openpond-sdk 0.9.6 → 0.9.7
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/dist/{chunk-EMANJDJR.js → chunk-CNGJF6AM.js} +61 -198
- package/dist/chunk-CNGJF6AM.js.map +7 -0
- package/dist/chunk-SR55WVPK.js +158 -0
- package/dist/chunk-SR55WVPK.js.map +7 -0
- package/dist/dataset-marketplace.js +216 -0
- package/dist/dataset-marketplace.js.map +7 -0
- package/dist/dataset-workspaces.js +86 -2
- package/dist/dataset-workspaces.js.map +4 -4
- package/dist/experiments.js +9 -7
- package/dist/experiments.js.map +1 -1
- package/dist/index.js +9 -8
- package/dist/index.js.map +1 -1
- package/dist/types/packages/sdk/src/dataset-marketplace.d.ts +945 -0
- package/dist/types/packages/sdk/src/dataset-marketplace.d.ts.map +1 -0
- package/dist/types/packages/sdk/src/dataset-population.d.ts +97 -0
- package/dist/types/packages/sdk/src/dataset-population.d.ts.map +1 -0
- package/dist/types/packages/sdk/src/dataset-workspace-client.d.ts +146 -145
- package/dist/types/packages/sdk/src/dataset-workspace-client.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/dataset-workspace-contracts.d.ts +84 -84
- package/dist/types/packages/sdk/src/dataset-workspaces.d.ts +1 -0
- package/dist/types/packages/sdk/src/dataset-workspaces.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/evaluation-project-context.d.ts +1 -1
- package/dist/types/packages/sdk/src/experiment-client.d.ts +23 -23
- package/dist/types/packages/sdk/src/experiment-contracts.d.ts +3 -3
- package/dist/types/packages/sdk/src/experiment-grading-contracts.d.ts +1 -1
- package/dist/types/packages/sdk/src/experiment-scoring-contracts.d.ts +3 -3
- package/dist/types/packages/sdk/src/grader-inspection.d.ts +8 -8
- package/dist/types/packages/sdk/src/learning-batch-assets.d.ts +2 -2
- package/dist/types/packages/sdk/src/learning-client.d.ts +1 -1
- package/dist/types/packages/sdk/src/managed-training-preparation.d.ts +2 -2
- package/dist/types/packages/sdk/src/model-batch-review-preparation.d.ts +7 -7
- package/dist/types/packages/sdk/src/model-batch-review.d.ts +23 -23
- package/dist/types/packages/sdk/src/model-learning-policy.d.ts +3 -3
- package/dist/types/packages/sdk/src/model-projects.d.ts +27 -27
- package/dist/types/packages/sdk/src/model-starter-attempts.d.ts +47 -47
- package/dist/types/packages/sdk/src/model-starter-catalog.d.ts +15 -15
- package/dist/types/packages/sdk/src/model-starter-execution.d.ts +4 -4
- package/dist/types/packages/sdk/src/model-starters.d.ts +18 -18
- package/dist/types/packages/sdk/src/model-taskset-runs-contracts.d.ts +16 -16
- package/dist/types/packages/sdk/src/model-taskset-runs.d.ts +15 -15
- package/dist/types/packages/sdk/src/task-inventory.d.ts +4 -4
- package/dist/types/packages/sdk/src/taskset-authored-contracts.d.ts +46 -46
- package/dist/types/packages/sdk/src/taskset-authored-files.d.ts +30 -30
- package/dist/types/packages/sdk/src/taskset-authored-portable-release.d.ts +2 -2
- package/dist/types/packages/sdk/src/taskset-draft-api-contracts.d.ts +84 -84
- package/dist/types/packages/sdk/src/taskset-draft-client.d.ts +176 -176
- package/dist/types/packages/sdk/src/taskset-draft-core.d.ts +46 -46
- package/dist/types/packages/sdk/src/taskset-draft-dataset-artifacts.d.ts +25 -25
- package/dist/types/packages/sdk/src/taskset-draft-dataset-sources.d.ts +41 -41
- package/dist/types/packages/sdk/src/taskset-draft-document.d.ts +45 -45
- package/dist/types/packages/sdk/src/taskset-draft-package-compiler.d.ts +56 -56
- package/dist/types/packages/sdk/src/taskset-draft-source.d.ts +71 -71
- package/dist/types/packages/sdk/src/taskset-draft-workspace.d.ts +84 -84
- package/dist/types/packages/sdk/src/taskset-package-client.d.ts +1733 -1733
- package/dist/types/packages/sdk/src/taskset-package-contracts.d.ts +1083 -1083
- package/dist/types/packages/sdk/src/taskset-package-grader-projection.d.ts +2 -2
- package/dist/types/packages/sdk/src/taskset-package-learning.d.ts +18 -18
- package/dist/types/packages/sdk/src/taskset-package-projection.d.ts +59 -59
- package/dist/types/packages/sdk/src/training-acceptance.d.ts +4 -4
- package/dist/types/packages/sdk/src/training-harness-sources.d.ts +1 -1
- package/dist/types/packages/sdk/src/training-learning-batch.d.ts +68 -68
- package/dist/types/packages/sdk/src/training-project-client.d.ts +12 -12
- package/dist/types/packages/sdk/src/training-project-contracts.d.ts +21 -21
- package/dist/types/packages/sdk/src/training.d.ts +24 -24
- package/package.json +5 -1
- package/dist/chunk-EMANJDJR.js.map +0 -7
|
@@ -12,7 +12,7 @@ export declare function importedPackageGraders(value: TasksetPackage): ({
|
|
|
12
12
|
rewardEligible: boolean;
|
|
13
13
|
privileged: boolean;
|
|
14
14
|
metadata: Record<string, unknown>;
|
|
15
|
-
kind: "file" | "content" | "
|
|
15
|
+
kind: "file" | "content" | "state" | "test" | "schema" | "diff" | "runtime_event";
|
|
16
16
|
config: Record<string, unknown>;
|
|
17
17
|
} | {
|
|
18
18
|
id: string;
|
|
@@ -30,7 +30,7 @@ export declare function importedPackageGraders(value: TasksetPackage): ({
|
|
|
30
30
|
modelId: string;
|
|
31
31
|
};
|
|
32
32
|
calibrationFixtureRefs: string[];
|
|
33
|
-
calibrationStatus: "
|
|
33
|
+
calibrationStatus: "passed" | "failed" | "pending";
|
|
34
34
|
temperature: number;
|
|
35
35
|
} | {
|
|
36
36
|
id: string;
|
|
@@ -467,7 +467,7 @@ export declare function validateTasksetLearningResources(taskset: TasksetRelease
|
|
|
467
467
|
};
|
|
468
468
|
familyKey: string;
|
|
469
469
|
inputHash: string;
|
|
470
|
-
split: "train" | "validation" | "
|
|
470
|
+
split: "train" | "validation" | "frozen_eval" | "test";
|
|
471
471
|
}[];
|
|
472
472
|
sealedAt: string;
|
|
473
473
|
sealedBy: string;
|
|
@@ -495,7 +495,7 @@ export declare function validateTasksetLearningResources(taskset: TasksetRelease
|
|
|
495
495
|
attemptId: string;
|
|
496
496
|
occurredAt: string;
|
|
497
497
|
familyKey: string | null;
|
|
498
|
-
split: "train" | "validation" | "
|
|
498
|
+
split: "train" | "validation" | "frozen_eval" | "test";
|
|
499
499
|
input: Record<string, z.core.util.JSONType>;
|
|
500
500
|
observedOutput: Record<string, z.core.util.JSONType> | null;
|
|
501
501
|
expected: Record<string, z.core.util.JSONType> | null;
|
|
@@ -560,9 +560,9 @@ export declare function validateTasksetLearningResources(taskset: TasksetRelease
|
|
|
560
560
|
} | null;
|
|
561
561
|
};
|
|
562
562
|
evidenceValidity: "pending" | "valid" | "invalid";
|
|
563
|
-
taskAdmissibility: "
|
|
564
|
-
observedQuality: "
|
|
565
|
-
targetApproval: "
|
|
563
|
+
taskAdmissibility: "pending" | "rejected" | "approved";
|
|
564
|
+
observedQuality: "passed" | "failed" | "unavailable" | "unscored";
|
|
565
|
+
targetApproval: "pending" | "rejected" | "approved" | "not_required";
|
|
566
566
|
approvedTarget: Record<string, z.core.util.JSONType> | null;
|
|
567
567
|
grade: {
|
|
568
568
|
schemaVersion: "openpond.rewardComposition.v1";
|
|
@@ -580,8 +580,8 @@ export declare function validateTasksetLearningResources(taskset: TasksetRelease
|
|
|
580
580
|
contentHash: string;
|
|
581
581
|
revision: number;
|
|
582
582
|
};
|
|
583
|
-
role: "
|
|
584
|
-
status: "failed" | "
|
|
583
|
+
role: "training" | "evaluation";
|
|
584
|
+
status: "failed" | "unavailable" | "scored" | "pending";
|
|
585
585
|
rawScore: number | null;
|
|
586
586
|
normalizedScore: number | null;
|
|
587
587
|
passed: boolean | null;
|
|
@@ -594,7 +594,7 @@ export declare function validateTasksetLearningResources(taskset: TasksetRelease
|
|
|
594
594
|
score: number | null;
|
|
595
595
|
passed: boolean;
|
|
596
596
|
rewardEligible: boolean;
|
|
597
|
-
failureClass: "
|
|
597
|
+
failureClass: "policy_failure" | "grader_failure" | "environment_failure" | "infrastructure_failure" | "timeout" | "cancelled" | null;
|
|
598
598
|
feedback: string[];
|
|
599
599
|
visibleEvidenceRefs: string[];
|
|
600
600
|
privilegedEvidenceRefs: string[];
|
|
@@ -614,12 +614,12 @@ export declare function validateTasksetLearningResources(taskset: TasksetRelease
|
|
|
614
614
|
} | undefined;
|
|
615
615
|
}[];
|
|
616
616
|
training: {
|
|
617
|
-
status: "
|
|
617
|
+
status: "scored" | "unscorable" | "not_configured";
|
|
618
618
|
score: number | null;
|
|
619
619
|
passed: boolean | null;
|
|
620
620
|
};
|
|
621
621
|
evaluation: {
|
|
622
|
-
status: "
|
|
622
|
+
status: "scored" | "unscorable" | "not_configured";
|
|
623
623
|
score: number | null;
|
|
624
624
|
passed: boolean | null;
|
|
625
625
|
};
|
|
@@ -641,8 +641,8 @@ export declare function validateTasksetLearningResources(taskset: TasksetRelease
|
|
|
641
641
|
contentHash: string;
|
|
642
642
|
revision: number;
|
|
643
643
|
};
|
|
644
|
-
role: "
|
|
645
|
-
status: "failed" | "
|
|
644
|
+
role: "training" | "evaluation";
|
|
645
|
+
status: "failed" | "unavailable" | "scored" | "pending";
|
|
646
646
|
rawScore: number | null;
|
|
647
647
|
normalizedScore: number | null;
|
|
648
648
|
passed: boolean | null;
|
|
@@ -655,7 +655,7 @@ export declare function validateTasksetLearningResources(taskset: TasksetRelease
|
|
|
655
655
|
score: number | null;
|
|
656
656
|
passed: boolean;
|
|
657
657
|
rewardEligible: boolean;
|
|
658
|
-
failureClass: "
|
|
658
|
+
failureClass: "policy_failure" | "grader_failure" | "environment_failure" | "infrastructure_failure" | "timeout" | "cancelled" | null;
|
|
659
659
|
feedback: string[];
|
|
660
660
|
visibleEvidenceRefs: string[];
|
|
661
661
|
privilegedEvidenceRefs: string[];
|
|
@@ -675,12 +675,12 @@ export declare function validateTasksetLearningResources(taskset: TasksetRelease
|
|
|
675
675
|
} | undefined;
|
|
676
676
|
}[];
|
|
677
677
|
training: {
|
|
678
|
-
status: "
|
|
678
|
+
status: "scored" | "unscorable" | "not_configured";
|
|
679
679
|
score: number | null;
|
|
680
680
|
passed: boolean | null;
|
|
681
681
|
};
|
|
682
682
|
evaluation: {
|
|
683
|
-
status: "
|
|
683
|
+
status: "scored" | "unscorable" | "not_configured";
|
|
684
684
|
score: number | null;
|
|
685
685
|
passed: boolean | null;
|
|
686
686
|
};
|
|
@@ -695,14 +695,14 @@ export declare function validateTasksetLearningResources(taskset: TasksetRelease
|
|
|
695
695
|
id: string;
|
|
696
696
|
revision: number;
|
|
697
697
|
name: string;
|
|
698
|
-
kind: "
|
|
698
|
+
kind: "provider" | "benchmark" | "work" | "direct" | "opentelemetry";
|
|
699
699
|
taskDefinition: {
|
|
700
700
|
id: string;
|
|
701
701
|
contentHash: string;
|
|
702
702
|
revision: number;
|
|
703
703
|
};
|
|
704
704
|
enabled: boolean;
|
|
705
|
-
allowedSplits: ("train" | "validation" | "
|
|
705
|
+
allowedSplits: ("train" | "validation" | "frozen_eval" | "test")[];
|
|
706
706
|
mapping: {
|
|
707
707
|
schemaVersion: "openpond.taskSourceMapping.v1";
|
|
708
708
|
exampleId: string;
|
|
@@ -713,7 +713,7 @@ export declare function validateTasksetLearningResources(taskset: TasksetRelease
|
|
|
713
713
|
expected: string | null;
|
|
714
714
|
evaluatorContext: string | null;
|
|
715
715
|
familyKey: string;
|
|
716
|
-
split: "train" | "validation" | "
|
|
716
|
+
split: "train" | "validation" | "frozen_eval" | "test";
|
|
717
717
|
} | null;
|
|
718
718
|
adapterVersion: string | null;
|
|
719
719
|
contentHash: string;
|
|
@@ -46,7 +46,7 @@ export declare function prepareImportedTasksetPackage(input: {
|
|
|
46
46
|
tasks: {
|
|
47
47
|
id: string;
|
|
48
48
|
clusterKey: string;
|
|
49
|
-
split: "train" | "validation" | "
|
|
49
|
+
split: "train" | "validation" | "frozen_eval" | "test";
|
|
50
50
|
input: Record<string, unknown>;
|
|
51
51
|
expectedOutput: Record<string, unknown> | null;
|
|
52
52
|
policyVisibleContext: Record<string, unknown>;
|
|
@@ -82,7 +82,7 @@ export declare function prepareImportedTasksetPackage(input: {
|
|
|
82
82
|
hardGate: boolean;
|
|
83
83
|
rewardEligible: boolean;
|
|
84
84
|
privileged: boolean;
|
|
85
|
-
kind: "content" | "
|
|
85
|
+
kind: "content" | "state" | "schema" | "runtime_event" | "artifact";
|
|
86
86
|
config: Record<string, unknown>;
|
|
87
87
|
} | {
|
|
88
88
|
id: string;
|
|
@@ -100,7 +100,7 @@ export declare function prepareImportedTasksetPackage(input: {
|
|
|
100
100
|
mediaType: string;
|
|
101
101
|
visibility: "policy" | "verifier" | "host_private";
|
|
102
102
|
};
|
|
103
|
-
calibrationStatus: "
|
|
103
|
+
calibrationStatus: "passed" | "failed" | "pending";
|
|
104
104
|
model?: {
|
|
105
105
|
providerId: string;
|
|
106
106
|
modelId: string;
|
|
@@ -227,7 +227,7 @@ export declare function prepareImportedTasksetPackage(input: {
|
|
|
227
227
|
hardGate: boolean;
|
|
228
228
|
rewardEligible: boolean;
|
|
229
229
|
privileged: boolean;
|
|
230
|
-
kind: "content" | "
|
|
230
|
+
kind: "content" | "state" | "schema" | "runtime_event" | "artifact";
|
|
231
231
|
config: Record<string, unknown>;
|
|
232
232
|
} | {
|
|
233
233
|
id: string;
|
|
@@ -245,7 +245,7 @@ export declare function prepareImportedTasksetPackage(input: {
|
|
|
245
245
|
mediaType: string;
|
|
246
246
|
visibility: "policy" | "verifier" | "host_private";
|
|
247
247
|
};
|
|
248
|
-
calibrationStatus: "
|
|
248
|
+
calibrationStatus: "passed" | "failed" | "pending";
|
|
249
249
|
model?: {
|
|
250
250
|
providerId: string;
|
|
251
251
|
modelId: string;
|
|
@@ -414,7 +414,7 @@ export declare function prepareImportedTasksetPackage(input: {
|
|
|
414
414
|
name: string;
|
|
415
415
|
description: string;
|
|
416
416
|
implementation: {
|
|
417
|
-
kind: "content" | "
|
|
417
|
+
kind: "content" | "state" | "schema" | "runtime_event" | "artifact";
|
|
418
418
|
config: Record<string, unknown>;
|
|
419
419
|
} | {
|
|
420
420
|
kind: "model_judge";
|
|
@@ -426,7 +426,7 @@ export declare function prepareImportedTasksetPackage(input: {
|
|
|
426
426
|
mediaType: string;
|
|
427
427
|
visibility: "policy" | "verifier" | "host_private";
|
|
428
428
|
};
|
|
429
|
-
calibrationStatus: "
|
|
429
|
+
calibrationStatus: "passed" | "failed" | "pending";
|
|
430
430
|
model?: {
|
|
431
431
|
providerId: string;
|
|
432
432
|
modelId: string;
|
|
@@ -512,7 +512,7 @@ export declare function prepareImportedTasksetPackage(input: {
|
|
|
512
512
|
contentHash: string;
|
|
513
513
|
revision: number;
|
|
514
514
|
};
|
|
515
|
-
role: "
|
|
515
|
+
role: "training" | "evaluation";
|
|
516
516
|
normalization: {
|
|
517
517
|
kind: "identity";
|
|
518
518
|
} | {
|
|
@@ -599,7 +599,7 @@ export declare function prepareImportedTasksetPackage(input: {
|
|
|
599
599
|
hardGate: boolean;
|
|
600
600
|
rewardEligible: boolean;
|
|
601
601
|
privileged: boolean;
|
|
602
|
-
kind: "content" | "
|
|
602
|
+
kind: "content" | "state" | "schema" | "runtime_event" | "artifact";
|
|
603
603
|
config: Record<string, unknown>;
|
|
604
604
|
} | {
|
|
605
605
|
id: string;
|
|
@@ -617,7 +617,7 @@ export declare function prepareImportedTasksetPackage(input: {
|
|
|
617
617
|
mediaType: string;
|
|
618
618
|
visibility: "policy" | "verifier" | "host_private";
|
|
619
619
|
};
|
|
620
|
-
calibrationStatus: "
|
|
620
|
+
calibrationStatus: "passed" | "failed" | "pending";
|
|
621
621
|
model?: {
|
|
622
622
|
providerId: string;
|
|
623
623
|
modelId: string;
|
|
@@ -766,7 +766,7 @@ export declare function prepareImportedTasksetPackage(input: {
|
|
|
766
766
|
};
|
|
767
767
|
familyKey: string;
|
|
768
768
|
inputHash: string;
|
|
769
|
-
split: "train" | "validation" | "
|
|
769
|
+
split: "train" | "validation" | "frozen_eval" | "test";
|
|
770
770
|
}[];
|
|
771
771
|
sealedAt: string;
|
|
772
772
|
sealedBy: string;
|
|
@@ -794,7 +794,7 @@ export declare function prepareImportedTasksetPackage(input: {
|
|
|
794
794
|
attemptId: string;
|
|
795
795
|
occurredAt: string;
|
|
796
796
|
familyKey: string | null;
|
|
797
|
-
split: "train" | "validation" | "
|
|
797
|
+
split: "train" | "validation" | "frozen_eval" | "test";
|
|
798
798
|
input: Record<string, import("zod").JSONType>;
|
|
799
799
|
observedOutput: Record<string, import("zod").JSONType> | null;
|
|
800
800
|
expected: Record<string, import("zod").JSONType> | null;
|
|
@@ -859,9 +859,9 @@ export declare function prepareImportedTasksetPackage(input: {
|
|
|
859
859
|
} | null;
|
|
860
860
|
};
|
|
861
861
|
evidenceValidity: "pending" | "valid" | "invalid";
|
|
862
|
-
taskAdmissibility: "
|
|
863
|
-
observedQuality: "
|
|
864
|
-
targetApproval: "
|
|
862
|
+
taskAdmissibility: "pending" | "rejected" | "approved";
|
|
863
|
+
observedQuality: "passed" | "failed" | "unavailable" | "unscored";
|
|
864
|
+
targetApproval: "pending" | "rejected" | "approved" | "not_required";
|
|
865
865
|
approvedTarget: Record<string, import("zod").JSONType> | null;
|
|
866
866
|
grade: {
|
|
867
867
|
schemaVersion: "openpond.rewardComposition.v1";
|
|
@@ -879,8 +879,8 @@ export declare function prepareImportedTasksetPackage(input: {
|
|
|
879
879
|
contentHash: string;
|
|
880
880
|
revision: number;
|
|
881
881
|
};
|
|
882
|
-
role: "
|
|
883
|
-
status: "failed" | "
|
|
882
|
+
role: "training" | "evaluation";
|
|
883
|
+
status: "failed" | "unavailable" | "scored" | "pending";
|
|
884
884
|
rawScore: number | null;
|
|
885
885
|
normalizedScore: number | null;
|
|
886
886
|
passed: boolean | null;
|
|
@@ -893,7 +893,7 @@ export declare function prepareImportedTasksetPackage(input: {
|
|
|
893
893
|
score: number | null;
|
|
894
894
|
passed: boolean;
|
|
895
895
|
rewardEligible: boolean;
|
|
896
|
-
failureClass: "
|
|
896
|
+
failureClass: "policy_failure" | "grader_failure" | "environment_failure" | "infrastructure_failure" | "timeout" | "cancelled" | null;
|
|
897
897
|
feedback: string[];
|
|
898
898
|
visibleEvidenceRefs: string[];
|
|
899
899
|
privilegedEvidenceRefs: string[];
|
|
@@ -913,12 +913,12 @@ export declare function prepareImportedTasksetPackage(input: {
|
|
|
913
913
|
} | undefined;
|
|
914
914
|
}[];
|
|
915
915
|
training: {
|
|
916
|
-
status: "
|
|
916
|
+
status: "scored" | "unscorable" | "not_configured";
|
|
917
917
|
score: number | null;
|
|
918
918
|
passed: boolean | null;
|
|
919
919
|
};
|
|
920
920
|
evaluation: {
|
|
921
|
-
status: "
|
|
921
|
+
status: "scored" | "unscorable" | "not_configured";
|
|
922
922
|
score: number | null;
|
|
923
923
|
passed: boolean | null;
|
|
924
924
|
};
|
|
@@ -940,8 +940,8 @@ export declare function prepareImportedTasksetPackage(input: {
|
|
|
940
940
|
contentHash: string;
|
|
941
941
|
revision: number;
|
|
942
942
|
};
|
|
943
|
-
role: "
|
|
944
|
-
status: "failed" | "
|
|
943
|
+
role: "training" | "evaluation";
|
|
944
|
+
status: "failed" | "unavailable" | "scored" | "pending";
|
|
945
945
|
rawScore: number | null;
|
|
946
946
|
normalizedScore: number | null;
|
|
947
947
|
passed: boolean | null;
|
|
@@ -954,7 +954,7 @@ export declare function prepareImportedTasksetPackage(input: {
|
|
|
954
954
|
score: number | null;
|
|
955
955
|
passed: boolean;
|
|
956
956
|
rewardEligible: boolean;
|
|
957
|
-
failureClass: "
|
|
957
|
+
failureClass: "policy_failure" | "grader_failure" | "environment_failure" | "infrastructure_failure" | "timeout" | "cancelled" | null;
|
|
958
958
|
feedback: string[];
|
|
959
959
|
visibleEvidenceRefs: string[];
|
|
960
960
|
privilegedEvidenceRefs: string[];
|
|
@@ -974,12 +974,12 @@ export declare function prepareImportedTasksetPackage(input: {
|
|
|
974
974
|
} | undefined;
|
|
975
975
|
}[];
|
|
976
976
|
training: {
|
|
977
|
-
status: "
|
|
977
|
+
status: "scored" | "unscorable" | "not_configured";
|
|
978
978
|
score: number | null;
|
|
979
979
|
passed: boolean | null;
|
|
980
980
|
};
|
|
981
981
|
evaluation: {
|
|
982
|
-
status: "
|
|
982
|
+
status: "scored" | "unscorable" | "not_configured";
|
|
983
983
|
score: number | null;
|
|
984
984
|
passed: boolean | null;
|
|
985
985
|
};
|
|
@@ -994,14 +994,14 @@ export declare function prepareImportedTasksetPackage(input: {
|
|
|
994
994
|
id: string;
|
|
995
995
|
revision: number;
|
|
996
996
|
name: string;
|
|
997
|
-
kind: "
|
|
997
|
+
kind: "provider" | "benchmark" | "work" | "direct" | "opentelemetry";
|
|
998
998
|
taskDefinition: {
|
|
999
999
|
id: string;
|
|
1000
1000
|
contentHash: string;
|
|
1001
1001
|
revision: number;
|
|
1002
1002
|
};
|
|
1003
1003
|
enabled: boolean;
|
|
1004
|
-
allowedSplits: ("train" | "validation" | "
|
|
1004
|
+
allowedSplits: ("train" | "validation" | "frozen_eval" | "test")[];
|
|
1005
1005
|
mapping: {
|
|
1006
1006
|
schemaVersion: "openpond.taskSourceMapping.v1";
|
|
1007
1007
|
exampleId: string;
|
|
@@ -1012,7 +1012,7 @@ export declare function prepareImportedTasksetPackage(input: {
|
|
|
1012
1012
|
expected: string | null;
|
|
1013
1013
|
evaluatorContext: string | null;
|
|
1014
1014
|
familyKey: string;
|
|
1015
|
-
split: "train" | "validation" | "
|
|
1015
|
+
split: "train" | "validation" | "frozen_eval" | "test";
|
|
1016
1016
|
} | null;
|
|
1017
1017
|
adapterVersion: string | null;
|
|
1018
1018
|
contentHash: string;
|
|
@@ -1068,8 +1068,8 @@ export declare function prepareImportedTasksetPackage(input: {
|
|
|
1068
1068
|
releaseId: string;
|
|
1069
1069
|
releaseHash: string;
|
|
1070
1070
|
managedReleasePath: string;
|
|
1071
|
-
adaptationSplit: "train" | "validation" | "
|
|
1072
|
-
evaluationSplit: "train" | "validation" | "
|
|
1071
|
+
adaptationSplit: "train" | "validation" | "frozen_eval" | "test";
|
|
1072
|
+
evaluationSplit: "train" | "validation" | "frozen_eval" | "test";
|
|
1073
1073
|
primaryMetric: "foreground_tokens" | "success_rate" | "latency_ms" | "cost_usd";
|
|
1074
1074
|
qualityGate: "none" | "non_regression" | "all_pass";
|
|
1075
1075
|
source: "builtin" | "imported";
|
|
@@ -1082,16 +1082,16 @@ export declare function prepareImportedTasksetPackage(input: {
|
|
|
1082
1082
|
publishedAt: string;
|
|
1083
1083
|
metadata: Record<string, unknown>;
|
|
1084
1084
|
} | null;
|
|
1085
|
-
status: "draft" | "awaiting_disclosure_approval" | "awaiting_materialization_approval" | "materializing" | "validating" | "needs_review" | "baselining" | "ready" | "blocked" | "
|
|
1085
|
+
status: "failed" | "draft" | "awaiting_disclosure_approval" | "awaiting_materialization_approval" | "materializing" | "validating" | "needs_review" | "baselining" | "ready" | "blocked" | "archived";
|
|
1086
1086
|
sourceRefs: ({
|
|
1087
1087
|
id: string;
|
|
1088
1088
|
profileId: string;
|
|
1089
1089
|
title: string;
|
|
1090
1090
|
sourceHash: string;
|
|
1091
1091
|
occurredAt: string;
|
|
1092
|
-
licensingStatus: "
|
|
1093
|
-
secretScanStatus: "
|
|
1094
|
-
piiScanStatus: "
|
|
1092
|
+
licensingStatus: "pending" | "blocked" | "approved" | "review";
|
|
1093
|
+
secretScanStatus: "passed" | "pending" | "blocked";
|
|
1094
|
+
piiScanStatus: "passed" | "pending" | "blocked" | "review";
|
|
1095
1095
|
metadata: Record<string, unknown>;
|
|
1096
1096
|
schemaVersion: "openpond.huggingFaceDatasetSource.v1";
|
|
1097
1097
|
kind: "huggingface";
|
|
@@ -1110,9 +1110,9 @@ export declare function prepareImportedTasksetPackage(input: {
|
|
|
1110
1110
|
title: string;
|
|
1111
1111
|
sourceHash: string;
|
|
1112
1112
|
occurredAt: string;
|
|
1113
|
-
licensingStatus: "
|
|
1114
|
-
secretScanStatus: "
|
|
1115
|
-
piiScanStatus: "
|
|
1113
|
+
licensingStatus: "pending" | "blocked" | "approved" | "review";
|
|
1114
|
+
secretScanStatus: "passed" | "pending" | "blocked";
|
|
1115
|
+
piiScanStatus: "passed" | "pending" | "blocked" | "review";
|
|
1116
1116
|
metadata: Record<string, unknown>;
|
|
1117
1117
|
schemaVersion: "openpond.uploadedFileDatasetSource.v1";
|
|
1118
1118
|
kind: "uploaded_file";
|
|
@@ -1127,9 +1127,9 @@ export declare function prepareImportedTasksetPackage(input: {
|
|
|
1127
1127
|
title: string;
|
|
1128
1128
|
sourceHash: string;
|
|
1129
1129
|
occurredAt: string;
|
|
1130
|
-
licensingStatus: "
|
|
1131
|
-
secretScanStatus: "
|
|
1132
|
-
piiScanStatus: "
|
|
1130
|
+
licensingStatus: "pending" | "blocked" | "approved" | "review";
|
|
1131
|
+
secretScanStatus: "passed" | "pending" | "blocked";
|
|
1132
|
+
piiScanStatus: "passed" | "pending" | "blocked" | "review";
|
|
1133
1133
|
metadata: Record<string, unknown>;
|
|
1134
1134
|
schemaVersion: "openpond.generatedDatasetSource.v1";
|
|
1135
1135
|
kind: "generated";
|
|
@@ -1143,9 +1143,9 @@ export declare function prepareImportedTasksetPackage(input: {
|
|
|
1143
1143
|
title: string;
|
|
1144
1144
|
sourceHash: string;
|
|
1145
1145
|
occurredAt: string;
|
|
1146
|
-
licensingStatus: "
|
|
1147
|
-
secretScanStatus: "
|
|
1148
|
-
piiScanStatus: "
|
|
1146
|
+
licensingStatus: "pending" | "blocked" | "approved" | "review";
|
|
1147
|
+
secretScanStatus: "passed" | "pending" | "blocked";
|
|
1148
|
+
piiScanStatus: "passed" | "pending" | "blocked" | "review";
|
|
1149
1149
|
metadata: Record<string, unknown>;
|
|
1150
1150
|
schemaVersion: "openpond.learningBatchDatasetSource.v1";
|
|
1151
1151
|
kind: "learning_batch";
|
|
@@ -1179,9 +1179,9 @@ export declare function prepareImportedTasksetPackage(input: {
|
|
|
1179
1179
|
purpose: "task_authoring_and_evaluation";
|
|
1180
1180
|
};
|
|
1181
1181
|
connectedAppIds: string[];
|
|
1182
|
-
secretScanStatus: "
|
|
1183
|
-
piiScanStatus: "
|
|
1184
|
-
licensingStatus: "
|
|
1182
|
+
secretScanStatus: "passed" | "pending" | "blocked";
|
|
1183
|
+
piiScanStatus: "passed" | "pending" | "blocked" | "review";
|
|
1184
|
+
licensingStatus: "pending" | "blocked" | "approved" | "review";
|
|
1185
1185
|
metadata: Record<string, unknown>;
|
|
1186
1186
|
})[];
|
|
1187
1187
|
policy: {
|
|
@@ -1197,7 +1197,7 @@ export declare function prepareImportedTasksetPackage(input: {
|
|
|
1197
1197
|
stateful: boolean;
|
|
1198
1198
|
deterministicSeeds: boolean;
|
|
1199
1199
|
toolNames: string[];
|
|
1200
|
-
lifecycle: ("
|
|
1200
|
+
lifecycle: ("grade" | "create" | "reset" | "step" | "cleanup")[];
|
|
1201
1201
|
defaultTimeoutMs: number;
|
|
1202
1202
|
networkPolicy: "none" | "declared_read_only" | "declared_scoped";
|
|
1203
1203
|
metadata: Record<string, unknown>;
|
|
@@ -1236,8 +1236,8 @@ export declare function prepareImportedTasksetPackage(input: {
|
|
|
1236
1236
|
capabilities: {
|
|
1237
1237
|
schemaVersion: "openpond.tasksetCapabilities.v1";
|
|
1238
1238
|
taskKind: "chat" | "single_agent" | "multi_agent" | "custom_program";
|
|
1239
|
-
supportedSignals: ("label" | "
|
|
1240
|
-
compatibleMethods: ("none" | "
|
|
1239
|
+
supportedSignals: ("label" | "reward" | "preference" | "feedback" | "demonstration" | "correction")[];
|
|
1240
|
+
compatibleMethods: ("none" | "sft" | "dpo" | "grpo" | "ppo" | "sdft" | "opd" | "opsd" | "sdpo" | "retrieval")[];
|
|
1241
1241
|
rewardKinds: ("none" | "exact" | "deterministic" | "model_judge" | "human")[];
|
|
1242
1242
|
requiresTools: boolean;
|
|
1243
1243
|
requiresState: boolean;
|
|
@@ -1250,7 +1250,7 @@ export declare function prepareImportedTasksetPackage(input: {
|
|
|
1250
1250
|
schemaVersion: "openpond.taskData.v1";
|
|
1251
1251
|
id: string;
|
|
1252
1252
|
clusterKey: string;
|
|
1253
|
-
split: "train" | "validation" | "
|
|
1253
|
+
split: "train" | "validation" | "frozen_eval" | "test";
|
|
1254
1254
|
input: Record<string, unknown>;
|
|
1255
1255
|
expectedOutput: Record<string, unknown> | null;
|
|
1256
1256
|
policyVisibleContext: Record<string, unknown>;
|
|
@@ -1266,7 +1266,7 @@ export declare function prepareImportedTasksetPackage(input: {
|
|
|
1266
1266
|
mediaType: string;
|
|
1267
1267
|
sha256: string;
|
|
1268
1268
|
sizeBytes: number;
|
|
1269
|
-
split: "train" | "validation" | "
|
|
1269
|
+
split: "train" | "validation" | "frozen_eval" | "test";
|
|
1270
1270
|
metadata: Record<string, unknown>;
|
|
1271
1271
|
}[] | undefined;
|
|
1272
1272
|
resourceRefs?: string[] | undefined;
|
|
@@ -1287,7 +1287,7 @@ export declare function prepareImportedTasksetPackage(input: {
|
|
|
1287
1287
|
rewardEligible: boolean;
|
|
1288
1288
|
privileged: boolean;
|
|
1289
1289
|
metadata: Record<string, unknown>;
|
|
1290
|
-
kind: "file" | "content" | "
|
|
1290
|
+
kind: "file" | "content" | "state" | "test" | "schema" | "diff" | "runtime_event";
|
|
1291
1291
|
config: Record<string, unknown>;
|
|
1292
1292
|
} | {
|
|
1293
1293
|
id: string;
|
|
@@ -1305,7 +1305,7 @@ export declare function prepareImportedTasksetPackage(input: {
|
|
|
1305
1305
|
modelId: string;
|
|
1306
1306
|
};
|
|
1307
1307
|
calibrationFixtureRefs: string[];
|
|
1308
|
-
calibrationStatus: "
|
|
1308
|
+
calibrationStatus: "passed" | "failed" | "pending";
|
|
1309
1309
|
temperature: number;
|
|
1310
1310
|
} | {
|
|
1311
1311
|
id: string;
|
|
@@ -1338,7 +1338,7 @@ export declare function prepareImportedTasksetPackage(input: {
|
|
|
1338
1338
|
graderFixtures: {
|
|
1339
1339
|
id: string;
|
|
1340
1340
|
taskId: string;
|
|
1341
|
-
label: "positive" | "negative" | "boundary" | "adversarial" | "prompt_injection"
|
|
1341
|
+
label: "infrastructure_failure" | "positive" | "negative" | "boundary" | "adversarial" | "prompt_injection";
|
|
1342
1342
|
output: Record<string, unknown>;
|
|
1343
1343
|
infrastructureError: string | null;
|
|
1344
1344
|
expectedPassed: boolean;
|
|
@@ -1502,7 +1502,7 @@ export declare function prepareImportedTasksetPackage(input: {
|
|
|
1502
1502
|
tasksetId: string;
|
|
1503
1503
|
tasksetHash: string;
|
|
1504
1504
|
ready: boolean;
|
|
1505
|
-
recommendedMethod: "none" | "
|
|
1505
|
+
recommendedMethod: "none" | "sft" | "dpo" | "grpo" | "ppo" | "sdft" | "opd" | "opsd" | "sdpo" | "retrieval";
|
|
1506
1506
|
trainingPath: {
|
|
1507
1507
|
primaryMethod: "sft" | "dpo" | "grpo" | "ppo" | "sdft" | "opsd" | "sdpo";
|
|
1508
1508
|
bootstrap: {
|
|
@@ -1552,7 +1552,7 @@ export declare function prepareImportedTasksetPackage(input: {
|
|
|
1552
1552
|
schemaVersion: "openpond.datasetSemanticSchema.v1";
|
|
1553
1553
|
fields: {
|
|
1554
1554
|
name: string;
|
|
1555
|
-
semanticRole: "metadata" | "
|
|
1555
|
+
semanticRole: "metadata" | "reward" | "feedback" | "messages" | "row_id" | "cluster_id" | "prompt" | "demonstration" | "chosen" | "rejected" | "expected_output" | "privileged_context" | "tag";
|
|
1556
1556
|
logicalType: "string" | "boolean" | "messages" | "integer" | "float" | "json";
|
|
1557
1557
|
nullable: boolean;
|
|
1558
1558
|
policy: "metadata" | "visible" | "privileged";
|
|
@@ -1561,7 +1561,7 @@ export declare function prepareImportedTasksetPackage(input: {
|
|
|
1561
1561
|
};
|
|
1562
1562
|
shards: {
|
|
1563
1563
|
id: string;
|
|
1564
|
-
split: "train" | "validation" | "
|
|
1564
|
+
split: "train" | "validation" | "frozen_eval" | "test";
|
|
1565
1565
|
path: string;
|
|
1566
1566
|
contentHash: string;
|
|
1567
1567
|
schemaHash: string;
|
|
@@ -1570,7 +1570,7 @@ export declare function prepareImportedTasksetPackage(input: {
|
|
|
1570
1570
|
rowGroupCount: number;
|
|
1571
1571
|
}[];
|
|
1572
1572
|
rowCount: number;
|
|
1573
|
-
splitCounts: Record<"train" | "validation" | "
|
|
1573
|
+
splitCounts: Record<"train" | "validation" | "frozen_eval" | "test", number>;
|
|
1574
1574
|
sourceReceiptRefs: string[];
|
|
1575
1575
|
mappingHash: string;
|
|
1576
1576
|
qualityReportHash: string;
|
|
@@ -1593,7 +1593,7 @@ export declare function prepareImportedTasksetPackage(input: {
|
|
|
1593
1593
|
};
|
|
1594
1594
|
generatedFiles: {
|
|
1595
1595
|
path: string;
|
|
1596
|
-
role: "environment" | "
|
|
1596
|
+
role: "environment" | "fixture" | "verifier";
|
|
1597
1597
|
content: string;
|
|
1598
1598
|
}[];
|
|
1599
1599
|
};
|
|
@@ -108,7 +108,7 @@ export declare function createTrainingAcceptanceClient(request: (path: string, i
|
|
|
108
108
|
id: string;
|
|
109
109
|
contentHash: string;
|
|
110
110
|
};
|
|
111
|
-
state: "
|
|
111
|
+
state: "failed" | "cancelled" | "queued" | "running" | "cancelling" | "completed" | "budget_exhausted" | "blocked" | "cleaning";
|
|
112
112
|
updatedAt: string;
|
|
113
113
|
}[];
|
|
114
114
|
nextCursor: string | null;
|
|
@@ -173,14 +173,14 @@ export declare function createTrainingAcceptanceClient(request: (path: string, i
|
|
|
173
173
|
contentHash: string;
|
|
174
174
|
};
|
|
175
175
|
revision: number;
|
|
176
|
-
state: "
|
|
176
|
+
state: "failed" | "cancelled" | "queued" | "running" | "cancelling" | "completed" | "budget_exhausted" | "blocked" | "cleaning";
|
|
177
177
|
attempts: {
|
|
178
178
|
id: string;
|
|
179
179
|
checkId: string;
|
|
180
180
|
subject: "baseline" | "candidate";
|
|
181
181
|
ordinal: number;
|
|
182
182
|
dispatchId: string;
|
|
183
|
-
state: "failed" | "
|
|
183
|
+
state: "failed" | "cancelled" | "queued" | "running" | "completed" | "budget_exhausted" | "cleaning" | "dispatching" | "grading";
|
|
184
184
|
artifact: {
|
|
185
185
|
id: string;
|
|
186
186
|
contentHash: string;
|
|
@@ -204,7 +204,7 @@ export declare function createTrainingAcceptanceClient(request: (path: string, i
|
|
|
204
204
|
expectedCount: number;
|
|
205
205
|
scoredCount: number;
|
|
206
206
|
score: number | null;
|
|
207
|
-
status: "failed" | "
|
|
207
|
+
status: "failed" | "cancelled" | "running" | "completed" | "pending";
|
|
208
208
|
evidenceHash: string;
|
|
209
209
|
} | null;
|
|
210
210
|
executionReceipt: {
|
|
@@ -113,7 +113,7 @@ export declare function createTrainingHarnessSourceClient(request: (path: string
|
|
|
113
113
|
step: true;
|
|
114
114
|
collect: true;
|
|
115
115
|
destroy: true;
|
|
116
|
-
resetScope: "
|
|
116
|
+
resetScope: "attempt" | "task";
|
|
117
117
|
};
|
|
118
118
|
graderInterface: {
|
|
119
119
|
visibleEvidence: string[];
|