Model selection proposal puts Nemotron training target on hold
The documents put the proposed Nemotron 3 Nano training target on hold because its official card discloses Qwen use. No replacement model has approval.
Proposed ancestry requirements
The proposed selection rule requires US-led candidates and excludes Chinese-developed models, including base checkpoints and training teachers. Unknown ancestry remains pending. The proposed release review also covers tokenizers, synthetic data, filtering models, reward models and other learned dependencies.
Comparison guides proposed research
The documents reference an index comparing ten US lab/provider candidates and ten open-source paths across ten parameters. The proposed research direction prioritizes Ai2 or Marin training software with a newly admitted corpus for training from scratch.
Training milestone changes
The proposed training milestone now calls for an ancestry-approved model experiment using admitted SIP data, evaluated exports and disclosed provenance. The architecture retains Nemotron profiles, recipes and milestones as research alternatives without execution approval.
Audit recommendation changes
The selection index supersedes the audit's original Nemotron recommendation. The audit remains a reference for proposed rights and contract design. Adding SIP material cannot establish that every upstream training source was enrolled, attributed or paid.
Files in this change
| File | Change | Lines |
|---|---|---|
README.md | modified | +4 −3 |
docs/FRONTIER_LAB_IP_AND_CREATOR_REWARDS_AUDIT.md | modified | +1 −1 |
docs/SOLUTION_ARCHITECTURE.md | modified | +3 −1 |
docs/SOVEREIGN_SOURCE_AI_CODEX_SPEC.md | modified | +3 −0 |