Research document proposes an open-source stack for an independent LLM
The new research document proposes PyTorch, OLMo-core and Dolma for training an independent model on separately approved material. It reports no package installations, model training or compatibility benchmarks for this research.
Proposed training stack
The proposal calls for a tokenizer trained on approved material, newly initialized model weights and complete experiment records. It retains the existing Nemotron 3 Ultra reference selection and leaves its disclosed ancestry conflict unresolved.
Source permissions and rewards
The proposed research connects source permissions, recorded training use, evidence-backed answers and contributor rewards. Default datasets and released checkpoints require separate admission under the proposal.
Evaluation and reproducibility
The document proposes EleutherAI's evaluation harness and self-hosted MLflow to measure and preserve results. It proposes vLLM for serving after compatibility checks. It identifies restricted corpus access as a reproducibility limitation that experiment reports should disclose.
Training alternatives and learning resources
The document compares training kits and maps computer-science topics to proposed experiments. It proposes one trainer and a small dense model for the first research program. The README now lists the research document.
Files in this change
| File | Change | Lines |
|---|---|---|
README.md | modified | +1 −0 |
docs/OPEN_SOURCE_RESEARCH_LLM_STACK.md | added | +230 −0 |