openpond-sdk 0.9.4 → 0.9.6
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-DNBAJUVR.js → chunk-6D4JIOOC.js} +8 -3
- package/dist/{chunk-DNBAJUVR.js.map → chunk-6D4JIOOC.js.map} +2 -2
- package/dist/dataset-workspaces.js +31 -2
- package/dist/dataset-workspaces.js.map +2 -2
- package/dist/experiments.js +33 -1
- package/dist/experiments.js.map +3 -3
- package/dist/learning.js +55 -0
- package/dist/learning.js.map +4 -4
- package/dist/taskset-packages.js +1 -1
- package/dist/training-projects.js +4 -1
- package/dist/training-projects.js.map +2 -2
- package/dist/types/packages/sdk/src/dataset-workspace-client.d.ts +1080 -1
- 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 +34 -0
- package/dist/types/packages/sdk/src/dataset-workspace-contracts.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/experiment-client.d.ts +16 -0
- package/dist/types/packages/sdk/src/experiment-client.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/experiment-inspection.d.ts +50 -0
- package/dist/types/packages/sdk/src/experiment-inspection.d.ts.map +1 -0
- package/dist/types/packages/sdk/src/experiments.d.ts +1 -0
- package/dist/types/packages/sdk/src/experiments.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/grader-inspection.d.ts +568 -0
- package/dist/types/packages/sdk/src/grader-inspection.d.ts.map +1 -0
- package/dist/types/packages/sdk/src/learning.d.ts +1 -0
- package/dist/types/packages/sdk/src/learning.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/taskset-authored-files.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/training-project-client.d.ts +4 -0
- package/dist/types/packages/sdk/src/training-project-client.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/training-project-contracts.d.ts +12 -0
- package/dist/types/packages/sdk/src/training-project-contracts.d.ts.map +1 -1
- package/package.json +1 -1
|
@@ -1,4 +1,4 @@
|
|
|
1
|
-
import { DatasetWorkspacePublishSchema, DatasetWorkspaceWriteSchema } from "./dataset-workspace-contracts.js";
|
|
1
|
+
import { DatasetWorkspaceBeginVersionSchema, DatasetWorkspacePublishSchema, DatasetWorkspaceWriteSchema } from "./dataset-workspace-contracts.js";
|
|
2
2
|
import type { z } from "zod";
|
|
3
3
|
/** Hosted independent authoring. Writes use explicit revision/operation identity;
|
|
4
4
|
* callers retain the operation ID when retrying an uncertain response. */
|
|
@@ -32,6 +32,7 @@ export declare class OpenPondDatasetWorkspaceClient {
|
|
|
32
32
|
} | null;
|
|
33
33
|
updatedAt: string;
|
|
34
34
|
originProjectId?: string | undefined;
|
|
35
|
+
ownerScope?: "workspace" | "personal" | undefined;
|
|
35
36
|
}[];
|
|
36
37
|
nextCursor: string | null;
|
|
37
38
|
}>;
|
|
@@ -560,6 +561,1082 @@ export declare class OpenPondDatasetWorkspaceClient {
|
|
|
560
561
|
packageHash: string;
|
|
561
562
|
} | null;
|
|
562
563
|
originProjectId?: string | undefined;
|
|
564
|
+
ownerScope?: "workspace" | "personal" | undefined;
|
|
565
|
+
}>;
|
|
566
|
+
beginVersion(id: string, input: z.input<typeof DatasetWorkspaceBeginVersionSchema>, signal?: AbortSignal): Promise<{
|
|
567
|
+
schemaVersion: "openpond.datasetWorkspaceReceipt.v1";
|
|
568
|
+
teamId: string;
|
|
569
|
+
datasetId: string;
|
|
570
|
+
revision: number;
|
|
571
|
+
workspace: {
|
|
572
|
+
schemaVersion: "openpond.tasksetDraftWorkspace.v1";
|
|
573
|
+
draft: {
|
|
574
|
+
schemaVersion: "openpond.tasksetDraft.v1";
|
|
575
|
+
id: string;
|
|
576
|
+
revision: number;
|
|
577
|
+
profileId: string;
|
|
578
|
+
modelScope: {
|
|
579
|
+
modelId: string;
|
|
580
|
+
expectedModelRevision: number;
|
|
581
|
+
source?: {
|
|
582
|
+
schemaVersion: "openpond.modelTasksetDraftPreparation.v1";
|
|
583
|
+
draftId: string;
|
|
584
|
+
requestHash: string;
|
|
585
|
+
sourcePackageHash: string;
|
|
586
|
+
sourceTasksetRef: {
|
|
587
|
+
id: string;
|
|
588
|
+
contentHash: string;
|
|
589
|
+
revision: number;
|
|
590
|
+
};
|
|
591
|
+
tasksetId: string;
|
|
592
|
+
tasksetRevision: number;
|
|
593
|
+
lineage: {
|
|
594
|
+
schemaVersion: "openpond.modelTasksetAuthoring.v1";
|
|
595
|
+
owner: {
|
|
596
|
+
scopeId: string;
|
|
597
|
+
modelId: string;
|
|
598
|
+
};
|
|
599
|
+
root: {
|
|
600
|
+
id: string;
|
|
601
|
+
contentHash: string;
|
|
602
|
+
revision: number;
|
|
603
|
+
};
|
|
604
|
+
parent: {
|
|
605
|
+
id: string;
|
|
606
|
+
contentHash: string;
|
|
607
|
+
revision: number;
|
|
608
|
+
};
|
|
609
|
+
};
|
|
610
|
+
authoringGraph?: "bound" | undefined;
|
|
611
|
+
} | undefined;
|
|
612
|
+
} | null;
|
|
613
|
+
name: string;
|
|
614
|
+
objective: string;
|
|
615
|
+
purpose: "general" | "benchmark";
|
|
616
|
+
benchmark: {
|
|
617
|
+
schemaVersion: "openpond.tasksetBenchmark.v1";
|
|
618
|
+
definitionId: string;
|
|
619
|
+
releaseId: string;
|
|
620
|
+
releaseHash: string;
|
|
621
|
+
managedReleasePath: string;
|
|
622
|
+
adaptationSplit: "train" | "validation" | "test" | "frozen_eval";
|
|
623
|
+
evaluationSplit: "train" | "validation" | "test" | "frozen_eval";
|
|
624
|
+
primaryMetric: "foreground_tokens" | "success_rate" | "latency_ms" | "cost_usd";
|
|
625
|
+
qualityGate: "none" | "non_regression" | "all_pass";
|
|
626
|
+
source: "builtin" | "imported";
|
|
627
|
+
metadata: Record<string, unknown>;
|
|
628
|
+
} | null;
|
|
629
|
+
preferenceComparison: {
|
|
630
|
+
schemaVersion: "openpond.tasksetPreferenceComparisonBinding.v1";
|
|
631
|
+
releaseId: string;
|
|
632
|
+
releaseHash: string;
|
|
633
|
+
publishedAt: string;
|
|
634
|
+
metadata: Record<string, unknown>;
|
|
635
|
+
} | null;
|
|
636
|
+
status: "draft" | "validating" | "needs_review" | "published";
|
|
637
|
+
sourceRefs: ({
|
|
638
|
+
id: string;
|
|
639
|
+
profileId: string;
|
|
640
|
+
title: string;
|
|
641
|
+
sourceHash: string;
|
|
642
|
+
occurredAt: string;
|
|
643
|
+
licensingStatus: "blocked" | "pending" | "approved" | "review";
|
|
644
|
+
secretScanStatus: "blocked" | "pending" | "passed";
|
|
645
|
+
piiScanStatus: "blocked" | "pending" | "review" | "passed";
|
|
646
|
+
metadata: Record<string, unknown>;
|
|
647
|
+
schemaVersion: "openpond.huggingFaceDatasetSource.v1";
|
|
648
|
+
kind: "huggingface";
|
|
649
|
+
repositoryId: string;
|
|
650
|
+
repositoryUrl: string;
|
|
651
|
+
revision: string;
|
|
652
|
+
configuration: string;
|
|
653
|
+
upstreamSplits: string[];
|
|
654
|
+
gated: boolean;
|
|
655
|
+
private: boolean;
|
|
656
|
+
declaredLicense: string | null;
|
|
657
|
+
sourceFileHashes: string[];
|
|
658
|
+
} | {
|
|
659
|
+
id: string;
|
|
660
|
+
profileId: string;
|
|
661
|
+
title: string;
|
|
662
|
+
sourceHash: string;
|
|
663
|
+
occurredAt: string;
|
|
664
|
+
licensingStatus: "blocked" | "pending" | "approved" | "review";
|
|
665
|
+
secretScanStatus: "blocked" | "pending" | "passed";
|
|
666
|
+
piiScanStatus: "blocked" | "pending" | "review" | "passed";
|
|
667
|
+
metadata: Record<string, unknown>;
|
|
668
|
+
schemaVersion: "openpond.uploadedFileDatasetSource.v1";
|
|
669
|
+
kind: "uploaded_file";
|
|
670
|
+
originalFileNames: string[];
|
|
671
|
+
mediaTypes: string[];
|
|
672
|
+
sourceFileHashes: string[];
|
|
673
|
+
totalBytes: number;
|
|
674
|
+
parserVersion: string;
|
|
675
|
+
} | {
|
|
676
|
+
id: string;
|
|
677
|
+
profileId: string;
|
|
678
|
+
title: string;
|
|
679
|
+
sourceHash: string;
|
|
680
|
+
occurredAt: string;
|
|
681
|
+
licensingStatus: "blocked" | "pending" | "approved" | "review";
|
|
682
|
+
secretScanStatus: "blocked" | "pending" | "passed";
|
|
683
|
+
piiScanStatus: "blocked" | "pending" | "review" | "passed";
|
|
684
|
+
metadata: Record<string, unknown>;
|
|
685
|
+
schemaVersion: "openpond.generatedDatasetSource.v1";
|
|
686
|
+
kind: "generated";
|
|
687
|
+
generatorId: string;
|
|
688
|
+
generatorVersion: string;
|
|
689
|
+
seed: number;
|
|
690
|
+
generatorHash: string;
|
|
691
|
+
} | {
|
|
692
|
+
id: string;
|
|
693
|
+
profileId: string;
|
|
694
|
+
title: string;
|
|
695
|
+
sourceHash: string;
|
|
696
|
+
occurredAt: string;
|
|
697
|
+
licensingStatus: "blocked" | "pending" | "approved" | "review";
|
|
698
|
+
secretScanStatus: "blocked" | "pending" | "passed";
|
|
699
|
+
piiScanStatus: "blocked" | "pending" | "review" | "passed";
|
|
700
|
+
metadata: Record<string, unknown>;
|
|
701
|
+
schemaVersion: "openpond.learningBatchDatasetSource.v1";
|
|
702
|
+
kind: "learning_batch";
|
|
703
|
+
batch: {
|
|
704
|
+
id: string;
|
|
705
|
+
revision: number;
|
|
706
|
+
contentHash: string;
|
|
707
|
+
};
|
|
708
|
+
taskDefinition: {
|
|
709
|
+
id: string;
|
|
710
|
+
revision: number;
|
|
711
|
+
contentHash: string;
|
|
712
|
+
};
|
|
713
|
+
admittedBy: string;
|
|
714
|
+
} | {
|
|
715
|
+
schemaVersion: "openpond.trainingSource.v1";
|
|
716
|
+
id: string;
|
|
717
|
+
profileId: string;
|
|
718
|
+
sessionId: string;
|
|
719
|
+
turnIds: string[];
|
|
720
|
+
workspaceId: string | null;
|
|
721
|
+
sourceHash: string;
|
|
722
|
+
clusterKey: string;
|
|
723
|
+
title: string;
|
|
724
|
+
occurredAt: string;
|
|
725
|
+
consent: {
|
|
726
|
+
status: "pending" | "granted" | "denied" | "revoked";
|
|
727
|
+
scope: "metadata_only" | "selected_turns" | "full_session";
|
|
728
|
+
grantedBy: string | null;
|
|
729
|
+
grantedAt: string | null;
|
|
730
|
+
purpose: "task_authoring_and_evaluation";
|
|
731
|
+
};
|
|
732
|
+
connectedAppIds: string[];
|
|
733
|
+
secretScanStatus: "blocked" | "pending" | "passed";
|
|
734
|
+
piiScanStatus: "blocked" | "pending" | "review" | "passed";
|
|
735
|
+
licensingStatus: "blocked" | "pending" | "approved" | "review";
|
|
736
|
+
metadata: Record<string, unknown>;
|
|
737
|
+
})[];
|
|
738
|
+
policy: {
|
|
739
|
+
policyVisibleFields: string[];
|
|
740
|
+
privilegedFields: string[];
|
|
741
|
+
hiddenGraderRefs: string[];
|
|
742
|
+
connectedAppScopes: string[];
|
|
743
|
+
};
|
|
744
|
+
environment: {
|
|
745
|
+
protocolVersion: "openpond.taskEnvironment.v1";
|
|
746
|
+
kind: "chat" | "agent" | "program" | "stateful_harness" | "work";
|
|
747
|
+
entrypoint: string;
|
|
748
|
+
stateful: boolean;
|
|
749
|
+
deterministicSeeds: boolean;
|
|
750
|
+
toolNames: string[];
|
|
751
|
+
lifecycle: ("create" | "reset" | "step" | "grade" | "cleanup")[];
|
|
752
|
+
defaultTimeoutMs: number;
|
|
753
|
+
networkPolicy: "none" | "declared_read_only" | "declared_scoped";
|
|
754
|
+
metadata: Record<string, unknown>;
|
|
755
|
+
actionBindings?: {
|
|
756
|
+
schemaVersion: "openpond.harnessActionBinding.v1";
|
|
757
|
+
actionId: string;
|
|
758
|
+
modelToolName: string;
|
|
759
|
+
description: string;
|
|
760
|
+
inputSchema: Record<string, unknown>;
|
|
761
|
+
actionSchemaHash: string;
|
|
762
|
+
agentRelease: {
|
|
763
|
+
id: string;
|
|
764
|
+
contentHash: string;
|
|
765
|
+
};
|
|
766
|
+
implementationHash: string;
|
|
767
|
+
runtimeBindingId: string;
|
|
768
|
+
capabilityReceiptHash: string;
|
|
769
|
+
sideEffect: "read" | "write";
|
|
770
|
+
studentVisible: boolean;
|
|
771
|
+
timeoutMs: number;
|
|
772
|
+
episodeArgumentBindings: {
|
|
773
|
+
argument: string;
|
|
774
|
+
source: "case_id";
|
|
775
|
+
}[];
|
|
776
|
+
}[] | undefined;
|
|
777
|
+
resources?: {
|
|
778
|
+
id: string;
|
|
779
|
+
kind: "file" | "configuration" | "catalog" | "code_module";
|
|
780
|
+
path: string;
|
|
781
|
+
visibility: "privileged" | "policy_visible" | "policy_hidden";
|
|
782
|
+
required: boolean;
|
|
783
|
+
metadata: Record<string, unknown>;
|
|
784
|
+
mediaType?: string | null | undefined;
|
|
785
|
+
}[] | undefined;
|
|
786
|
+
};
|
|
787
|
+
output: {
|
|
788
|
+
mode: "text" | "structured_json" | "artifacts";
|
|
789
|
+
jsonSchema: Record<string, unknown> | null;
|
|
790
|
+
renderer: {
|
|
791
|
+
module: string;
|
|
792
|
+
exportName: string;
|
|
793
|
+
configRef: string | null;
|
|
794
|
+
} | null;
|
|
795
|
+
};
|
|
796
|
+
capabilities: {
|
|
797
|
+
schemaVersion: "openpond.tasksetCapabilities.v1";
|
|
798
|
+
taskKind: "chat" | "single_agent" | "multi_agent" | "custom_program";
|
|
799
|
+
supportedSignals: ("label" | "demonstration" | "reward" | "feedback" | "preference" | "correction")[];
|
|
800
|
+
compatibleMethods: ("none" | "retrieval" | "sft" | "dpo" | "grpo" | "ppo" | "sdft" | "opd" | "opsd" | "sdpo")[];
|
|
801
|
+
rewardKinds: ("none" | "exact" | "deterministic" | "model_judge" | "human")[];
|
|
802
|
+
requiresTools: boolean;
|
|
803
|
+
requiresState: boolean;
|
|
804
|
+
requiresPrivilegedGrading: boolean;
|
|
805
|
+
environmentPlacements: ("local" | "remote" | "colocated" | "provider_native")[];
|
|
806
|
+
exportable: boolean;
|
|
807
|
+
portabilityBlockers: string[];
|
|
808
|
+
};
|
|
809
|
+
metrics: {
|
|
810
|
+
schemaVersion: "openpond.tasksetMetricPolicy.v1";
|
|
811
|
+
primaryMetric: string;
|
|
812
|
+
aggregation: "custom" | "mean_score" | "pass_rate" | "weighted_mean";
|
|
813
|
+
missingReward: "zero" | "exclude";
|
|
814
|
+
customAggregator: {
|
|
815
|
+
module: string;
|
|
816
|
+
exportName: string;
|
|
817
|
+
contentHash: string;
|
|
818
|
+
timeoutMs: number;
|
|
819
|
+
networkPolicy: "none";
|
|
820
|
+
} | null;
|
|
821
|
+
taskWeights?: Record<string, number> | undefined;
|
|
822
|
+
};
|
|
823
|
+
review: {
|
|
824
|
+
enabled: boolean;
|
|
825
|
+
candidateCount: number;
|
|
826
|
+
minimumSamples: number;
|
|
827
|
+
allowTies: boolean;
|
|
828
|
+
allowRejectAll: boolean;
|
|
829
|
+
rubric: string;
|
|
830
|
+
criteria: {
|
|
831
|
+
id: string;
|
|
832
|
+
label: string;
|
|
833
|
+
description: string;
|
|
834
|
+
weight: number;
|
|
835
|
+
}[];
|
|
836
|
+
};
|
|
837
|
+
tasks: {
|
|
838
|
+
schemaVersion: "openpond.taskData.v1";
|
|
839
|
+
id: string;
|
|
840
|
+
clusterKey: string;
|
|
841
|
+
split: "train" | "validation" | "test" | "frozen_eval";
|
|
842
|
+
input: Record<string, unknown>;
|
|
843
|
+
expectedOutput: Record<string, unknown> | null;
|
|
844
|
+
policyVisibleContext: Record<string, unknown>;
|
|
845
|
+
privilegedContextRef: string | null;
|
|
846
|
+
tags: string[];
|
|
847
|
+
metadata: Record<string, unknown>;
|
|
848
|
+
sourceRefs: string[];
|
|
849
|
+
assets?: {
|
|
850
|
+
id: string;
|
|
851
|
+
sourceRefId: string;
|
|
852
|
+
artifactRef: string;
|
|
853
|
+
fileName: string;
|
|
854
|
+
mediaType: string;
|
|
855
|
+
sha256: string;
|
|
856
|
+
sizeBytes: number;
|
|
857
|
+
split: "train" | "validation" | "test" | "frozen_eval";
|
|
858
|
+
metadata: Record<string, unknown>;
|
|
859
|
+
}[] | undefined;
|
|
860
|
+
resourceRefs?: string[] | undefined;
|
|
861
|
+
requiredOutputs?: {
|
|
862
|
+
path: string;
|
|
863
|
+
mediaType: string;
|
|
864
|
+
metadata: Record<string, unknown>;
|
|
865
|
+
schemaRef?: string | null | undefined;
|
|
866
|
+
maxBytes?: number | undefined;
|
|
867
|
+
}[] | undefined;
|
|
868
|
+
}[];
|
|
869
|
+
graders: ({
|
|
870
|
+
id: string;
|
|
871
|
+
version: string;
|
|
872
|
+
label: string;
|
|
873
|
+
weight: number;
|
|
874
|
+
hardGate: boolean;
|
|
875
|
+
rewardEligible: boolean;
|
|
876
|
+
privileged: boolean;
|
|
877
|
+
metadata: Record<string, unknown>;
|
|
878
|
+
kind: "file" | "content" | "test" | "schema" | "diff" | "runtime_event" | "state";
|
|
879
|
+
config: Record<string, unknown>;
|
|
880
|
+
} | {
|
|
881
|
+
id: string;
|
|
882
|
+
version: string;
|
|
883
|
+
label: string;
|
|
884
|
+
weight: number;
|
|
885
|
+
hardGate: boolean;
|
|
886
|
+
rewardEligible: boolean;
|
|
887
|
+
privileged: boolean;
|
|
888
|
+
metadata: Record<string, unknown>;
|
|
889
|
+
kind: "model_judge";
|
|
890
|
+
rubric: string;
|
|
891
|
+
judge: {
|
|
892
|
+
providerId: "openpond" | "codex" | "anthropic" | "openai" | "xai" | "google" | "openrouter" | "deepseek" | "zai" | "moonshot" | "together" | "groq" | "custom-openai-compatible";
|
|
893
|
+
modelId: string;
|
|
894
|
+
};
|
|
895
|
+
calibrationFixtureRefs: string[];
|
|
896
|
+
calibrationStatus: "failed" | "pending" | "passed";
|
|
897
|
+
temperature: number;
|
|
898
|
+
} | {
|
|
899
|
+
id: string;
|
|
900
|
+
version: string;
|
|
901
|
+
label: string;
|
|
902
|
+
weight: number;
|
|
903
|
+
hardGate: boolean;
|
|
904
|
+
rewardEligible: boolean;
|
|
905
|
+
privileged: boolean;
|
|
906
|
+
metadata: Record<string, unknown>;
|
|
907
|
+
kind: "human";
|
|
908
|
+
rubric: string;
|
|
909
|
+
reviewerRole: string;
|
|
910
|
+
} | {
|
|
911
|
+
id: string;
|
|
912
|
+
version: string;
|
|
913
|
+
label: string;
|
|
914
|
+
weight: number;
|
|
915
|
+
hardGate: boolean;
|
|
916
|
+
rewardEligible: boolean;
|
|
917
|
+
privileged: boolean;
|
|
918
|
+
metadata: Record<string, unknown>;
|
|
919
|
+
kind: "custom_verifier";
|
|
920
|
+
module: string;
|
|
921
|
+
exportName: string;
|
|
922
|
+
timeoutMs: number;
|
|
923
|
+
networkPolicy: "none";
|
|
924
|
+
runtime?: "isolated_javascript" | "sandbox_process" | undefined;
|
|
925
|
+
})[];
|
|
926
|
+
graderFixtures: {
|
|
927
|
+
id: string;
|
|
928
|
+
taskId: string;
|
|
929
|
+
label: "positive" | "negative" | "boundary" | "adversarial" | "prompt_injection" | "infrastructure_failure";
|
|
930
|
+
output: Record<string, unknown>;
|
|
931
|
+
infrastructureError: string | null;
|
|
932
|
+
expectedPassed: boolean;
|
|
933
|
+
expectedRewardEligible: boolean;
|
|
934
|
+
metadata: Record<string, unknown>;
|
|
935
|
+
}[];
|
|
936
|
+
learningSignals: {
|
|
937
|
+
demonstrations: {
|
|
938
|
+
id: string;
|
|
939
|
+
taskId: string | null;
|
|
940
|
+
sourceRefs: string[];
|
|
941
|
+
artifactRef: string;
|
|
942
|
+
approved: boolean;
|
|
943
|
+
confidence: number;
|
|
944
|
+
metadata: Record<string, unknown>;
|
|
945
|
+
kind: "demonstration";
|
|
946
|
+
prompt: string | null;
|
|
947
|
+
response: string | null;
|
|
948
|
+
}[];
|
|
949
|
+
preferences: {
|
|
950
|
+
id: string;
|
|
951
|
+
taskId: string | null;
|
|
952
|
+
sourceRefs: string[];
|
|
953
|
+
artifactRef: string;
|
|
954
|
+
approved: boolean;
|
|
955
|
+
confidence: number;
|
|
956
|
+
metadata: Record<string, unknown>;
|
|
957
|
+
kind: "preference";
|
|
958
|
+
prompt: string;
|
|
959
|
+
chosen: string;
|
|
960
|
+
rejected: string;
|
|
961
|
+
rationale: string | null;
|
|
962
|
+
}[];
|
|
963
|
+
corrections: {
|
|
964
|
+
id: string;
|
|
965
|
+
taskId: string | null;
|
|
966
|
+
sourceRefs: string[];
|
|
967
|
+
artifactRef: string;
|
|
968
|
+
approved: boolean;
|
|
969
|
+
confidence: number;
|
|
970
|
+
metadata: Record<string, unknown>;
|
|
971
|
+
kind: "correction";
|
|
972
|
+
original: string;
|
|
973
|
+
corrected: string;
|
|
974
|
+
rationale: string | null;
|
|
975
|
+
}[];
|
|
976
|
+
feedback: {
|
|
977
|
+
id: string;
|
|
978
|
+
taskId: string | null;
|
|
979
|
+
sourceRefs: string[];
|
|
980
|
+
artifactRef: string;
|
|
981
|
+
approved: boolean;
|
|
982
|
+
confidence: number;
|
|
983
|
+
metadata: Record<string, unknown>;
|
|
984
|
+
kind: "feedback";
|
|
985
|
+
feedback: string;
|
|
986
|
+
polarity: "neutral" | "positive" | "negative" | "mixed";
|
|
987
|
+
}[];
|
|
988
|
+
rewards: {
|
|
989
|
+
id: string;
|
|
990
|
+
taskId: string | null;
|
|
991
|
+
sourceRefs: string[];
|
|
992
|
+
artifactRef: string;
|
|
993
|
+
approved: boolean;
|
|
994
|
+
confidence: number;
|
|
995
|
+
metadata: Record<string, unknown>;
|
|
996
|
+
kind: "reward";
|
|
997
|
+
task: string;
|
|
998
|
+
rules: {
|
|
999
|
+
id: string;
|
|
1000
|
+
points: number;
|
|
1001
|
+
condition: string;
|
|
1002
|
+
}[];
|
|
1003
|
+
otherwisePoints: number;
|
|
1004
|
+
executable: boolean;
|
|
1005
|
+
}[];
|
|
1006
|
+
labels: {
|
|
1007
|
+
id: string;
|
|
1008
|
+
taskId: string | null;
|
|
1009
|
+
sourceRefs: string[];
|
|
1010
|
+
artifactRef: string;
|
|
1011
|
+
approved: boolean;
|
|
1012
|
+
confidence: number;
|
|
1013
|
+
metadata: Record<string, unknown>;
|
|
1014
|
+
kind: "label";
|
|
1015
|
+
labelKind: "rubric";
|
|
1016
|
+
task: string;
|
|
1017
|
+
criteria: {
|
|
1018
|
+
id: string;
|
|
1019
|
+
label: string;
|
|
1020
|
+
description: string;
|
|
1021
|
+
}[];
|
|
1022
|
+
calibrationExamples: {
|
|
1023
|
+
positive: string;
|
|
1024
|
+
negative: string;
|
|
1025
|
+
boundary: string;
|
|
1026
|
+
};
|
|
1027
|
+
}[];
|
|
1028
|
+
};
|
|
1029
|
+
publishedTasksetRef: {
|
|
1030
|
+
id: string;
|
|
1031
|
+
revision: number;
|
|
1032
|
+
contentHash: string;
|
|
1033
|
+
} | null;
|
|
1034
|
+
createdAt: string;
|
|
1035
|
+
updatedAt: string;
|
|
1036
|
+
metadata: Record<string, unknown>;
|
|
1037
|
+
datasetArtifact?: {
|
|
1038
|
+
schemaVersion: "openpond.datasetArtifact.v1";
|
|
1039
|
+
id: string;
|
|
1040
|
+
tasksetId: string;
|
|
1041
|
+
tasksetRevision: number;
|
|
1042
|
+
contentHash: string;
|
|
1043
|
+
format: "parquet";
|
|
1044
|
+
schema: {
|
|
1045
|
+
schemaVersion: "openpond.datasetSemanticSchema.v1";
|
|
1046
|
+
fields: {
|
|
1047
|
+
name: string;
|
|
1048
|
+
semanticRole: "metadata" | "row_id" | "cluster_id" | "prompt" | "messages" | "demonstration" | "chosen" | "rejected" | "expected_output" | "privileged_context" | "reward" | "feedback" | "tag";
|
|
1049
|
+
logicalType: "string" | "boolean" | "messages" | "integer" | "float" | "json";
|
|
1050
|
+
nullable: boolean;
|
|
1051
|
+
policy: "metadata" | "visible" | "privileged";
|
|
1052
|
+
}[];
|
|
1053
|
+
schemaHash: string;
|
|
1054
|
+
};
|
|
1055
|
+
shards: {
|
|
1056
|
+
id: string;
|
|
1057
|
+
split: "train" | "validation" | "test" | "frozen_eval";
|
|
1058
|
+
path: string;
|
|
1059
|
+
contentHash: string;
|
|
1060
|
+
schemaHash: string;
|
|
1061
|
+
sizeBytes: number;
|
|
1062
|
+
rowCount: number;
|
|
1063
|
+
rowGroupCount: number;
|
|
1064
|
+
}[];
|
|
1065
|
+
rowCount: number;
|
|
1066
|
+
splitCounts: Record<"train" | "validation" | "test" | "frozen_eval", number>;
|
|
1067
|
+
sourceReceiptRefs: string[];
|
|
1068
|
+
mappingHash: string;
|
|
1069
|
+
qualityReportHash: string;
|
|
1070
|
+
createdAt: string;
|
|
1071
|
+
} | null | undefined;
|
|
1072
|
+
};
|
|
1073
|
+
files: {
|
|
1074
|
+
path: string;
|
|
1075
|
+
contentHash: string;
|
|
1076
|
+
sizeBytes: number;
|
|
1077
|
+
base64: string;
|
|
1078
|
+
}[];
|
|
1079
|
+
contentHash: string;
|
|
1080
|
+
};
|
|
1081
|
+
publication: {
|
|
1082
|
+
tasksetId: string;
|
|
1083
|
+
release: {
|
|
1084
|
+
id: string;
|
|
1085
|
+
revision: number;
|
|
1086
|
+
contentHash: string;
|
|
1087
|
+
};
|
|
1088
|
+
packageHash: string;
|
|
1089
|
+
} | null;
|
|
1090
|
+
originProjectId?: string | undefined;
|
|
1091
|
+
ownerScope?: "workspace" | "personal" | undefined;
|
|
1092
|
+
}>;
|
|
1093
|
+
versions(id: string, options?: {
|
|
1094
|
+
beforeRevision?: number;
|
|
1095
|
+
signal?: AbortSignal;
|
|
1096
|
+
}): Promise<{
|
|
1097
|
+
teamId: string;
|
|
1098
|
+
datasetId: string;
|
|
1099
|
+
items: {
|
|
1100
|
+
workspaceRevision: number;
|
|
1101
|
+
publication: {
|
|
1102
|
+
tasksetId: string;
|
|
1103
|
+
release: {
|
|
1104
|
+
id: string;
|
|
1105
|
+
revision: number;
|
|
1106
|
+
contentHash: string;
|
|
1107
|
+
};
|
|
1108
|
+
packageHash: string;
|
|
1109
|
+
};
|
|
1110
|
+
publishedAt: string;
|
|
1111
|
+
}[];
|
|
1112
|
+
nextBeforeRevision: number | null;
|
|
1113
|
+
}>;
|
|
1114
|
+
version(id: string, workspaceRevision: number, signal?: AbortSignal): Promise<{
|
|
1115
|
+
schemaVersion: "openpond.datasetWorkspaceReceipt.v1";
|
|
1116
|
+
teamId: string;
|
|
1117
|
+
datasetId: string;
|
|
1118
|
+
revision: number;
|
|
1119
|
+
workspace: {
|
|
1120
|
+
schemaVersion: "openpond.tasksetDraftWorkspace.v1";
|
|
1121
|
+
draft: {
|
|
1122
|
+
schemaVersion: "openpond.tasksetDraft.v1";
|
|
1123
|
+
id: string;
|
|
1124
|
+
revision: number;
|
|
1125
|
+
profileId: string;
|
|
1126
|
+
modelScope: {
|
|
1127
|
+
modelId: string;
|
|
1128
|
+
expectedModelRevision: number;
|
|
1129
|
+
source?: {
|
|
1130
|
+
schemaVersion: "openpond.modelTasksetDraftPreparation.v1";
|
|
1131
|
+
draftId: string;
|
|
1132
|
+
requestHash: string;
|
|
1133
|
+
sourcePackageHash: string;
|
|
1134
|
+
sourceTasksetRef: {
|
|
1135
|
+
id: string;
|
|
1136
|
+
contentHash: string;
|
|
1137
|
+
revision: number;
|
|
1138
|
+
};
|
|
1139
|
+
tasksetId: string;
|
|
1140
|
+
tasksetRevision: number;
|
|
1141
|
+
lineage: {
|
|
1142
|
+
schemaVersion: "openpond.modelTasksetAuthoring.v1";
|
|
1143
|
+
owner: {
|
|
1144
|
+
scopeId: string;
|
|
1145
|
+
modelId: string;
|
|
1146
|
+
};
|
|
1147
|
+
root: {
|
|
1148
|
+
id: string;
|
|
1149
|
+
contentHash: string;
|
|
1150
|
+
revision: number;
|
|
1151
|
+
};
|
|
1152
|
+
parent: {
|
|
1153
|
+
id: string;
|
|
1154
|
+
contentHash: string;
|
|
1155
|
+
revision: number;
|
|
1156
|
+
};
|
|
1157
|
+
};
|
|
1158
|
+
authoringGraph?: "bound" | undefined;
|
|
1159
|
+
} | undefined;
|
|
1160
|
+
} | null;
|
|
1161
|
+
name: string;
|
|
1162
|
+
objective: string;
|
|
1163
|
+
purpose: "general" | "benchmark";
|
|
1164
|
+
benchmark: {
|
|
1165
|
+
schemaVersion: "openpond.tasksetBenchmark.v1";
|
|
1166
|
+
definitionId: string;
|
|
1167
|
+
releaseId: string;
|
|
1168
|
+
releaseHash: string;
|
|
1169
|
+
managedReleasePath: string;
|
|
1170
|
+
adaptationSplit: "train" | "validation" | "test" | "frozen_eval";
|
|
1171
|
+
evaluationSplit: "train" | "validation" | "test" | "frozen_eval";
|
|
1172
|
+
primaryMetric: "foreground_tokens" | "success_rate" | "latency_ms" | "cost_usd";
|
|
1173
|
+
qualityGate: "none" | "non_regression" | "all_pass";
|
|
1174
|
+
source: "builtin" | "imported";
|
|
1175
|
+
metadata: Record<string, unknown>;
|
|
1176
|
+
} | null;
|
|
1177
|
+
preferenceComparison: {
|
|
1178
|
+
schemaVersion: "openpond.tasksetPreferenceComparisonBinding.v1";
|
|
1179
|
+
releaseId: string;
|
|
1180
|
+
releaseHash: string;
|
|
1181
|
+
publishedAt: string;
|
|
1182
|
+
metadata: Record<string, unknown>;
|
|
1183
|
+
} | null;
|
|
1184
|
+
status: "draft" | "validating" | "needs_review" | "published";
|
|
1185
|
+
sourceRefs: ({
|
|
1186
|
+
id: string;
|
|
1187
|
+
profileId: string;
|
|
1188
|
+
title: string;
|
|
1189
|
+
sourceHash: string;
|
|
1190
|
+
occurredAt: string;
|
|
1191
|
+
licensingStatus: "blocked" | "pending" | "approved" | "review";
|
|
1192
|
+
secretScanStatus: "blocked" | "pending" | "passed";
|
|
1193
|
+
piiScanStatus: "blocked" | "pending" | "review" | "passed";
|
|
1194
|
+
metadata: Record<string, unknown>;
|
|
1195
|
+
schemaVersion: "openpond.huggingFaceDatasetSource.v1";
|
|
1196
|
+
kind: "huggingface";
|
|
1197
|
+
repositoryId: string;
|
|
1198
|
+
repositoryUrl: string;
|
|
1199
|
+
revision: string;
|
|
1200
|
+
configuration: string;
|
|
1201
|
+
upstreamSplits: string[];
|
|
1202
|
+
gated: boolean;
|
|
1203
|
+
private: boolean;
|
|
1204
|
+
declaredLicense: string | null;
|
|
1205
|
+
sourceFileHashes: string[];
|
|
1206
|
+
} | {
|
|
1207
|
+
id: string;
|
|
1208
|
+
profileId: string;
|
|
1209
|
+
title: string;
|
|
1210
|
+
sourceHash: string;
|
|
1211
|
+
occurredAt: string;
|
|
1212
|
+
licensingStatus: "blocked" | "pending" | "approved" | "review";
|
|
1213
|
+
secretScanStatus: "blocked" | "pending" | "passed";
|
|
1214
|
+
piiScanStatus: "blocked" | "pending" | "review" | "passed";
|
|
1215
|
+
metadata: Record<string, unknown>;
|
|
1216
|
+
schemaVersion: "openpond.uploadedFileDatasetSource.v1";
|
|
1217
|
+
kind: "uploaded_file";
|
|
1218
|
+
originalFileNames: string[];
|
|
1219
|
+
mediaTypes: string[];
|
|
1220
|
+
sourceFileHashes: string[];
|
|
1221
|
+
totalBytes: number;
|
|
1222
|
+
parserVersion: string;
|
|
1223
|
+
} | {
|
|
1224
|
+
id: string;
|
|
1225
|
+
profileId: string;
|
|
1226
|
+
title: string;
|
|
1227
|
+
sourceHash: string;
|
|
1228
|
+
occurredAt: string;
|
|
1229
|
+
licensingStatus: "blocked" | "pending" | "approved" | "review";
|
|
1230
|
+
secretScanStatus: "blocked" | "pending" | "passed";
|
|
1231
|
+
piiScanStatus: "blocked" | "pending" | "review" | "passed";
|
|
1232
|
+
metadata: Record<string, unknown>;
|
|
1233
|
+
schemaVersion: "openpond.generatedDatasetSource.v1";
|
|
1234
|
+
kind: "generated";
|
|
1235
|
+
generatorId: string;
|
|
1236
|
+
generatorVersion: string;
|
|
1237
|
+
seed: number;
|
|
1238
|
+
generatorHash: string;
|
|
1239
|
+
} | {
|
|
1240
|
+
id: string;
|
|
1241
|
+
profileId: string;
|
|
1242
|
+
title: string;
|
|
1243
|
+
sourceHash: string;
|
|
1244
|
+
occurredAt: string;
|
|
1245
|
+
licensingStatus: "blocked" | "pending" | "approved" | "review";
|
|
1246
|
+
secretScanStatus: "blocked" | "pending" | "passed";
|
|
1247
|
+
piiScanStatus: "blocked" | "pending" | "review" | "passed";
|
|
1248
|
+
metadata: Record<string, unknown>;
|
|
1249
|
+
schemaVersion: "openpond.learningBatchDatasetSource.v1";
|
|
1250
|
+
kind: "learning_batch";
|
|
1251
|
+
batch: {
|
|
1252
|
+
id: string;
|
|
1253
|
+
revision: number;
|
|
1254
|
+
contentHash: string;
|
|
1255
|
+
};
|
|
1256
|
+
taskDefinition: {
|
|
1257
|
+
id: string;
|
|
1258
|
+
revision: number;
|
|
1259
|
+
contentHash: string;
|
|
1260
|
+
};
|
|
1261
|
+
admittedBy: string;
|
|
1262
|
+
} | {
|
|
1263
|
+
schemaVersion: "openpond.trainingSource.v1";
|
|
1264
|
+
id: string;
|
|
1265
|
+
profileId: string;
|
|
1266
|
+
sessionId: string;
|
|
1267
|
+
turnIds: string[];
|
|
1268
|
+
workspaceId: string | null;
|
|
1269
|
+
sourceHash: string;
|
|
1270
|
+
clusterKey: string;
|
|
1271
|
+
title: string;
|
|
1272
|
+
occurredAt: string;
|
|
1273
|
+
consent: {
|
|
1274
|
+
status: "pending" | "granted" | "denied" | "revoked";
|
|
1275
|
+
scope: "metadata_only" | "selected_turns" | "full_session";
|
|
1276
|
+
grantedBy: string | null;
|
|
1277
|
+
grantedAt: string | null;
|
|
1278
|
+
purpose: "task_authoring_and_evaluation";
|
|
1279
|
+
};
|
|
1280
|
+
connectedAppIds: string[];
|
|
1281
|
+
secretScanStatus: "blocked" | "pending" | "passed";
|
|
1282
|
+
piiScanStatus: "blocked" | "pending" | "review" | "passed";
|
|
1283
|
+
licensingStatus: "blocked" | "pending" | "approved" | "review";
|
|
1284
|
+
metadata: Record<string, unknown>;
|
|
1285
|
+
})[];
|
|
1286
|
+
policy: {
|
|
1287
|
+
policyVisibleFields: string[];
|
|
1288
|
+
privilegedFields: string[];
|
|
1289
|
+
hiddenGraderRefs: string[];
|
|
1290
|
+
connectedAppScopes: string[];
|
|
1291
|
+
};
|
|
1292
|
+
environment: {
|
|
1293
|
+
protocolVersion: "openpond.taskEnvironment.v1";
|
|
1294
|
+
kind: "chat" | "agent" | "program" | "stateful_harness" | "work";
|
|
1295
|
+
entrypoint: string;
|
|
1296
|
+
stateful: boolean;
|
|
1297
|
+
deterministicSeeds: boolean;
|
|
1298
|
+
toolNames: string[];
|
|
1299
|
+
lifecycle: ("create" | "reset" | "step" | "grade" | "cleanup")[];
|
|
1300
|
+
defaultTimeoutMs: number;
|
|
1301
|
+
networkPolicy: "none" | "declared_read_only" | "declared_scoped";
|
|
1302
|
+
metadata: Record<string, unknown>;
|
|
1303
|
+
actionBindings?: {
|
|
1304
|
+
schemaVersion: "openpond.harnessActionBinding.v1";
|
|
1305
|
+
actionId: string;
|
|
1306
|
+
modelToolName: string;
|
|
1307
|
+
description: string;
|
|
1308
|
+
inputSchema: Record<string, unknown>;
|
|
1309
|
+
actionSchemaHash: string;
|
|
1310
|
+
agentRelease: {
|
|
1311
|
+
id: string;
|
|
1312
|
+
contentHash: string;
|
|
1313
|
+
};
|
|
1314
|
+
implementationHash: string;
|
|
1315
|
+
runtimeBindingId: string;
|
|
1316
|
+
capabilityReceiptHash: string;
|
|
1317
|
+
sideEffect: "read" | "write";
|
|
1318
|
+
studentVisible: boolean;
|
|
1319
|
+
timeoutMs: number;
|
|
1320
|
+
episodeArgumentBindings: {
|
|
1321
|
+
argument: string;
|
|
1322
|
+
source: "case_id";
|
|
1323
|
+
}[];
|
|
1324
|
+
}[] | undefined;
|
|
1325
|
+
resources?: {
|
|
1326
|
+
id: string;
|
|
1327
|
+
kind: "file" | "configuration" | "catalog" | "code_module";
|
|
1328
|
+
path: string;
|
|
1329
|
+
visibility: "privileged" | "policy_visible" | "policy_hidden";
|
|
1330
|
+
required: boolean;
|
|
1331
|
+
metadata: Record<string, unknown>;
|
|
1332
|
+
mediaType?: string | null | undefined;
|
|
1333
|
+
}[] | undefined;
|
|
1334
|
+
};
|
|
1335
|
+
output: {
|
|
1336
|
+
mode: "text" | "structured_json" | "artifacts";
|
|
1337
|
+
jsonSchema: Record<string, unknown> | null;
|
|
1338
|
+
renderer: {
|
|
1339
|
+
module: string;
|
|
1340
|
+
exportName: string;
|
|
1341
|
+
configRef: string | null;
|
|
1342
|
+
} | null;
|
|
1343
|
+
};
|
|
1344
|
+
capabilities: {
|
|
1345
|
+
schemaVersion: "openpond.tasksetCapabilities.v1";
|
|
1346
|
+
taskKind: "chat" | "single_agent" | "multi_agent" | "custom_program";
|
|
1347
|
+
supportedSignals: ("label" | "demonstration" | "reward" | "feedback" | "preference" | "correction")[];
|
|
1348
|
+
compatibleMethods: ("none" | "retrieval" | "sft" | "dpo" | "grpo" | "ppo" | "sdft" | "opd" | "opsd" | "sdpo")[];
|
|
1349
|
+
rewardKinds: ("none" | "exact" | "deterministic" | "model_judge" | "human")[];
|
|
1350
|
+
requiresTools: boolean;
|
|
1351
|
+
requiresState: boolean;
|
|
1352
|
+
requiresPrivilegedGrading: boolean;
|
|
1353
|
+
environmentPlacements: ("local" | "remote" | "colocated" | "provider_native")[];
|
|
1354
|
+
exportable: boolean;
|
|
1355
|
+
portabilityBlockers: string[];
|
|
1356
|
+
};
|
|
1357
|
+
metrics: {
|
|
1358
|
+
schemaVersion: "openpond.tasksetMetricPolicy.v1";
|
|
1359
|
+
primaryMetric: string;
|
|
1360
|
+
aggregation: "custom" | "mean_score" | "pass_rate" | "weighted_mean";
|
|
1361
|
+
missingReward: "zero" | "exclude";
|
|
1362
|
+
customAggregator: {
|
|
1363
|
+
module: string;
|
|
1364
|
+
exportName: string;
|
|
1365
|
+
contentHash: string;
|
|
1366
|
+
timeoutMs: number;
|
|
1367
|
+
networkPolicy: "none";
|
|
1368
|
+
} | null;
|
|
1369
|
+
taskWeights?: Record<string, number> | undefined;
|
|
1370
|
+
};
|
|
1371
|
+
review: {
|
|
1372
|
+
enabled: boolean;
|
|
1373
|
+
candidateCount: number;
|
|
1374
|
+
minimumSamples: number;
|
|
1375
|
+
allowTies: boolean;
|
|
1376
|
+
allowRejectAll: boolean;
|
|
1377
|
+
rubric: string;
|
|
1378
|
+
criteria: {
|
|
1379
|
+
id: string;
|
|
1380
|
+
label: string;
|
|
1381
|
+
description: string;
|
|
1382
|
+
weight: number;
|
|
1383
|
+
}[];
|
|
1384
|
+
};
|
|
1385
|
+
tasks: {
|
|
1386
|
+
schemaVersion: "openpond.taskData.v1";
|
|
1387
|
+
id: string;
|
|
1388
|
+
clusterKey: string;
|
|
1389
|
+
split: "train" | "validation" | "test" | "frozen_eval";
|
|
1390
|
+
input: Record<string, unknown>;
|
|
1391
|
+
expectedOutput: Record<string, unknown> | null;
|
|
1392
|
+
policyVisibleContext: Record<string, unknown>;
|
|
1393
|
+
privilegedContextRef: string | null;
|
|
1394
|
+
tags: string[];
|
|
1395
|
+
metadata: Record<string, unknown>;
|
|
1396
|
+
sourceRefs: string[];
|
|
1397
|
+
assets?: {
|
|
1398
|
+
id: string;
|
|
1399
|
+
sourceRefId: string;
|
|
1400
|
+
artifactRef: string;
|
|
1401
|
+
fileName: string;
|
|
1402
|
+
mediaType: string;
|
|
1403
|
+
sha256: string;
|
|
1404
|
+
sizeBytes: number;
|
|
1405
|
+
split: "train" | "validation" | "test" | "frozen_eval";
|
|
1406
|
+
metadata: Record<string, unknown>;
|
|
1407
|
+
}[] | undefined;
|
|
1408
|
+
resourceRefs?: string[] | undefined;
|
|
1409
|
+
requiredOutputs?: {
|
|
1410
|
+
path: string;
|
|
1411
|
+
mediaType: string;
|
|
1412
|
+
metadata: Record<string, unknown>;
|
|
1413
|
+
schemaRef?: string | null | undefined;
|
|
1414
|
+
maxBytes?: number | undefined;
|
|
1415
|
+
}[] | undefined;
|
|
1416
|
+
}[];
|
|
1417
|
+
graders: ({
|
|
1418
|
+
id: string;
|
|
1419
|
+
version: string;
|
|
1420
|
+
label: string;
|
|
1421
|
+
weight: number;
|
|
1422
|
+
hardGate: boolean;
|
|
1423
|
+
rewardEligible: boolean;
|
|
1424
|
+
privileged: boolean;
|
|
1425
|
+
metadata: Record<string, unknown>;
|
|
1426
|
+
kind: "file" | "content" | "test" | "schema" | "diff" | "runtime_event" | "state";
|
|
1427
|
+
config: Record<string, unknown>;
|
|
1428
|
+
} | {
|
|
1429
|
+
id: string;
|
|
1430
|
+
version: string;
|
|
1431
|
+
label: string;
|
|
1432
|
+
weight: number;
|
|
1433
|
+
hardGate: boolean;
|
|
1434
|
+
rewardEligible: boolean;
|
|
1435
|
+
privileged: boolean;
|
|
1436
|
+
metadata: Record<string, unknown>;
|
|
1437
|
+
kind: "model_judge";
|
|
1438
|
+
rubric: string;
|
|
1439
|
+
judge: {
|
|
1440
|
+
providerId: "openpond" | "codex" | "anthropic" | "openai" | "xai" | "google" | "openrouter" | "deepseek" | "zai" | "moonshot" | "together" | "groq" | "custom-openai-compatible";
|
|
1441
|
+
modelId: string;
|
|
1442
|
+
};
|
|
1443
|
+
calibrationFixtureRefs: string[];
|
|
1444
|
+
calibrationStatus: "failed" | "pending" | "passed";
|
|
1445
|
+
temperature: number;
|
|
1446
|
+
} | {
|
|
1447
|
+
id: string;
|
|
1448
|
+
version: string;
|
|
1449
|
+
label: string;
|
|
1450
|
+
weight: number;
|
|
1451
|
+
hardGate: boolean;
|
|
1452
|
+
rewardEligible: boolean;
|
|
1453
|
+
privileged: boolean;
|
|
1454
|
+
metadata: Record<string, unknown>;
|
|
1455
|
+
kind: "human";
|
|
1456
|
+
rubric: string;
|
|
1457
|
+
reviewerRole: string;
|
|
1458
|
+
} | {
|
|
1459
|
+
id: string;
|
|
1460
|
+
version: string;
|
|
1461
|
+
label: string;
|
|
1462
|
+
weight: number;
|
|
1463
|
+
hardGate: boolean;
|
|
1464
|
+
rewardEligible: boolean;
|
|
1465
|
+
privileged: boolean;
|
|
1466
|
+
metadata: Record<string, unknown>;
|
|
1467
|
+
kind: "custom_verifier";
|
|
1468
|
+
module: string;
|
|
1469
|
+
exportName: string;
|
|
1470
|
+
timeoutMs: number;
|
|
1471
|
+
networkPolicy: "none";
|
|
1472
|
+
runtime?: "isolated_javascript" | "sandbox_process" | undefined;
|
|
1473
|
+
})[];
|
|
1474
|
+
graderFixtures: {
|
|
1475
|
+
id: string;
|
|
1476
|
+
taskId: string;
|
|
1477
|
+
label: "positive" | "negative" | "boundary" | "adversarial" | "prompt_injection" | "infrastructure_failure";
|
|
1478
|
+
output: Record<string, unknown>;
|
|
1479
|
+
infrastructureError: string | null;
|
|
1480
|
+
expectedPassed: boolean;
|
|
1481
|
+
expectedRewardEligible: boolean;
|
|
1482
|
+
metadata: Record<string, unknown>;
|
|
1483
|
+
}[];
|
|
1484
|
+
learningSignals: {
|
|
1485
|
+
demonstrations: {
|
|
1486
|
+
id: string;
|
|
1487
|
+
taskId: string | null;
|
|
1488
|
+
sourceRefs: string[];
|
|
1489
|
+
artifactRef: string;
|
|
1490
|
+
approved: boolean;
|
|
1491
|
+
confidence: number;
|
|
1492
|
+
metadata: Record<string, unknown>;
|
|
1493
|
+
kind: "demonstration";
|
|
1494
|
+
prompt: string | null;
|
|
1495
|
+
response: string | null;
|
|
1496
|
+
}[];
|
|
1497
|
+
preferences: {
|
|
1498
|
+
id: string;
|
|
1499
|
+
taskId: string | null;
|
|
1500
|
+
sourceRefs: string[];
|
|
1501
|
+
artifactRef: string;
|
|
1502
|
+
approved: boolean;
|
|
1503
|
+
confidence: number;
|
|
1504
|
+
metadata: Record<string, unknown>;
|
|
1505
|
+
kind: "preference";
|
|
1506
|
+
prompt: string;
|
|
1507
|
+
chosen: string;
|
|
1508
|
+
rejected: string;
|
|
1509
|
+
rationale: string | null;
|
|
1510
|
+
}[];
|
|
1511
|
+
corrections: {
|
|
1512
|
+
id: string;
|
|
1513
|
+
taskId: string | null;
|
|
1514
|
+
sourceRefs: string[];
|
|
1515
|
+
artifactRef: string;
|
|
1516
|
+
approved: boolean;
|
|
1517
|
+
confidence: number;
|
|
1518
|
+
metadata: Record<string, unknown>;
|
|
1519
|
+
kind: "correction";
|
|
1520
|
+
original: string;
|
|
1521
|
+
corrected: string;
|
|
1522
|
+
rationale: string | null;
|
|
1523
|
+
}[];
|
|
1524
|
+
feedback: {
|
|
1525
|
+
id: string;
|
|
1526
|
+
taskId: string | null;
|
|
1527
|
+
sourceRefs: string[];
|
|
1528
|
+
artifactRef: string;
|
|
1529
|
+
approved: boolean;
|
|
1530
|
+
confidence: number;
|
|
1531
|
+
metadata: Record<string, unknown>;
|
|
1532
|
+
kind: "feedback";
|
|
1533
|
+
feedback: string;
|
|
1534
|
+
polarity: "neutral" | "positive" | "negative" | "mixed";
|
|
1535
|
+
}[];
|
|
1536
|
+
rewards: {
|
|
1537
|
+
id: string;
|
|
1538
|
+
taskId: string | null;
|
|
1539
|
+
sourceRefs: string[];
|
|
1540
|
+
artifactRef: string;
|
|
1541
|
+
approved: boolean;
|
|
1542
|
+
confidence: number;
|
|
1543
|
+
metadata: Record<string, unknown>;
|
|
1544
|
+
kind: "reward";
|
|
1545
|
+
task: string;
|
|
1546
|
+
rules: {
|
|
1547
|
+
id: string;
|
|
1548
|
+
points: number;
|
|
1549
|
+
condition: string;
|
|
1550
|
+
}[];
|
|
1551
|
+
otherwisePoints: number;
|
|
1552
|
+
executable: boolean;
|
|
1553
|
+
}[];
|
|
1554
|
+
labels: {
|
|
1555
|
+
id: string;
|
|
1556
|
+
taskId: string | null;
|
|
1557
|
+
sourceRefs: string[];
|
|
1558
|
+
artifactRef: string;
|
|
1559
|
+
approved: boolean;
|
|
1560
|
+
confidence: number;
|
|
1561
|
+
metadata: Record<string, unknown>;
|
|
1562
|
+
kind: "label";
|
|
1563
|
+
labelKind: "rubric";
|
|
1564
|
+
task: string;
|
|
1565
|
+
criteria: {
|
|
1566
|
+
id: string;
|
|
1567
|
+
label: string;
|
|
1568
|
+
description: string;
|
|
1569
|
+
}[];
|
|
1570
|
+
calibrationExamples: {
|
|
1571
|
+
positive: string;
|
|
1572
|
+
negative: string;
|
|
1573
|
+
boundary: string;
|
|
1574
|
+
};
|
|
1575
|
+
}[];
|
|
1576
|
+
};
|
|
1577
|
+
publishedTasksetRef: {
|
|
1578
|
+
id: string;
|
|
1579
|
+
revision: number;
|
|
1580
|
+
contentHash: string;
|
|
1581
|
+
} | null;
|
|
1582
|
+
createdAt: string;
|
|
1583
|
+
updatedAt: string;
|
|
1584
|
+
metadata: Record<string, unknown>;
|
|
1585
|
+
datasetArtifact?: {
|
|
1586
|
+
schemaVersion: "openpond.datasetArtifact.v1";
|
|
1587
|
+
id: string;
|
|
1588
|
+
tasksetId: string;
|
|
1589
|
+
tasksetRevision: number;
|
|
1590
|
+
contentHash: string;
|
|
1591
|
+
format: "parquet";
|
|
1592
|
+
schema: {
|
|
1593
|
+
schemaVersion: "openpond.datasetSemanticSchema.v1";
|
|
1594
|
+
fields: {
|
|
1595
|
+
name: string;
|
|
1596
|
+
semanticRole: "metadata" | "row_id" | "cluster_id" | "prompt" | "messages" | "demonstration" | "chosen" | "rejected" | "expected_output" | "privileged_context" | "reward" | "feedback" | "tag";
|
|
1597
|
+
logicalType: "string" | "boolean" | "messages" | "integer" | "float" | "json";
|
|
1598
|
+
nullable: boolean;
|
|
1599
|
+
policy: "metadata" | "visible" | "privileged";
|
|
1600
|
+
}[];
|
|
1601
|
+
schemaHash: string;
|
|
1602
|
+
};
|
|
1603
|
+
shards: {
|
|
1604
|
+
id: string;
|
|
1605
|
+
split: "train" | "validation" | "test" | "frozen_eval";
|
|
1606
|
+
path: string;
|
|
1607
|
+
contentHash: string;
|
|
1608
|
+
schemaHash: string;
|
|
1609
|
+
sizeBytes: number;
|
|
1610
|
+
rowCount: number;
|
|
1611
|
+
rowGroupCount: number;
|
|
1612
|
+
}[];
|
|
1613
|
+
rowCount: number;
|
|
1614
|
+
splitCounts: Record<"train" | "validation" | "test" | "frozen_eval", number>;
|
|
1615
|
+
sourceReceiptRefs: string[];
|
|
1616
|
+
mappingHash: string;
|
|
1617
|
+
qualityReportHash: string;
|
|
1618
|
+
createdAt: string;
|
|
1619
|
+
} | null | undefined;
|
|
1620
|
+
};
|
|
1621
|
+
files: {
|
|
1622
|
+
path: string;
|
|
1623
|
+
contentHash: string;
|
|
1624
|
+
sizeBytes: number;
|
|
1625
|
+
base64: string;
|
|
1626
|
+
}[];
|
|
1627
|
+
contentHash: string;
|
|
1628
|
+
};
|
|
1629
|
+
publication: {
|
|
1630
|
+
tasksetId: string;
|
|
1631
|
+
release: {
|
|
1632
|
+
id: string;
|
|
1633
|
+
revision: number;
|
|
1634
|
+
contentHash: string;
|
|
1635
|
+
};
|
|
1636
|
+
packageHash: string;
|
|
1637
|
+
} | null;
|
|
1638
|
+
originProjectId?: string | undefined;
|
|
1639
|
+
ownerScope?: "workspace" | "personal" | undefined;
|
|
563
1640
|
}>;
|
|
564
1641
|
save(input: z.input<typeof DatasetWorkspaceWriteSchema>, signal?: AbortSignal): Promise<{
|
|
565
1642
|
schemaVersion: "openpond.datasetWorkspaceReceipt.v1";
|
|
@@ -1086,6 +2163,7 @@ export declare class OpenPondDatasetWorkspaceClient {
|
|
|
1086
2163
|
packageHash: string;
|
|
1087
2164
|
} | null;
|
|
1088
2165
|
originProjectId?: string | undefined;
|
|
2166
|
+
ownerScope?: "workspace" | "personal" | undefined;
|
|
1089
2167
|
}>;
|
|
1090
2168
|
validate(id: string, expectedRevision: number, signal?: AbortSignal): Promise<{
|
|
1091
2169
|
teamId: string;
|
|
@@ -1624,6 +2702,7 @@ export declare class OpenPondDatasetWorkspaceClient {
|
|
|
1624
2702
|
packageHash: string;
|
|
1625
2703
|
} | null;
|
|
1626
2704
|
originProjectId?: string | undefined;
|
|
2705
|
+
ownerScope?: "workspace" | "personal" | undefined;
|
|
1627
2706
|
}>;
|
|
1628
2707
|
private readback;
|
|
1629
2708
|
private request;
|