Models
Inspect trained versions, candidate lineage, evaluations, and deployment decisions.
Open Models to inspect saved versions produced by training or supported modifications. This collection is distinct from the hosted model catalog in Inference.
For workflow and adapter choices, start with Model adapters. Model setup retains the exact starting model and produces a saved version where the operation supports one.
Create or configure a Model
Choose SFT, RL or Decision model first, then its compatible adapter. SFT offers LoRA, Engram adapter and Neuron training; RL and Decision use their respective current LoRA implementation. Choose the exact starting base or saved version alongside that compatibility check.
Choose Save in the workflow and adapter chooser to open a resumable setup page with five sections: Model, Dataset, Graders, Continual learning and Run settings. Next saves the current section and opens the next; earlier sections remain editable. Save configuration retains the authoring draft at its setup URL. Moving between sections preserves reviewed data and exact source lineage; saved drafts do not acquire a new default method. Decision scoring retains its release and does not train. Saving and navigation start no compute and enable no schedule; Train and supported Enable continual learning actions require separate explicit review.
Existing steering versions and historical runs retain their original links. New standard Model setup offers the workflows and adapters above. Training history belongs under Models → Runs; new Experiments compare outputs without training the selected target.
Continual learning
The Model's learning policy connects its supported training setup to reviewed source data, a schedule, spending limits and candidate approval. It creates separate training runs and candidate versions over time; it does not overwrite the evidence from earlier runs. A Dataset can supply more than one Model.
Open the Dataset's Continual learning tab to find related Models and open these same settings. Select the Model before editing: the policy belongs to that Model and its exact configuration, not to an independent Dataset schedule. Review the source/release binding, starting version, minimum approved-task count, next check with timezone, budget and candidate behavior together.
To hand off published Dataset tasks, open the Dataset's Continual learning tab and choose Submit Dataset tasks for review for this Model. Select only training tasks from the exact bound release, then follow the retained receipt to the Model's Tasks queue. Only the Model owner with managed-training permission can submit this handoff. It does not approve the examples or start training.
Ordinary submitted evidence needs an explicit admission in the Model's Tasks review queue. Inspect awaiting-review, ready and used counts; enabling learning does not submit or approve all Dataset rows. Conversation policies can admit independently verified examples under their configured rules, with unresolved exceptions and candidate decisions in Inbox.
A scheduled check waits when the eligible population, budget or runtime is not ready. Pausing future cycles and stopping work already in progress are separate actions. Use retained run history when recovering a lost reply so that retries do not become duplicate training runs.
Read Continual learning for the cycle, or Continual learning setup to use imported agent conversations. Scheduling support is specific to the training configuration; an adapter being selectable does not make it available for automatic recurring updates.
Manual exact-parent continuation, interrupted-run recovery and scheduled updates are separate. Engram and Decision can use their supported saved parents in a reviewed operation; that does not give them a recurring policy. Saved neuron patches support inference, not further training.
Versions, training runs and evaluation
Review a candidate
A retained preparation can be revisited or discarded before submission. After Start, read the original Job state before retrying a lost request. Closing an editor does not cancel submitted work; use Stop and follow provider cleanup. Configure training from Models, retaining the exact starting version, Dataset and grader releases. Evaluation results can inform that decision; training examples still need separate selection and approval.
Open the version and follow its source training run, starting model, dataset, recipe, grader releases, and evaluation evidence. Inspect artifacts and any readiness or compatibility issues before using it in another run or serving it.
Compare the candidate with the current model on held-out tasks under the same scoring rules. Inspect critical failures and cost or latency changes as well as the aggregate score. An absent or failed check is not evidence of acceptance.
Decide how to use it
Keep the current model, investigate the candidate further, or explicitly accept the candidate where the available controls allow it. Training completion and candidate acceptance are separate from creating or activating a serving deployment. Open Model deployments for endpoint management and its usage and API guide.
Previous runs retain their captured versions and scores. A later model or grader change does not rewrite that history. See Experiments for fresh comparisons and Training for the creation workflow.
Sources
- OpenPond SDK contracts: saved versions and immutable references.
- Model adapters and Model deployments: adapter-specific results and serving.
Compatibility
A visible model record is not a callable deployment. Serving requires a compatible version, available capacity, and separate deployment authority.