Continual learning for AI models
Connect reviewed tasks, recurring model updates, held-out evaluation and separately approved inference with OpenPond continual learning.
Continual learning adapts a model as new data arrives while checking that useful previous behavior is retained. For AI applications, that means evaluating both newly corrected cases and independent tasks from earlier work.
OpenPond post-training connects reviewed data, supported model updates and independent evaluation into a retained cycle. You decide what evidence may be used and what can happen to each candidate. Serverless inference serves compatible saved versions through an API after a separate deployment decision.
Start with one nightly loop
- Bring in useful work. Import conversations from connected agent logs, send tasks and results from your harness through the CLI/SDK, or add tasks manually. Fix-it Agent captures supported tool failures and recovery evidence for review.
- Review the feedback. Choose graders, admit ordinary training tasks, and review corrections or exceptions. Keep separate tasks for your evals.
- Pick a time. In the Dataset's Continual learning tab, select a compatible Model and set its nightly time, timezone, minimum new examples and spending limits.
- Compare and decide. Follow each cycle's results and compare the trained candidate with your baseline on the same held-out evals. Approve a useful candidate or keep the current version.
For GRPO, choose Shared, Stacked or Merged LoRA when configuring a compatible continuation. To compare methods, train separate candidates, then compare them in evaluation Experiments with the same held-out tasks and graders. One schedule uses its saved method rather than trying all three automatically. A nightly check waits when there is not enough approved work or evaluation data.
Follow the quick setup for source choices, adding tasks, human feedback and result review. The detailed guides below explain the individual controls.
What you can configure
- Data: a published Dataset or approved conversation sources, with training and held-out families kept separate.
- Grading and admission: checked graders, human admission for ordinary Model tasks, or qualified verification for supported conversation policies.
- Timing and limits: minimum new examples, schedule, timezone and spending.
- Candidate decision: owner review or explicitly authorized automation, with serving authority checked separately.
Choose a guide
| Guide | What you will learn |
|---|---|
| How it works | Admission, scheduled checks, training, holdout protection and candidate outcomes. |
| Set it up | Dataset and conversation setup, checked graders, schedules, limits and review choices. |
| Research and evaluation | Adaptation, forgetting, replay and how to assess evidence over time. |
| Adaptation time | Correction receipt to first verified improvement, with resolved and pending evidence. |
| Repeat rate | Confirmed repeats among scored relevant opportunities, separately for Evaluation and Live. |
Start from your data
Use a Dataset's Continual learning tab to select a compatible Model and configure the saved policy. New Dataset creation first saves the tasks and source; open the saved Dataset to connect its learning settings. New Model setup includes Continual learning alongside Model, Dataset, Graders and Run settings on the configuration page.
The Model owns the policy, its exact training configuration and consumption history. A Dataset can supply several Models. These entry points do not create an independent Dataset scheduler.
| Evidence | Admission boundary |
|---|---|
| Ordinary published Dataset tasks | Explicitly submit selected training tasks, then record human admission in Model Tasks. |
| Supported conversation evidence | The saved policy grades and independently verifies examples; eligible examples can receive qualified admission. |
| Conversation exceptions | Inbox reviewers exclude an item or propose a correction that must pass verification. |
Saving a row, publishing a version or enabling a schedule does not automatically approve every example. See the Dataset handoff for the exact selected-task submission.
Which adapters support continual learning?
Here, continual learning means updates through an explicitly enabled Model policy. The full adapter support table also shows which methods can continue from a saved result in a separately reviewed operation.
| Workflow / adapter | Continual learning | Manual continuation scope |
|---|---|---|
| SFT / LoRA | Yes | Generic cross-job SFT adapter continuation is unavailable. |
| SFT / Engram adapter | No | Exact compatible parent module with a fresh optimizer. |
| SFT / Neuron training | No | Saved-delta training continuation is unsupported; inference is separate. |
| RL / LoRA (GRPO) | Yes | Shared, Stacked or Merged mode under exact parent/rank/runtime checks. |
| Decision model / specialized LoRA | No | Exact adapter parent in a separate reviewed operation; scoring does not train. |
Existing inference-only steering does not create a recurring training policy. Shared, Stacked and Merged are RL continuation modes rather than additional workflows. Recovering an interrupted operation is distinct from starting a new parent-training stage.
GRPO's Shared, Stacked and Merged modes have exact parent, rank, capacity and runtime requirements. A supported policy does not mean every model qualifies or that a nightly check must launch training. Ordinary Model policies require human-admitted training examples. The separate conversation-driven automatic policy currently trains with GRPO only.
Use Harness refinement to change agent instructions and tools around a model. That workflow has its own review boundary and does not itself update model weights.
Know when nothing should run
A scheduled check can legitimately wait because:
- too few approved, unused examples are ready;
- an exception needs a human decision;
- protected holdout data, runtime readiness or authority is missing;
- budget is unavailable or another iteration is active.
Inspect the waiting reason and next check. A completed training job is still only a candidate; a failed or inconclusive comparison keeps the current version.
Common questions
How is continual learning different from fine-tuning?
Fine-tuning adapts a model using a selected training dataset. Continual learning organizes repeated updates as new evidence arrives, with checks for both new learning and retained behavior. SFT or reinforcement learning can be part of a supported cycle; a scheduled job alone does not demonstrate that a model improved.
Does sending a task through the SDK start training?
No. A task append adds evidence to the selected Dataset. Review, training admission and an explicitly authorized run or enabled learning policy remain separate. See task ingestion and setup.
How do I check whether an update helped?
Compare the candidate and baseline on the same held-out tasks and graders, including older task families. OpenPond also proposes adaptation time and repeat rate as operational measures. These are proposed measures, not universal research standards or evidence that every update improves a model.
Can I deploy the model for inference after training?
Use Model deployments to review a compatible saved version, its runtime and costs, then activate its endpoint. Training completion and candidate acceptance do not automatically activate serving. Not every adapter or saved artifact has a qualified serving runtime.
Compatibility
Continual learning requires compatible training, admitted evidence, a ready runtime and saved budget/authority. A schedule does not guarantee a training job.