SkyRL assessment adds a proposed post-training option
The new assessment proposes SkyRL as an optional, self-hosted post-training framework for the research assistant. The work remains research and a proposed pilot, with no installation, GPU training or deployment performed.
Proposed training option
The assessment proposes keeping Nemotron 3 Ultra and the NVIDIA training path as the reference plan. SkyRL compatibility with Ultra remains unverified.
Compatibility checks
The proposal requires reproducible checks before adopting SkyRL for Ultra. These cover checkpoint loading, numerical agreement, optimization, weight transfer, save and resume, export, memory and throughput.
Model ancestry limits
The assessment identifies Qwen-based examples that fail the project's no-Qwen requirement. It also records Ultra's disclosed Qwen and DeepSeek training inputs and states that SkyRL cannot remove that ancestry.
Separate permissions and rewards
The proposed integration separates numerical training scores from contractual creator rewards. It assigns permission enforcement, source evidence, contribution agreements and accounting to SIP. The proposed adoption process reviews framework licenses, model components, data permissions and service terms separately.
Documentation entry
The README now lists the SkyRL assessment, including its coverage of licensing, model ancestry, Ultra compatibility, attribution behavior and creator payments.
Files in this change
| File | Change | Lines |
|---|---|---|
README.md | modified | +1 −0 |
docs/SKYRL_SOVEREIGN_SOURCE_RESEARCH.md | added | +165 −0 |