AlphaEdit

AlphaEdit applies constrained updates to selected original weight matrices. It edits associations while using preservation statistics to limit other changes. The immutable base remains the reference, and a saved Model installs the exact matrix changes and sequential editing state. This is distinct from adding a prompt rule or training a LoRA adapter.

For example, change a fictional product association from its old category to a new category, then test paraphrases, neighboring products and unrelated text. Do not use a single successful rewrite as evidence of reliable factual editing.

Prepare two distinct data roles

The current dense BF16 Qwen3-8B path requires sealed reviewed supervised batches for edits and preservation. Each task has one approved demonstration matching its expected output in the training split.

An edit input is exactly:

Its expected output is exactly {"text":"desk lighting"}. The prompt template has exactly one {} placeholder, where the subject is inserted. The approved demonstration is the JSON representation of that complete object. After validating the reviewed object, AlphaEdit uses its literal text value as the numerical target; it does not train on the JSON envelope.

A preservation input is exactly {"text":"A separate passage about a different subject."}. Its expected output is {"text":"A separate passage about a different subject."}; the output's text must equal the input's literal text. Use distinct preservation examples, not the desired rewrite target. Dataset instructions are explicitly prepended by the projection; edit instructions cannot contain braces that would conflict with the single-placeholder grammar. Referenced assets or hidden resources are not accepted as implicit training inputs.

Set up the experiment

Choose New → AlphaEdit in Experiments.

Layer choices, preservation statistics and template shape are part of the method, not interchangeable prompt options. Start with a small edit and inspect actual numerical evidence before accumulating sequential edits.

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.

Read the evidence

Results separate latent optimization loss from measured starting, edited and fresh-reload target negative log likelihood (NLL), preservation next-token KL, artifact hashes and reload tolerance. These use the reviewed inputs. They do not replace held-out paraphrase, nearby-fact and unrelated-task evaluation.

Compatible continuation requires the exact saved matrices and sequential editing state with matching preservation provenance. A weight file alone is not the complete continuation authority.

API, SDK and harness

Use compileAlphaEditExecution, prepareAlphaEditExecution and startAlphaEditExecution. CLI commands are training hosted-compile-alphaedit, hosted-prepare-alphaedit, hosted-start-alphaedit and hosted-alphaedit-execution. Save with hosted-save-alphaedit-model and control compatible serving with hosted-control-alphaedit-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.