SkyRL research adds a proposed training workflow and evaluation criteria
The research document now explains a proposed SkyRL training loop using admitted documents, approved tools and reviewed scoring criteria. It separates learning scores from proposed creator payment eligibility.
Proposed training workflow
The proposed workflow lets a model attempt tasks, scores the outcomes and updates its parameters. Evaluation then uses tasks excluded from training. SkyRL coordinates attempts and updates; the operator supplies the model, task material and scoring rules.
Evidence and permission checks
A proposed scientific task checks supporting citations, required credit and source-use receipts against actual tool activity. Missing evidence should prompt abstention. Proposed tool-service permission checks would block unauthorized access independently of learned behavior.
Scoring and pilot comparison
The proposed approach allows rule-based learning scores without a teacher model, while semantic judgments require reviewed criteria and reference answers. The proposed small-model pilot compares the original model, supervised fine-tuning and supervised fine-tuning plus reinforcement learning. Base-model lineage and Ultra compatibility remain separate proposed gates.
Separate creator payment rules
Under the proposed SIP ledger rules, accepted agreements and verified use determine financial eligibility. A higher training score would not create a royalty.
Files in this change
| File | Change | Lines |
|---|---|---|
docs/SKYRL_SOVEREIGN_SOURCE_RESEARCH.md | modified | +21 −1 |