Model modifications
Preview · Execution and serving require a qualified combination of checkpoint, modification and runtime. The new modification runtimes are not yet qualified; a saved configuration does not make them available.
Configure weight training, activation experiments, conditional memory and decision models.
Open Experiments, choose New experiment, then choose a Modification. Each card has a separate Read guide link. Follow Model → Data / Prompts / Questions → Settings → Review to select the starting model and configure that operation. The input tab changes with the method; prompts stay in the current editor rather than opening a new chat. You can also start from an existing Model. Opening the editor does not start a chat or paid compute.
Choose what changes
| Modification | What changes | Example |
|---|---|---|
| SFT with LoRA | Trains adapter weights from reviewed answers; the original checkpoint stays frozen. | Teach a support model your answer format. |
| RL / GRPO | Updates a policy adapter using rewards over groups of generated responses. | Reward correct answers with a code verifier. |
| Inspect activations | Records selected internal activations without a weight update. | Inspect which neurons activate at each prompt token. |
| Neuron steering | Applies selected activation doses during inference; checkpoint weights stay unchanged. | Compare the same prompt under Off, zero and signed doses. |
| Neuron training | Learns deltas to selected MLP output-weight columns against an immutable base. | Adapt selected neurons using target and separate background domains. |
| Engram adapter | Trains added n-gram memory tables, projections and gates over a frozen base. | Learn recurring product names and domain terminology. |
| Decision model | Trains an adapter for defined decisions and scores every allowed choice. | Return probabilities for refund, troubleshoot or escalate. |
| AlphaEdit | Applies constrained updates to selected original weight matrices. | Edit a factual association and inspect preservation and reload evidence. |
SFT is a training objective; LoRA is a parameterization. Inspection, steering, learned adapters and weight edits produce different kinds of evidence and artifacts. LoReFT is deferred and is not an available execution option.
Prepare and run
Use the existing Dataset review/import controls for training data. Keep each role explicit: desired target answers, background-domain examples, preservation inputs and held-out evaluation tasks are not interchangeable.
Review the compiled inputs, compatible runtime, compute and maximum spend, then start the retained preparation explicitly. Retain the original operation, plan and revision after a lost response. Reopen the same execution before retrying; an uncertain response is not permission to launch another optimizer.
Inspection has its own results. From a completed compatible hosted inspection, select actual Off-observed neurons and choose Train selected neurons. This opens the same editor with the exact inspected model, result and coordinates. Choose reviewed target/background data before starting training. Imported external observations do not grant hosted training authority.
AlphaEdit uses reviewed {subject,promptTemplate} inputs with one {} subject
placeholder, reviewed {text: string} targets and a separate literal-text
preservation batch.
Its results separate latent optimization losses from measured starting,
edited and fresh-reload target NLL, preservation next-token KL and reload
diagnostics. Those scores use the reviewed inputs, not a held-out evaluation.
Neuron training retains selection measurements, selected coordinates and the actual learned column delta. An already saved neuron delta supports inference; training continuation from that delta is currently unsupported. AlphaEdit can continue from an exact compatible saved parent and its sequential editing state.
Save a Model and serve it
Open a completed result and use Save Model. The saved version retains its exact base, modification and artifact identity. Open it in Models to inspect its source and serving state, then review and activate compatible serving separately. Saving does not allocate capacity.
Select that saved version in Desktop chat when using its supported serving path. A saved weight Model always installs its immutable patch; select the unmodified base separately for a baseline. Decision scoring returns typed candidate distributions rather than generated chat JSON.
Use Stop and inspect outstanding requests and cleanup until capacity is actually released. Training completion, successful tensor upload, settled billing and released provider resources are separate facts.
API, SDK and harness
The UI and headless clients use the same preparation, execution, history and saved-Model operations. The official CLI and its maintained skill expose the workflow to your harness; use the running app's configured account/home.
| Work | TypeScript SDK | Official CLI |
|---|---|---|
| Activation/steering | compileModificationExecution, prepareModificationExecution, startModificationExecution | training hosted-compile-modification, hosted-prepare-modification, hosted-start-modification |
| Engram | compileEngramExecution, prepareEngramExecution, startEngramExecution | training hosted-compile-engram, hosted-prepare-engram, hosted-start-engram |
| Decision | compileDecisionExecution, prepareDecisionExecution, scoreDecisionModel | training hosted-compile-decision, hosted-prepare-decision, hosted-score-decision |
| AlphaEdit | compileAlphaEditExecution, prepareAlphaEditExecution, startAlphaEditExecution | training hosted-compile-alphaedit, hosted-prepare-alphaedit, hosted-start-alphaedit |
| Neuron training | compileNeuronTrainingExecution, prepareNeuronTrainingExecution, startNeuronTrainingExecution | training hosted-compile-neuron-training, hosted-prepare-neuron-training, hosted-start-neuron-training |
Read openpond training --help for exact JSON inputs, scoped capability,
history, cancellation, save and serving commands. Supply --team-id and the
configured --home when reusing the app account. Paid Start and serving
activation require --yes. Python exposes matching typed sync/async operations.
Capability discovery is authoritative for the exact model and runtime. Current AlphaEdit and neuron-training implementations target dense BF16 Qwen3-8B; Decision uses its separate Nimble/Qwen3.5 source and tokenizer. Quantized bases, arbitrary model sizes and arbitrary modification stacking are not implied. Source implementation and package checks do not establish GPU compatibility or model quality. Keep target/paraphrase/nearby/unrelated evaluation separate from training measurements before deciding to use a modified Model.