openpond-sdk 0.12.18 → 0.12.19

package/TRAINING_WORKFLOWS.md

@@ -623,3 +623,18 @@ This read returns the owned execution's exact request, lifecycle phase and any r
 On a compatible host, `compileModificationExecution(input)` compiles a bounded raw form with the exact registered tokenizer. `prepareModificationExecution(input)` retains the compiled request, Project, name and maximum spend without starting compute. After explicit review, `startModificationExecution(input)` uses the returned execution ID, revision, plan hash, approved cap and a stable operation ID. `cancelModificationExecution(input)` retains that exact execution identity while cleanup proceeds.
 
 The CLI equivalents are `training hosted-compile-modification`, `hosted-prepare-modification`, `hosted-start-modification --yes` and `hosted-cancel-modification`, each taking `<input.json> --team-id <id>`. Native Python provides corresponding snake_case methods in both clients. Use the exported input schemas or installed help; preserve the original command and operation ID after an uncertain reply. Start is the funded execution boundary. An unavailable capability never grants permission to start an unsupported model.
+
+## Continual-learning grader selection
+
+An experiment draft's `trainingForm.learningIntent.graders` optionally stores
+exact learning revision references for recurring tasks. Omitting the field uses
+the reviewed batch's whole grader pool. An explicit subset must retain required
+gates and the training/evaluation roles needed by the hosted loop. Hosts validate
+these references against the exact reviewed batch before publishing the policy.
+
+The learning save hashes this selection and retains it in the immutable policy;
+it does not rewrite the manual experiment's batch or grader binding. The retained
+evaluation remains an independent candidate check. A nightly schedule uses the
+authored time and timezone, waits for approved unused examples, and respects the
+saved task threshold and spend limits. It does not itself enable conversation
+collection, automatic data approval, or automatic serving.

package/dist/training.js

@@ -3401,12 +3401,16 @@ var EngramExperimentFormDraftSchema = z33.object({
 }).strict();
 
 // src/training-experiment-setup.ts
-import { LearningRevisionRefSchema } from "@openpond/evals/learning";
+import { LearningRevisionRefSchema as LearningRevisionRefSchema2 } from "@openpond/evals/learning";
 
 // src/training-experiment-learning-intent.ts
 import { z as z34 } from "zod";
+import { LearningRevisionRefSchema } from "@openpond/evals/learning";
 var TrainingExperimentLearningIntentSchema = z34.object({
   enabled: z34.boolean(),
+  /** Applies only after an accepted compatible GRPO checkpoint exists. */
+  continuationAdapterMode: z34.enum(["shared", "residual_append", "sequential_merge"]).optional(),
+  graders: z34.array(LearningRevisionRefSchema).max(100).optional(),
   mode: z34.enum(["nightly", "approved_count"]),
   minimumTasks: z34.string().max(100),
   localTime: z34.string().max(100),
@@ -3501,7 +3505,7 @@ async function parseAndVerifyExperimentStartingModel(raw, expected = {}) {
 // src/training-experiment-setup.ts
 var ExperimentOperationTypeSchema = z36.enum(["evaluate", "sft", "grpo", "neuron_steering", "inspect_activations", "engram_adapter", "decision_model", "alphaedit", "selective_neuron_training"]);
 function experimentOperationLabel(type) {
-  return { evaluate: "Evaluate", sft: "SFT + LoRA", grpo: "RL (GRPO) + LoRA", neuron_steering: "Neuron steering", inspect_activations: "Inspect activations", engram_adapter: "Engram memory", decision_model: "Decision model", alphaedit: "AlphaEdit", selective_neuron_training: "Neuron training" }[type];
+  return { evaluate: "Evaluate", sft: "SFT", grpo: "RL", neuron_steering: "Neuron steering", inspect_activations: "Inspect activations", engram_adapter: "Engram memory", decision_model: "Decision model", alphaedit: "AlphaEdit", selective_neuron_training: "Neuron training" }[type];
 }
 var TrainingExperimentSetupRequestSchema = TrainingPreparationRequestSchema.omit({ configuration: true, origin: true }).extend({
   schemaVersion: z36.literal("openpond.trainingExperimentSetupRequest.v1"),
@@ -3585,7 +3589,7 @@ var TrainingExperimentFormDraftSchema = z36.object({
     inputHash: z36.string().regex(/^[a-f0-9]{64}$/),
     trainingInputHash: z36.string().regex(/^[a-f0-9]{64}$/),
     configuration: HostedModelProjectSummarySchema,
-    policy: LearningRevisionRefSchema
+    policy: LearningRevisionRefSchema2
   }).strict().optional()
 }).strict();
 var ExperimentDraftSchema = z36.object({
@@ -5985,7 +5989,7 @@ import { z as z67 } from "zod";
 
 // src/training-experiment-learning.ts
 import { z as z51 } from "zod";
-import { LearningRevisionRefSchema as LearningRevisionRefSchema2 } from "@openpond/evals/learning";
+import { LearningRevisionRefSchema as LearningRevisionRefSchema3 } from "@openpond/evals/learning";
 var TrainingExperimentLearningRequestSchema = z51.object({
   schemaVersion: z51.literal("openpond.trainingExperimentLearning.v1"),
   teamId: z51.string().min(1).max(191),
@@ -5993,13 +5997,13 @@ var TrainingExperimentLearningRequestSchema = z51.object({
   draftId: z51.string().min(1).max(191),
   form: TrainingExperimentFormDraftSchema.omit({ learningConfiguration: true }),
   expectedConfiguration: TrainingPreparationRequestSchema.shape.configuration.optional(),
-  expectedPolicy: LearningRevisionRefSchema2.optional()
+  expectedPolicy: LearningRevisionRefSchema3.optional()
 }).strict();
 var TrainingExperimentLearningReceiptSchema = z51.object({
   schemaVersion: z51.literal("openpond.trainingExperimentLearningReceipt.v1"),
   inputHash: z51.string().regex(/^[a-f0-9]{64}$/),
   configuration: HostedModelProjectSummarySchema,
-  policy: LearningRevisionRefSchema2
+  policy: LearningRevisionRefSchema3
 }).strict();
 function trainingExperimentLearningRequest(input) {
   const { learningConfiguration, ...form } = input.form;
@@ -6054,7 +6058,8 @@ function trainingExperimentLearningSettings(form) {
     baseModel: form.baseModel,
     recipe: { ...form.recipe, dataset: { ...form.recipe.dataset && typeof form.recipe.dataset === "object" ? form.recipe.dataset : {}, validationSplit: form.validationSplit } },
     evaluationTasksetRef: form.evaluation,
-    managedRolloutPlacement: "remote"
+    managedRolloutPlacement: "remote",
+    ...form.recipe.method === "grpo" && form.learningIntent?.continuationAdapterMode ? { continuationAdapterMode: form.learningIntent.continuationAdapterMode } : {}
   });
 }
 function createTrainingExperimentLearningClient(request) {