New comparison index proposes training from scratch and holding Nemotron
The new index proposes Ai2 or Marin training tools for a model trained from scratch on permission-reviewed material. It proposes holding the Nemotron checkpoint because its documented Qwen ancestry fails the proposed selection screen.
Proposed training direction
The index prioritizes Ai2 and Marin training tools with newly initialized weights and admitted data. It finds insufficient evidence that any reviewed release combines source permission, verified ancestry, attribution and contractual creator rewards.
Proposed ancestry screening
The proposed screen excludes Chinese-developed checkpoints and derivatives, including synthetic training teachers. It treats unknown ancestry as pending. The index proposes holding Nemotron 3 Nano 30B-A3B because its model card documents Qwen use.
Comparison scope
The index compares ten lab or provider candidates and ten open-source paths across ten parameters. It separates software licenses, model licenses and training-data permissions. The review reports no benchmarks and approves no candidate for production.
Proposed implementation sequence
The index proposes building permission, attribution and accounting controls before small-scale training experiments. It keeps downloadable model options conditional on ancestry and permission evidence. The proposed creator reward system requires separate implementation.
Files in this change
| File | Change | Lines |
|---|---|---|
docs/USA_SOVEREIGN_AI_INDEX.md | added | +107 −0 |
docs/usa-sovereign-ai-index.html | added | +417 −0 |