CC BY 4.0 — attribution mandatory | Aggregated quotes, GPUPLN v0.1. Citation stored in attrs.attribution. |
| Synthetic seed (in-house) 442 records |
both | CC0 — CVIN own records | Marked synthetic:true on every record. |
| RunPod 88 records |
model | Platform terms — no redistribution granted | Official GraphQL quotes (secure / community). |
| Smoke / audit / security tests 20 records |
both | CC0 — CVIN own records | Kept because the ledger is append-only. Classified out of the "real" count. |
| In-house demo writes 1 records |
both | CC0 — CVIN own records | Written by CVIN itself for demos (*.cvin.local). Never a statement about a real device. |
3 · What you owe each source
rentgpu.link — CC BY 4.0, attribution mandatory. The citation string is stored on every record under attrs.attribution (cite_zh / cite_en / canonical_url). If you build on these quotes you must carry that attribution; CC BY material cannot be re-licensed as CC0.
vast.ai / Clore.ai / RunPod — no redistribution licence was granted to CVIN. These records are referenced as-is for reference only and are not released under the CC0 grant that covers CVIN's own records. If you intend to republish the dataset or build a commercial data product, obtain written permission from those platforms first, or filter them out. The exact filter command is in LICENSE-DATA.
4 · Read this before citing a number
Synthetic seed records are flagged per record with synthetic:true and are 100% of the demo devices. Real devices carry identity grade L0: cross_operator:false, and not_usable_for includes collateral registration and insurance underwriting. CVIN records facts; it does not produce credit scores or liability findings.
Machine-readable versions: GET /v1/provenance · GET /v1/market/models · OpenAPI 3.1.