These guides reflect the local
stateset-agents source tree reviewed on October 20, 2026.
It declares version 0.67.0 and includes uncommitted development work beyond the latest
local release commit, v0.66.1. Use the reviewed checkout for these workflows; the version
string alone does not establish that every feature is available in a published package.Choose your workflow
The improvement loop
Logs can be ingested and graded withingest and improve run. Curation writes
curated.jsonl, improve_summary.json, and next_steps.md. Inspect the result and its
format before using it as training data. A high reward on the selection set is not a
held-out quality result.
The studio’s finetune prepare snapshots the dataset and records a plan checksum. Launch
uses that exact run and checksum. Repeating a launch reconnects to its original submission;
it does not create a replacement paid run. Training completion, checkpoint selection, and
production promotion are separate outcomes.
Methods and execution backends
The package also implements GEPO, PPO, VAPO, and RLAIF research/trainer surfaces. Their
availability and maturity differ; see the RL framework guide.
The native River refund RL runner is a separate workflow using River’s rollout engine and
trainer. Choosing River in the studio does not translate native GRPO/GSPO/DAPO into River RL.
External engine adapters for verl, NeMo RL, and OpenRLHF bind experiment identity,
capabilities, and supported versions. Their live GPU conformance remains pending in the
reviewed backend guide. They are distinct from both StateSet’s native trainers and remote
compute providers; they do not replace every built-in trainer.
Evidence and recovery
Current work checks complete rollout groups, generated token/log-probability provenance, committed optimizer updates, checkpoint state, and captured dataset lineage. Local SFT completion additionally checks a structurally complete safetensors LoRA export. These checks help establish what executed and what artifact was retained; they do not certify quality. The latest retained River recovery study reports six acknowledged CISPO updates and 960/960 candidate successes against 954/960 for the shared SFT parent. Its prespecified replication gate failed (p = 0.125); no checkpoint was promoted. See the
study scope and limitations.
Interfaces
The MCP server supports grading, curation, discovery, and free preset previews; its tools do not launch training. Authorized agents can separately use the JSON CLI for reviewed launches. The newer/finetune HTTP operator routes inspect retained runs, import evidence, and request local
stops; they do not launch training.
The Agents API also serves the framework’s existing agent,
conversation, training-job, and inference contracts. Check the actual deployed capabilities
and service-specific authentication before treating a documentation page as an active service.