What you can do
Train, modify, evaluate and serve models, and configure supported continual-learning policies.
OpenPond is a lab for model training, adapters and continual learning. Ponder Training connects reviewed data, training runs, independent evaluation, saved Model versions and serving. Choose a workflow, then the adapter and starting model for the behavior you want to improve.
Choose a workflow and adapter
| Workflow | Current adapters | Purpose |
|---|---|---|
| SFT | LoRA, Engram, Neuron training | Learn from reviewed answer examples through the selected adapter's trainer. |
| RL | LoRA | Learn from rewards using GRPO. |
| Decision Model | Specialized LoRA | Train or score allowed choices from structured state. |
See the adapter support matrix for supported and unsupported pairs and exact model/runtime requirements. Workflow and adapter choices can be saved as an authoring draft. Unsupported pairs cannot execute; selecting an adapter does not establish runtime availability.
Continual learning
| Workflow / adapter | Continual learning | Manual parent continuation |
|---|---|---|
| SFT / LoRA | Yes | Generic cross-job checkpoint continuation is unavailable. |
| SFT / Engram | No | Exact compatible memory module with a fresh optimizer. |
| SFT / Neuron training | No | Unavailable; saved patches support inference. |
| RL / LoRA | Yes | Shared, Stacked or Merged LoRA under exact-parent constraints. |
| Decision Model / LoRA | No | Exact compatible adapter with a fresh optimizer. |
Scheduled learning, manual continuation and interrupted-run recovery are separate capabilities. A schedule does not authorize candidate acceptance or serving. Continual learning explains those boundaries.
The model workflow
| Object | What it holds |
|---|---|
| Dataset | Versioned tasks, references, provenance, graders and train/holdout roles. |
| Experiment | A configured test and its retained evidence; training history remains under Models → Runs. |
| Workflow | SFT, RL or Decision Model: the learning approach or scoring task. |
| Adapter | LoRA, Engram or Neuron training: the parameters or added memory that learn, under their exact compatibility rules. |
| Model | Saved versions with exact artifacts, lineage and a canonical learning policy. |
| Continual learning | Scheduled checks and supported updates from eligible new evidence, with budgets, independent comparison and candidate decisions. |
| Serving | A separately authorized deployment of a compatible qualified version. |
SFT learns from reviewed demonstrations; GRPO learns from rewards on fresh attempts. An adapter's availability depends on the exact model and runtime. Continual learning repeats only supported training paths; a visible method does not imply that it can run on a schedule.
A new Dataset row or conversation does not change model weights. Ordinary Model evidence needs explicit human admission. Conversation policies can admit independently verified examples under their saved rules, routing exceptions to Inbox. Candidate acceptance and serving activation remain separate from example approval.
Start with a measured workflow
- Create or import a Dataset, retain its provenance and separate training from held-out examples.
- Validate the Graders and run a baseline Experiment.
- Choose training or a compatible modification based on the failure. Review exact inputs, runtime and spending before execution.
- Compare the candidate on independent tasks and save the versions you accept.
- Serve a qualified version, or configure a supported continual-learning policy to check for future eligible batches.
Workspaces and Projects
A workspace sets membership and access. A Console Project groups Models, Datasets, graders, experiments and training resources. Choose your workspace before opening a resource; use the Project picker to narrow the collection. Grouping resources does not change their versions or publish them.
Tools around the model workflow
Use the SDK for programmatic integration, Evals for portable task and scoring contracts, and Harness for versioned agent instructions and tools. Desktop works with local projects and evidence; Cloud provides hosted inference, compute, storage and services.
Pal is the assistant available in OpenPond's Chat and Work surfaces. A Pal conversation and Harness refinement have their own tools and permissions; neither automatically trains a Ponder model. Connected agent activity can supply reviewed evidence for a separately configured learning workflow.
Next: Choose your starting point. Check availability for exact execution and account requirements.