StateSet’s RL and fine-tuning workflows live in the stateset-agents framework. Use the
product overview for the current source/version scope and
supported interfaces. The latest local work includes a shared terminal/agent control plane,
exact checkpoint recovery, dataset lineage, and retained River CISPO studies.
Start with one outcome
SFT and RL
Supervised fine-tuning learns from reviewed assistant demonstrations. Native RL samples
trajectories, computes task rewards, and updates a policy using actual rollout evidence.
The studio supports local SFT/GRPO/GSPO/DAPO; remote studio backends currently support SFT.
Additional trainers and the native River CISPO path have separate contracts and maturity.
Read GRPO and the training framework for trainer choices
and rollout integrity, or the RL platform guide
for the full data → training → validation → test workflow.
Inspect without launching
From the reviewed checkout’s Python environment:
These commands discover capabilities or render a static view. Real training needs a
separately reviewed launch and configured compute. A CPU preview is not successful training.
Keep the evidence boundaries
A completed export establishes a retained artifact, a validation result supports selection,
and a fresh paired test supports a scoped quality claim. None automatically approves a
production rollout. Current River replication reports higher pooled successes but failed
its prespecified acceptance gate; the River guide
retains both the observations and that failure.
The MCP surface provides curation and free previews.
The JSON fine-tuning CLI separately supports authorized launches; its HTTP operator routes
support retained-run inspection and cooperative local stops. Last modified on October 20, 2026
