Neuron training
Preview · Requires a qualified exact model and runtime. A draft, guide or saved configuration does not establish execution or serving availability.
Learn persistent weight deltas for selected neurons.
Neuron training learns deltas to selected MLP down-projection columns while keeping the original base checkpoint frozen. Applying the saved delta changes the effective weights. It is actual weight training, distinct from activation inspection and a fixed inference-time steering dose.
For example, teach a small set of selected neurons a product-answer convention using reviewed target examples, while measuring their gradients on separate background domains. Evaluate new questions and unrelated domains afterward.
Set up the experiment
Choose New → Neuron training, or use Train selected neurons from a completed compatible hosted inspection.
| Tab | What to provide |
|---|---|
| Model | Exact compatible dense BF16 Qwen3-8B starting checkpoint and name. |
| Data | Reviewed target demonstrations, separately named background-domain batches and optional preservation data. |
| Settings | Selected neuron count, learning rate, optimizer steps and sequence limit. Advanced settings include selection criteria, explicit coordinates and preservation controls. |
| Review | Exact data/selection recipe, runtime, limits and retained preparation. |
The training batches use approved supervised demonstrations; activation values
alone are not desired answers. Keep target and background roles explicit.
Automatic selection measures target/background behavior. Alternatively enter
explicit layer:neuron coordinates, one per line; their distinct count must
match the requested count. A prior inspection supplies coordinates, not a
replacement for reviewed training data.
When using preservation, provide its batch and enable its preservation weight and examples-per-step together. The current defaults leave preservation loss disabled; merely selecting data does not mean the loss was optimized.
Read the result
Inspect actual selection measurements, selected coordinates, optimizer updates, learned column deltas and fresh reload evidence. The original checkpoint should remain unchanged and gradients should stay confined to the selected training parameters. A successful run does not demonstrate that the neurons have one human-readable purpose.
The saved artifact contains the selected-coordinate delta. It supports compatible inference; continuing training from an already saved neuron delta is currently unsupported. Do not confuse its inference artifact with resumable optimizer state.
Review, run and recover
Use Review → Prepare review to resolve the exact inputs, compatible runtime and maximum spend. Inspect the retained preparation before approving Start. Opening the editor, changing tabs and saving a draft do not start paid compute. Preparing a run is separate from starting it.
If a response is lost, reopen the same experiment and retained execution before retrying. Preserve its operation identity and revision; creating a second run can duplicate work and charges. Use the existing refresh, retry or cancellation controls. Cancellation is a request: follow the execution through cleanup and inspect remaining capacity and funding before treating it as finished.
Runtime discovery is authoritative. A selectable action or saved configuration does not prove that a checkpoint/image/GPU combination is qualified. Unsupported combinations cannot start simply because sufficient credits are available.
Save and serve
After a successful execution with verified artifacts, choose Save Model. The saved version retains its exact base, modification and artifact identity. Saving is separate from allocating serving capacity. Open the saved Model, review compatible serving and approve activation explicitly. Use Stop and verify released capacity when finished.
Measure a baseline and candidate on the same held-out tasks and grader versions. Training loss, successful export and a working endpoint do not establish better answers. Keep runtime correctness and task quality as separate checks.
API, SDK and harness
Use compileNeuronTrainingExecution, prepareNeuronTrainingExecution and
startNeuronTrainingExecution. CLI commands are training hosted-compile-neuron-training, hosted-prepare-neuron-training and
hosted-start-neuron-training; read hosted-neuron-training-execution, save
with hosted-save-neuron-training-model, and control serving with
hosted-control-neuron-training-serving.
Use the official CLI with your configured account and workspace. Run
openpond training --help and the selected command's help for the current
request schema; start from the SDK's typed draft/request builders rather than
inventing an unversioned JSON payload. Web, API, TypeScript/Python SDKs and the
CLI retain the same preparation, execution and Model identities. Paid Start and
serving activation require explicit approval; a harness should retain these
identities for recovery.