Research summary proposes foundation model and teacher choices
The new research summary proposes KL3M for a source-provenance pilot and OLMo-core for training newly initialized weights on admitted sources. The project remains at specification stage and has not benchmarked or approved any model for production.
Proposed pilot
The summary proposes evaluating KL3M-003-3.7b for a small pilot focused on source permissions. Checkpoint provenance and certification scope still need confirmation.
Proposed foundation training
The summary proposes training new weights with OLMo-core on a separately admitted corpus. This proposal requires independent selection of data, tokenizers, filters and teachers.
Conditional teacher proposal
The summary proposes evaluating self-hosted gpt-oss-120b as a teacher if the project accepts unresolved upstream provenance. No frontier teacher has qualified under the strict requirement for verified absence of Chinese model ancestry.
Existing target and unresolved conflict
The specification retains Nemotron 3 Ultra as its proposed large-model target. The summary states that disclosed Qwen and DeepSeek dependencies fail the earlier ancestry exclusion.
Files in this change
| File | Change | Lines |
|---|---|---|
README.md | modified | +3 −0 |
docs/MODEL_SELECTION_RESEARCH_SUMMARY.md | added | +192 −0 |