Ponder training
Choose a workflow and compatible adapter, prepare reviewed data, and compare before serving.
Ponder connects reviewed data, model adapters, independent evaluation and saved Model versions. Choose what the model should learn or change, prepare the exact inputs, then explicitly approve the run and its spending. Choose SFT, RL or Decision model before selecting the adapter. Opening setup or saving a configuration does not start compute.
Start here
| Your next step | Guide |
|---|---|
| Prepare demonstrations, tasks and graders | Prepare training data |
| Learn from reviewed answers | SFT with LoRA, Engram adapter or Neuron training |
| Learn from rewards on fresh attempts | Reinforcement learning / GRPO |
| Train or score defined choices | Decision model / specialized LoRA |
| Review the result and its lineage | Models and candidate decisions |
| Activate a qualified saved version | Model deployments |
| Bring reviewed local work into the workflow | Use with Desktop |
The sidebar separates data preparation, adapters, continual learning and Model workflows. Workflow and adapter are separate choices; saved sources retain their exact method and lineage. Each method guide explains its mechanism, exact inputs, client commands, compatibility and result interpretation.
From data to a model in use
A Dataset holds versioned tasks, references and review evidence. An Experiment compares outputs without training the selected target. Models → Runs retains training history and its original evidence. A Model retains reusable versions and their exact artifacts.
Repeat supported learning
A Model's continual-learning policy can check for new admitted data on a schedule. Reviewed SFT batches train from the pinned base each time. GRPO can continue an accepted compatible parent with Shared, Stacked or Merged LoRA, subject to rank and runtime checks. The separate conversation-driven automatic training workflow currently uses GRPO.
Inspect the next check, timezone, waiting reason, review/ready/used counts and spending limits. Too few eligible examples means waiting without a training Job. Candidate acceptance and serving activation retain their own authority. See Set up continual learning and How it works.
Manual Engram and Decision parent continuation is separate from a scheduled policy. Saved neuron patches do not support training continuation. Recover an interrupted run through its original operation before starting a new stage.
Use Harness refinement for changes to agent instructions and tools. That separate review loop does not itself update model weights.
Compatibility
Training requires compatible data, model and runtime, plus a reviewed spending limit. A saved configuration does not prove runtime qualification.