Set up continual learning

Start from New Dataset or a saved Dataset's Continual learning tab. Choose the evidence first, then the compatible Model and policy. The Model owns the saved learning configuration; Dataset and Model entry points edit it.

Before enabling learning

Have these ready:

• permission to use the selected source and its destination workspace;
• an exact published Dataset/training setup with checked graders;
• a compatible trainable Model and runtime;
• independent held-out tasks and an acceptance plan;
• explicit spending and candidate authority.

Importing a conversation from a proprietary model does not make its weights available. Choose an available trainable Model as the target.

Start from a Dataset

1. In New Dataset, choose manual, Hugging Face, file, connected activity, CLI/SDK or an existing Dataset. One dialog keeps Overview, Graders and Continual learning together.
2. Configure the source in Overview. Hugging Face/file Import tasks updates that tab in place; expand the optional task list to inspect or edit it.
3. Select checked graders and review source permission, references and splits. Apply changed import settings with Update import before saving.
4. For new authored/imported tasks, choose Save version and continue in Continual learning. External-authoring and existing sources can select an already-saved Dataset.
5. Select a compatible Model, inspect its exact release binding, then configure or open its saved policy. Model/Experiment setup retains its existing Overview, Dataset, grader and learning selections.
6. For ordinary learning, explicitly Submit Dataset tasks for review, then admit eligible evidence in Model Tasks. Publication and policy enablement do not approve every row.

See Datasets for import limits and the exact submission handoff.

Connect local conversations

Use Connected agents to select the source, workspace and destination Project. Copy its configured command on the computer that owns the history. When reusing a Desktop account, use its configured OpenPond home with --home rather than assuming the CLI default has the same session. The official CLI distinguishes connection, upload and recurrence:

• Connect authorizes the destination and exits without uploading history.
• Upload is explicit, either through selected conversations or --upload.
• Continual imports default off. --continual enables the import schedule; it does not enable a Model learning policy.

Install the official CLI when needed:

These source commands connect only. Follow the source approval prompts and inspect the destination before choosing what to upload:

openpond import connect --source openclaw

For an explicit Codex bulk upload of the past week and daily future imports, the official command is:

Use your source setup's configured Project/workspace flags for a specific destination. For selected conversations, use the connection's selection flow rather than a bulk upload. OpenPond's own hosted Chat/Work activity uses the Pal source without a local command.

The source machine must be awake and connected. Scheduled imports run bounded jobs; missed runs while asleep are skipped. The next manual or scheduled job resumes retained work. Inspect coverage, errors and the last successful sync; an empty upload queue does not establish complete source coverage.

Choose graders and example review

For a conversation policy, open Sources & review:

• Select only the connections allowed to contribute.
• Review correctness, groundedness and conciseness rubrics, or compatible task-specific checks from the training configuration.
• Use Choose graders for explicit checked releases, or Auto-select for each run to choose from the Model's checked pool using bounded samples.
• Review the automatic selection model and limits. A cycle retains its chosen grader versions; auto-selection does not author arbitrary new graders.

Every required admission check must pass. Conciseness cannot compensate for failed correctness. Test graders on known successes, failures and ambiguous cases before relying on them.

Choose Pause this cycle for Inbox exceptions or Continue with independently verified examples. Both retain failed/blocked evidence. A proposed human or model correction must pass independent verification before it can train. Ordinary Model Tasks retain their separate human-admission requirement.

Choose the candidate decision

New conversation policies default to scheduled training with owner approval; existing policies preserve their saved authority. An optional Answer improvement model proposes corrections, which separate evaluation must verify.

Set timing and spending

• Ordinary nightly setup uses a local time and timezone; task-count and saved supported triggers retain their own behavior.
• Conversation From / Until / Timezone define a nightly start window. Started work may finish afterward; Run now starts immediately.
• Review minimum new examples, maximum tasks and GPU duration.
• Set maximum spending per run, day and month. Grading, improvements, verification and training count toward the saved limits.
• Review selection/replay and acceptance thresholds. After execution, inspect adaptation time and repeat rate as measured correction outcomes; they are separate from these settings.

Before enabling automatic activation, inspect the exact serving target and Save automatic learning authority. A failed runtime canary restores its recorded prior binding. Saving does not launch another manual cycle.

Check the saved policy

Confirm the selected Model, Dataset/source releases, graders, next check and timezone, ready/review/used counts, limits and candidate behavior. Follow retained cycle, training, Inbox and candidate links to inspect the evidence.

Pause future cycles prevents new cycles; Stop current cycle or the available iteration cancellation control acts on active work. Import recurrence is separate: disabling an import schedule does not stop an active job or revoke past evidence, and pausing Model learning does not stop collection.

For a new criterion, validate a new grader release. Keep prior runs pinned to their original settings.