openpond-sdk 0.12.2 → 0.12.4
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/actions-local.js +2 -1
- package/dist/actions-local.js.map +1 -7
- package/dist/actions.js +2 -1
- package/dist/actions.js.map +1 -7
- package/dist/advanced-refiner-evaluations.js +13 -12
- package/dist/advanced-refiner-evaluations.js.map +1 -7
- package/dist/candidate-evaluations.js +9 -8
- package/dist/candidate-evaluations.js.map +1 -7
- package/dist/{chunk-MLFKQ7E7.js → chunk-24N6A6SF.js} +4 -4
- package/dist/chunk-24N6A6SF.js.map +1 -0
- package/dist/{chunk-YCQVVFE5.js → chunk-3B5KSH3V.js} +2 -2
- package/dist/chunk-3B5KSH3V.js.map +1 -0
- package/dist/{chunk-2DMYAJLS.js → chunk-3DWPLYFJ.js} +2 -2
- package/dist/chunk-3DWPLYFJ.js.map +1 -0
- package/dist/{chunk-B7VYE7YQ.js → chunk-53LLNDXF.js} +3 -3
- package/dist/chunk-53LLNDXF.js.map +1 -0
- package/dist/{chunk-VSRTJDVB.js → chunk-B5XTGEBE.js} +25 -13
- package/dist/chunk-B5XTGEBE.js.map +1 -0
- package/dist/{chunk-HKWX624G.js → chunk-BVMR5OFW.js} +5 -5
- package/dist/chunk-BVMR5OFW.js.map +1 -0
- package/dist/{chunk-3KXDH6OB.js → chunk-CL3IOOVL.js} +1 -1
- package/dist/chunk-CL3IOOVL.js.map +1 -0
- package/dist/chunk-CMWZZZGW.js +255 -0
- package/dist/chunk-CMWZZZGW.js.map +1 -0
- package/dist/{chunk-VOL3WES2.js → chunk-DF2QXVZ3.js} +14 -169
- package/dist/chunk-DF2QXVZ3.js.map +1 -0
- package/dist/{chunk-AG5DDL7T.js → chunk-DH6O4JFB.js} +2 -2
- package/dist/chunk-DH6O4JFB.js.map +1 -0
- package/dist/{chunk-7GYXC2AI.js → chunk-EFUT3XKW.js} +4 -4
- package/dist/chunk-EFUT3XKW.js.map +1 -0
- package/dist/{chunk-DBRRQASP.js → chunk-ETX25FTB.js} +3 -3
- package/dist/chunk-ETX25FTB.js.map +1 -0
- package/dist/{chunk-MEKWESBF.js → chunk-FBZTDDRC.js} +2 -2
- package/dist/chunk-FBZTDDRC.js.map +1 -0
- package/dist/{chunk-7AZDNZP6.js → chunk-GAIDBPFY.js} +353 -90
- package/dist/chunk-GAIDBPFY.js.map +1 -0
- package/dist/{chunk-YVUIHH6Z.js → chunk-GPQLBVQC.js} +1 -1
- package/dist/chunk-GPQLBVQC.js.map +1 -0
- package/dist/{chunk-KAE7AMYE.js → chunk-H3367ISL.js} +2 -2
- package/dist/chunk-H3367ISL.js.map +1 -0
- package/dist/{chunk-COPVBJPL.js → chunk-H3JZ3P5T.js} +2 -1
- package/dist/chunk-H3JZ3P5T.js.map +1 -0
- package/dist/{chunk-3LP7DNLQ.js → chunk-H4NTJFEE.js} +2 -2
- package/dist/chunk-H4NTJFEE.js.map +1 -0
- package/dist/{chunk-TCHAWJMB.js → chunk-HGXVOFCL.js} +2 -2
- package/dist/chunk-HGXVOFCL.js.map +1 -0
- package/dist/{chunk-CEZEVILU.js → chunk-HJ2AIAKX.js} +1 -1
- package/dist/chunk-HJ2AIAKX.js.map +1 -0
- package/dist/{chunk-ENLQCKY6.js → chunk-I54WQBXL.js} +8 -8
- package/dist/chunk-I54WQBXL.js.map +1 -0
- package/dist/{chunk-AV5JV26B.js → chunk-K5LWSWV7.js} +4 -4
- package/dist/chunk-K5LWSWV7.js.map +1 -0
- package/dist/{chunk-IWT6JUZD.js → chunk-L6EJX2VQ.js} +2 -2
- package/dist/chunk-L6EJX2VQ.js.map +1 -0
- package/dist/{chunk-WVISXXIL.js → chunk-LHX7ENKZ.js} +3 -3
- package/dist/chunk-LHX7ENKZ.js.map +1 -0
- package/dist/{chunk-WNH4DOEC.js → chunk-LR5CTOC5.js} +30 -299
- package/dist/chunk-LR5CTOC5.js.map +1 -0
- package/dist/chunk-MO4HAJZG.js +8709 -0
- package/dist/chunk-MO4HAJZG.js.map +1 -0
- package/dist/{chunk-T4APHDES.js → chunk-O6FKHAWN.js} +1 -1
- package/dist/chunk-O6FKHAWN.js.map +1 -0
- package/dist/{chunk-O5ARJVT5.js → chunk-OR3TMIJN.js} +5 -5
- package/dist/chunk-OR3TMIJN.js.map +1 -0
- package/dist/{chunk-P27DQ65Z.js → chunk-OVGDDKO2.js} +1 -1
- package/dist/chunk-OVGDDKO2.js.map +1 -0
- package/dist/{chunk-O4GE6VTR.js → chunk-PWW5QKZS.js} +2 -2
- package/dist/chunk-PWW5QKZS.js.map +1 -0
- package/dist/{chunk-DUKW3R7O.js → chunk-Q4TKI2MV.js} +2 -2
- package/dist/chunk-Q4TKI2MV.js.map +1 -0
- package/dist/{chunk-L5PQPAU6.js → chunk-QRS7Y3EM.js} +7 -3
- package/dist/chunk-QRS7Y3EM.js.map +1 -0
- package/dist/{chunk-YVFERCUC.js → chunk-RRKRA237.js} +83 -5
- package/dist/chunk-RRKRA237.js.map +1 -0
- package/dist/{chunk-2U54EOXJ.js → chunk-TK3DXH7A.js} +1 -1
- package/dist/chunk-TK3DXH7A.js.map +1 -0
- package/dist/{chunk-CK3ONRRC.js → chunk-U7XWIOSB.js} +3 -3
- package/dist/chunk-U7XWIOSB.js.map +1 -0
- package/dist/{chunk-DG5CRENU.js → chunk-VHGC44HA.js} +5 -5
- package/dist/chunk-VHGC44HA.js.map +1 -0
- package/dist/{chunk-7WUJBFCH.js → chunk-VN5V2E2S.js} +24 -5
- package/dist/chunk-VN5V2E2S.js.map +1 -0
- package/dist/{chunk-YSKUXQ2N.js → chunk-WMIKJGXW.js} +3 -3
- package/dist/chunk-WMIKJGXW.js.map +1 -0
- package/dist/{chunk-ROBAX5TN.js → chunk-WQFU635Q.js} +8 -6
- package/dist/chunk-WQFU635Q.js.map +1 -0
- package/dist/{chunk-D4WTE6ZA.js → chunk-XON75UFN.js} +3 -3
- package/dist/chunk-XON75UFN.js.map +1 -0
- package/dist/chunk-YLJ4XMA6.js +35 -0
- package/dist/chunk-YLJ4XMA6.js.map +1 -0
- package/dist/{chunk-K3JCP4FO.js → chunk-Z252BNAZ.js} +1 -1
- package/dist/chunk-Z252BNAZ.js.map +1 -0
- package/dist/{chunk-OEPYLZXK.js → chunk-ZHQK3MOM.js} +4 -4
- package/dist/chunk-ZHQK3MOM.js.map +1 -0
- package/dist/{chunk-6Q4I33CH.js → chunk-ZIFZFP5I.js} +5 -5
- package/dist/chunk-ZIFZFP5I.js.map +1 -0
- package/dist/connected-evidence.js +22 -21
- package/dist/connected-evidence.js.map +1 -7
- package/dist/dataset-commerce.js +2 -0
- package/dist/dataset-commerce.js.map +1 -7
- package/dist/dataset-marketplace.js +2 -0
- package/dist/dataset-marketplace.js.map +1 -7
- package/dist/dataset-workspaces.js +25 -23
- package/dist/dataset-workspaces.js.map +1 -7
- package/dist/experiment-evaluation-schedules.js +28 -27
- package/dist/experiment-evaluation-schedules.js.map +1 -7
- package/dist/experiment-improvements.js +12 -11
- package/dist/experiment-improvements.js.map +1 -7
- package/dist/experiments.js +31 -26
- package/dist/experiments.js.map +1 -7
- package/dist/human-review.js +3 -2
- package/dist/human-review.js.map +1 -7
- package/dist/index.js +98 -50
- package/dist/index.js.map +1 -7
- package/dist/learning.js +4 -3
- package/dist/learning.js.map +1 -7
- package/dist/model-batch-review.js +2 -1
- package/dist/model-batch-review.js.map +1 -7
- package/dist/model-projects.js +3 -2
- package/dist/model-projects.js.map +1 -7
- package/dist/model-starter-attempts.js +5 -4
- package/dist/model-starter-attempts.js.map +1 -7
- package/dist/model-starter-catalog.js +15 -14
- package/dist/model-starter-catalog.js.map +1 -7
- package/dist/model-starters.js +15 -14
- package/dist/model-starters.js.map +1 -7
- package/dist/model-taskset-authoring.js +2 -1
- package/dist/model-taskset-authoring.js.map +1 -7
- package/dist/model-taskset-runs.js +6 -5
- package/dist/model-taskset-runs.js.map +1 -7
- package/dist/post-training.js +9 -8
- package/dist/post-training.js.map +1 -7
- package/dist/profile-actions.js +3 -2
- package/dist/profile-actions.js.map +1 -7
- package/dist/project-actions.js +3 -2
- package/dist/project-actions.js.map +1 -7
- package/dist/refiner.js +3 -2
- package/dist/refiner.js.map +1 -7
- package/dist/taskset-catalog.js +2 -1
- package/dist/taskset-catalog.js.map +1 -7
- package/dist/taskset-drafts.js +27 -26
- package/dist/taskset-drafts.js.map +1 -7
- package/dist/taskset-packages.js +29 -26
- package/dist/taskset-packages.js.map +1 -7
- package/dist/training-bundle.js +195 -34
- package/dist/training-bundle.js.map +1 -7
- package/dist/training-projects.js +2 -1
- package/dist/training-projects.js.map +1 -7
- package/dist/training.js +17 -10
- package/dist/training.js.map +1 -7
- package/dist/types/packages/cloud/src/api/core.d.ts.map +1 -1
- package/dist/types/packages/cloud/src/sandbox/client-handles.d.ts.map +1 -1
- package/dist/types/packages/cloud/src/sandbox/client.d.ts.map +1 -1
- package/dist/types/packages/cloud/src/sandbox/sandbox-instance-client.d.ts.map +1 -1
- package/dist/types/packages/cloud/src/sandbox/types/runtime-profiles.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/advanced-refiner-evaluations.d.ts +487 -487
- package/dist/types/packages/sdk/src/advanced-refiner-evaluations.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/candidate-evaluation-client.d.ts +381 -348
- package/dist/types/packages/sdk/src/candidate-evaluation-client.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/candidate-evaluations.d.ts +86 -54
- package/dist/types/packages/sdk/src/candidate-evaluations.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/connected-collection-status.d.ts +5 -5
- package/dist/types/packages/sdk/src/connected-dataset-source.d.ts +10 -10
- package/dist/types/packages/sdk/src/connected-evidence-client.d.ts +464 -460
- package/dist/types/packages/sdk/src/connected-evidence-client.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/connected-evidence-contracts.d.ts +23 -23
- package/dist/types/packages/sdk/src/connected-recorded-execution.d.ts +77 -74
- package/dist/types/packages/sdk/src/connected-recorded-execution.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/connected-sync.d.ts +30 -30
- package/dist/types/packages/sdk/src/connected-sync.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/dataset-commerce-contracts.d.ts +14 -14
- package/dist/types/packages/sdk/src/dataset-commerce-files.d.ts +5 -5
- package/dist/types/packages/sdk/src/dataset-commerce-files.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/dataset-commerce-responses.d.ts +82 -82
- package/dist/types/packages/sdk/src/dataset-commerce.d.ts +58 -58
- package/dist/types/packages/sdk/src/dataset-commerce.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/dataset-marketplace.d.ts +4 -4
- package/dist/types/packages/sdk/src/dataset-marketplace.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/dataset-population.d.ts +10 -10
- package/dist/types/packages/sdk/src/dataset-population.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/dataset-preparation-authoring-contracts.d.ts +172 -172
- package/dist/types/packages/sdk/src/dataset-preparation-client.d.ts +47 -47
- package/dist/types/packages/sdk/src/dataset-preparation-client.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/dataset-preparation-contracts.d.ts +90 -90
- package/dist/types/packages/sdk/src/dataset-preparation-qualification-contracts.d.ts +7 -7
- package/dist/types/packages/sdk/src/dataset-preparation-research-contracts.d.ts +26 -26
- package/dist/types/packages/sdk/src/dataset-preparation-starting-state.d.ts +8 -8
- package/dist/types/packages/sdk/src/dataset-preparation-workspace.d.ts +111 -110
- package/dist/types/packages/sdk/src/dataset-preparation-workspace.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/dataset-workspace-client.d.ts +896 -889
- 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 +225 -223
- package/dist/types/packages/sdk/src/dataset-workspace-contracts.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/dataset-workspace-operations.d.ts +226 -224
- package/dist/types/packages/sdk/src/dataset-workspace-operations.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/evaluation-coordination-http.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 +2090 -1969
- package/dist/types/packages/sdk/src/experiment-client.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/experiment-contracts.d.ts +177 -113
- package/dist/types/packages/sdk/src/experiment-contracts.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/experiment-evaluation-schedules.d.ts +597 -505
- package/dist/types/packages/sdk/src/experiment-evaluation-schedules.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/experiment-execution-selection.d.ts +5 -5
- package/dist/types/packages/sdk/src/experiment-execution-selection.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/experiment-field-mappings.d.ts +8 -8
- package/dist/types/packages/sdk/src/experiment-grading-contracts.d.ts +16 -16
- package/dist/types/packages/sdk/src/experiment-improvement-client.d.ts +2524 -2524
- package/dist/types/packages/sdk/src/experiment-improvement-client.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/experiment-improvement-contracts.d.ts +1557 -1557
- package/dist/types/packages/sdk/src/experiment-improvement-hosted-adoption.d.ts +150 -150
- package/dist/types/packages/sdk/src/experiment-inspection.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/experiment-run-contracts.d.ts +41 -25
- package/dist/types/packages/sdk/src/experiment-run-contracts.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/experiment-run-details.d.ts +315 -240
- package/dist/types/packages/sdk/src/experiment-run-details.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/experiment-scoring-contracts.d.ts +21 -21
- package/dist/types/packages/sdk/src/grader-inspection.d.ts +104 -104
- package/dist/types/packages/sdk/src/grader-inspection.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/harness-experiment-policy.d.ts +50 -20
- package/dist/types/packages/sdk/src/harness-experiment-policy.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/human-review-client.d.ts +288 -288
- package/dist/types/packages/sdk/src/human-review-client.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/index.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/learning-batch-assets.d.ts +15 -13
- package/dist/types/packages/sdk/src/learning-batch-assets.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/learning-client.d.ts +5 -5
- package/dist/types/packages/sdk/src/learning-client.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/learning-credentials.d.ts +4 -4
- package/dist/types/packages/sdk/src/learning-credentials.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/managed-training-preparation.d.ts +71 -69
- package/dist/types/packages/sdk/src/managed-training-preparation.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/model-batch-review-contracts.d.ts +65 -61
- package/dist/types/packages/sdk/src/model-batch-review-contracts.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/model-batch-review-preparation.d.ts +74 -72
- package/dist/types/packages/sdk/src/model-batch-review-preparation.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/model-batch-review.d.ts +96 -92
- package/dist/types/packages/sdk/src/model-batch-review.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/model-learning-intent.d.ts +1 -1
- package/dist/types/packages/sdk/src/model-learning-policy.d.ts +247 -247
- package/dist/types/packages/sdk/src/model-learning-policy.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/model-projects.d.ts +324 -324
- package/dist/types/packages/sdk/src/model-projects.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/model-starter-attempts.d.ts +187 -187
- package/dist/types/packages/sdk/src/model-starter-attempts.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/model-starter-catalog.d.ts +181 -179
- package/dist/types/packages/sdk/src/model-starter-catalog.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/model-starter-execution.d.ts +40 -40
- package/dist/types/packages/sdk/src/model-starters.d.ts +291 -287
- package/dist/types/packages/sdk/src/model-starters.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/model-taskset-authoring-contracts.d.ts +3 -3
- package/dist/types/packages/sdk/src/model-taskset-derivation.d.ts +114 -112
- package/dist/types/packages/sdk/src/model-taskset-derivation.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/model-taskset-resources.d.ts +18 -18
- package/dist/types/packages/sdk/src/model-taskset-runs-contracts.d.ts +191 -127
- package/dist/types/packages/sdk/src/model-taskset-runs-contracts.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/model-taskset-runs.d.ts +868 -824
- package/dist/types/packages/sdk/src/model-taskset-runs.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/post-training.d.ts +448 -372
- package/dist/types/packages/sdk/src/post-training.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/profile-actions.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/profile-evaluation-discovery.d.ts +150 -150
- package/dist/types/packages/sdk/src/profile-evaluation-discovery.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/profile-external-dataset-package.d.ts +254 -251
- package/dist/types/packages/sdk/src/profile-external-dataset-package.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/project-actions.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/protocol.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/run-diagnostics.d.ts +3 -3
- package/dist/types/packages/sdk/src/task-design-contracts.d.ts +50 -50
- package/dist/types/packages/sdk/src/task-inventory.d.ts +4 -4
- package/dist/types/packages/sdk/src/task-inventory.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/taskset-authored-contracts.d.ts +190 -189
- package/dist/types/packages/sdk/src/taskset-authored-contracts.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/taskset-authored-files.d.ts +149 -148
- package/dist/types/packages/sdk/src/taskset-authored-files.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/taskset-authored-portable-release.d.ts +15 -6
- package/dist/types/packages/sdk/src/taskset-authored-portable-release.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/taskset-calibration-fixtures.d.ts +64 -63
- package/dist/types/packages/sdk/src/taskset-calibration-fixtures.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/taskset-catalog.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/taskset-draft-api-contracts.d.ts +245 -243
- package/dist/types/packages/sdk/src/taskset-draft-api-contracts.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/taskset-draft-authoring.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/taskset-draft-client.d.ts +744 -738
- package/dist/types/packages/sdk/src/taskset-draft-client.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/taskset-draft-core.d.ts +133 -131
- package/dist/types/packages/sdk/src/taskset-draft-core.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/taskset-draft-dataset-artifacts.d.ts +80 -80
- package/dist/types/packages/sdk/src/taskset-draft-dataset-sources.d.ts +34 -34
- package/dist/types/packages/sdk/src/taskset-draft-document.d.ts +119 -117
- package/dist/types/packages/sdk/src/taskset-draft-document.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/taskset-draft-package-compiler.d.ts +500 -492
- package/dist/types/packages/sdk/src/taskset-draft-package-compiler.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/taskset-draft-source.d.ts +236 -234
- package/dist/types/packages/sdk/src/taskset-draft-source.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/taskset-draft-workspace.d.ts +228 -224
- package/dist/types/packages/sdk/src/taskset-draft-workspace.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/taskset-intake.d.ts +1 -1
- package/dist/types/packages/sdk/src/taskset-package-client.d.ts +4589 -4577
- package/dist/types/packages/sdk/src/taskset-package-client.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/taskset-package-contracts.d.ts +2068 -2054
- package/dist/types/packages/sdk/src/taskset-package-contracts.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/taskset-package-files.d.ts +1 -1
- package/dist/types/packages/sdk/src/taskset-package-grader-projection.d.ts +21 -21
- package/dist/types/packages/sdk/src/taskset-package-learning.d.ts +88 -86
- package/dist/types/packages/sdk/src/taskset-package-learning.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/taskset-package-projection.d.ts +387 -383
- package/dist/types/packages/sdk/src/taskset-package-projection.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/training-acceptance.d.ts +26 -26
- package/dist/types/packages/sdk/src/training-acceptance.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/training-bundle-contracts.d.ts +18 -10
- package/dist/types/packages/sdk/src/training-bundle-contracts.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/training-bundle.d.ts +6 -0
- package/dist/types/packages/sdk/src/training-bundle.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/training-candidate-decisions.d.ts +2 -2
- package/dist/types/packages/sdk/src/training-candidate-decisions.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/training-evaluation-results.d.ts +3 -3
- package/dist/types/packages/sdk/src/training-evaluation-source.d.ts +1 -1
- package/dist/types/packages/sdk/src/training-evaluation-source.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/training-grading-plan.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/training-harness-sources.d.ts +14 -14
- package/dist/types/packages/sdk/src/training-harness-sources.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/training-learning-batch.d.ts +344 -340
- package/dist/types/packages/sdk/src/training-learning-batch.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/training-preparations.d.ts +700 -700
- package/dist/types/packages/sdk/src/training-preparations.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/training-project-client.d.ts +272 -272
- 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 +161 -161
- package/dist/types/packages/sdk/src/training-qualification-preparation.d.ts +368 -0
- package/dist/types/packages/sdk/src/training-qualification-preparation.d.ts.map +1 -0
- package/dist/types/packages/sdk/src/training.d.ts +1574 -1387
- package/dist/types/packages/sdk/src/training.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/work.d.ts +4 -0
- package/dist/types/packages/sdk/src/work.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/workflows.d.ts.map +1 -1
- package/dist/workflows.js +3 -2
- package/dist/workflows.js.map +1 -7
- package/package.json +3 -3
- package/dist/chunk-2DMYAJLS.js.map +0 -7
- package/dist/chunk-2U54EOXJ.js.map +0 -7
- package/dist/chunk-3KXDH6OB.js.map +0 -7
- package/dist/chunk-3LP7DNLQ.js.map +0 -7
- package/dist/chunk-6Q4I33CH.js.map +0 -7
- package/dist/chunk-7AZDNZP6.js.map +0 -7
- package/dist/chunk-7GYXC2AI.js.map +0 -7
- package/dist/chunk-7WUJBFCH.js.map +0 -7
- package/dist/chunk-AG5DDL7T.js.map +0 -7
- package/dist/chunk-AV5JV26B.js.map +0 -7
- package/dist/chunk-B7VYE7YQ.js.map +0 -7
- package/dist/chunk-CEZEVILU.js.map +0 -7
- package/dist/chunk-CK3ONRRC.js.map +0 -7
- package/dist/chunk-COPVBJPL.js.map +0 -7
- package/dist/chunk-D4WTE6ZA.js.map +0 -7
- package/dist/chunk-DBRRQASP.js.map +0 -7
- package/dist/chunk-DG5CRENU.js.map +0 -7
- package/dist/chunk-DUKW3R7O.js.map +0 -7
- package/dist/chunk-ENLQCKY6.js.map +0 -7
- package/dist/chunk-FHCSDXIA.js +0 -1417
- package/dist/chunk-FHCSDXIA.js.map +0 -7
- package/dist/chunk-HKWX624G.js.map +0 -7
- package/dist/chunk-IWT6JUZD.js.map +0 -7
- package/dist/chunk-K3JCP4FO.js.map +0 -7
- package/dist/chunk-KAE7AMYE.js.map +0 -7
- package/dist/chunk-L5PQPAU6.js.map +0 -7
- package/dist/chunk-MEKWESBF.js.map +0 -7
- package/dist/chunk-MLFKQ7E7.js.map +0 -7
- package/dist/chunk-O4GE6VTR.js.map +0 -7
- package/dist/chunk-O5ARJVT5.js.map +0 -7
- package/dist/chunk-OEPYLZXK.js.map +0 -7
- package/dist/chunk-P27DQ65Z.js.map +0 -7
- package/dist/chunk-PGBFI5HC.js +0 -98
- package/dist/chunk-PGBFI5HC.js.map +0 -7
- package/dist/chunk-ROBAX5TN.js.map +0 -7
- package/dist/chunk-T4APHDES.js.map +0 -7
- package/dist/chunk-TCHAWJMB.js.map +0 -7
- package/dist/chunk-VOL3WES2.js.map +0 -7
- package/dist/chunk-VSRTJDVB.js.map +0 -7
- package/dist/chunk-WNH4DOEC.js.map +0 -7
- package/dist/chunk-WVISXXIL.js.map +0 -7
- package/dist/chunk-YCQVVFE5.js.map +0 -7
- package/dist/chunk-YSKUXQ2N.js.map +0 -7
- package/dist/chunk-YVFERCUC.js.map +0 -7
- package/dist/chunk-YVUIHH6Z.js.map +0 -7
|
@@ -23,15 +23,22 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
23
23
|
revision: number;
|
|
24
24
|
teamId: string;
|
|
25
25
|
modelProjectId: string | null;
|
|
26
|
+
name?: string | undefined;
|
|
27
|
+
project?: {
|
|
28
|
+
id: string;
|
|
29
|
+
revision: number;
|
|
30
|
+
contentHash: string;
|
|
31
|
+
targetId: string | null;
|
|
32
|
+
} | undefined;
|
|
26
33
|
operationId: string;
|
|
27
34
|
taskset: {
|
|
28
35
|
id: string;
|
|
29
36
|
contentHash: string;
|
|
30
37
|
revision: number;
|
|
31
38
|
};
|
|
32
|
-
policyKind: "
|
|
39
|
+
policyKind: "fixture" | "hosted_chat" | "hosted_harness";
|
|
33
40
|
manifestHash: string;
|
|
34
|
-
status: "cancelled" | "
|
|
41
|
+
status: "cancelled" | "cancelling" | "completed" | "failed" | "queued" | "running";
|
|
35
42
|
totalCount: number;
|
|
36
43
|
counts: {
|
|
37
44
|
pending: number;
|
|
@@ -51,6 +58,12 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
51
58
|
code: string;
|
|
52
59
|
message: string;
|
|
53
60
|
} | null;
|
|
61
|
+
};
|
|
62
|
+
request: {
|
|
63
|
+
schemaVersion: "openpond.modelTasksetRunRequest.v1";
|
|
64
|
+
operationId: string;
|
|
65
|
+
teamId: string;
|
|
66
|
+
modelProjectId: string | null;
|
|
54
67
|
name?: string | undefined;
|
|
55
68
|
project?: {
|
|
56
69
|
id: string;
|
|
@@ -58,12 +71,6 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
58
71
|
contentHash: string;
|
|
59
72
|
targetId: string | null;
|
|
60
73
|
} | undefined;
|
|
61
|
-
};
|
|
62
|
-
request: {
|
|
63
|
-
schemaVersion: "openpond.modelTasksetRunRequest.v1";
|
|
64
|
-
operationId: string;
|
|
65
|
-
teamId: string;
|
|
66
|
-
modelProjectId: string | null;
|
|
67
74
|
taskset: {
|
|
68
75
|
id: string;
|
|
69
76
|
contentHash: string;
|
|
@@ -72,20 +79,7 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
72
79
|
policy: {
|
|
73
80
|
kind: "hosted_chat";
|
|
74
81
|
modelId: string;
|
|
75
|
-
maxOutputTokens: number;
|
|
76
|
-
temperature: number;
|
|
77
|
-
topP: number;
|
|
78
82
|
localRuntime?: {
|
|
79
|
-
providerId: "claude-code";
|
|
80
|
-
configurationHash: string;
|
|
81
|
-
executableHash: string;
|
|
82
|
-
version: string;
|
|
83
|
-
capabilityHash: string;
|
|
84
|
-
maximumRequestCostUsd: number;
|
|
85
|
-
requestTimeoutMs: number;
|
|
86
|
-
maximumTurns: number;
|
|
87
|
-
holdOpen: boolean;
|
|
88
|
-
} | {
|
|
89
83
|
providerId: "codex";
|
|
90
84
|
configurationHash: string;
|
|
91
85
|
accountHash: string;
|
|
@@ -97,6 +91,16 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
97
91
|
configurationHash: string;
|
|
98
92
|
maximumRequestCostUsd: number;
|
|
99
93
|
requestTimeoutMs: number;
|
|
94
|
+
} | {
|
|
95
|
+
providerId: "claude-code";
|
|
96
|
+
configurationHash: string;
|
|
97
|
+
executableHash: string;
|
|
98
|
+
version: string;
|
|
99
|
+
capabilityHash: string;
|
|
100
|
+
maximumRequestCostUsd: number;
|
|
101
|
+
requestTimeoutMs: number;
|
|
102
|
+
maximumTurns: number;
|
|
103
|
+
holdOpen: boolean;
|
|
100
104
|
} | undefined;
|
|
101
105
|
candidate?: {
|
|
102
106
|
artifactId: string;
|
|
@@ -118,12 +122,47 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
118
122
|
sourcePackageHash: string;
|
|
119
123
|
} | undefined;
|
|
120
124
|
messages?: {
|
|
121
|
-
role: "
|
|
125
|
+
role: "assistant" | "system" | "user";
|
|
122
126
|
content: string;
|
|
123
127
|
}[] | undefined;
|
|
128
|
+
maxOutputTokens: number;
|
|
129
|
+
temperature: number;
|
|
130
|
+
topP: number;
|
|
131
|
+
} | {
|
|
132
|
+
kind: "fixture";
|
|
124
133
|
} | {
|
|
125
134
|
kind: "hosted_harness";
|
|
126
135
|
modelId: string;
|
|
136
|
+
maxOutputTokens?: number | undefined;
|
|
137
|
+
temperature?: number | undefined;
|
|
138
|
+
topP?: number | undefined;
|
|
139
|
+
localRuntime?: {
|
|
140
|
+
providerId: "codex";
|
|
141
|
+
configurationHash: string;
|
|
142
|
+
accountHash: string;
|
|
143
|
+
reasoningEffort: "high" | "low" | "medium" | "xhigh";
|
|
144
|
+
maximumRequestCostUsd: number;
|
|
145
|
+
requestTimeoutMs: number;
|
|
146
|
+
} | undefined;
|
|
147
|
+
candidate?: {
|
|
148
|
+
artifactId: string;
|
|
149
|
+
contentHash: string;
|
|
150
|
+
preparationId: string;
|
|
151
|
+
dispatchId: string;
|
|
152
|
+
baseProfileId: string;
|
|
153
|
+
workerImageDigest: string;
|
|
154
|
+
} | undefined;
|
|
155
|
+
sourceCandidate?: {
|
|
156
|
+
candidateId: string;
|
|
157
|
+
frozenRevision: number;
|
|
158
|
+
frozenHash: string;
|
|
159
|
+
harnessRelease: {
|
|
160
|
+
id: string;
|
|
161
|
+
contentHash: string;
|
|
162
|
+
};
|
|
163
|
+
protectedClosureHash: string;
|
|
164
|
+
} | undefined;
|
|
165
|
+
reasoningEffort?: "high" | "low" | "max" | "medium" | "off" | "xhigh" | undefined;
|
|
127
166
|
profileRepositoryId: string;
|
|
128
167
|
source: {
|
|
129
168
|
profileId: string;
|
|
@@ -171,7 +210,35 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
171
210
|
};
|
|
172
211
|
declaredProfileCatalogHash: string | null;
|
|
173
212
|
protectedProfileClosureHash: string;
|
|
174
|
-
|
|
213
|
+
recordedOrigin?: {
|
|
214
|
+
execution: {
|
|
215
|
+
id: string;
|
|
216
|
+
contentHash: string;
|
|
217
|
+
};
|
|
218
|
+
originalEvidence: {
|
|
219
|
+
id: string;
|
|
220
|
+
contentHash: string;
|
|
221
|
+
};
|
|
222
|
+
members: {
|
|
223
|
+
caseId: string;
|
|
224
|
+
seed: string;
|
|
225
|
+
fixtureId: string;
|
|
226
|
+
sourceId: string;
|
|
227
|
+
snapshotHash: string;
|
|
228
|
+
boundaryRevisionHash: string;
|
|
229
|
+
}[];
|
|
230
|
+
} | undefined;
|
|
231
|
+
gradingSource?: {
|
|
232
|
+
pass: {
|
|
233
|
+
id: string;
|
|
234
|
+
contentHash: string;
|
|
235
|
+
};
|
|
236
|
+
evidence: {
|
|
237
|
+
id: string;
|
|
238
|
+
contentHash: string;
|
|
239
|
+
};
|
|
240
|
+
} | undefined;
|
|
241
|
+
split: "frozen_eval" | "test" | "train" | "validation";
|
|
175
242
|
population: {
|
|
176
243
|
taskId: string;
|
|
177
244
|
seed: string;
|
|
@@ -185,16 +252,16 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
185
252
|
};
|
|
186
253
|
feedbackKey: string;
|
|
187
254
|
configurationHash: string;
|
|
188
|
-
output: "boolean" | "
|
|
255
|
+
output: "boolean" | "category" | "score";
|
|
189
256
|
categories: string[];
|
|
190
257
|
}[];
|
|
191
258
|
fieldMappingsHash: string;
|
|
192
259
|
fieldMappings: {
|
|
193
260
|
graderId: string;
|
|
194
261
|
fields: {
|
|
195
|
-
destination: "
|
|
262
|
+
destination: "evaluatorContext" | "expectedOutput" | "input" | "output";
|
|
196
263
|
path: string[];
|
|
197
|
-
source: "
|
|
264
|
+
source: "evaluatorContext" | "expectedOutput" | "input" | "output";
|
|
198
265
|
sourcePath: string[];
|
|
199
266
|
}[];
|
|
200
267
|
}[];
|
|
@@ -205,34 +272,6 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
205
272
|
requireComplete: boolean;
|
|
206
273
|
};
|
|
207
274
|
contentHash: string;
|
|
208
|
-
recordedOrigin?: {
|
|
209
|
-
execution: {
|
|
210
|
-
id: string;
|
|
211
|
-
contentHash: string;
|
|
212
|
-
};
|
|
213
|
-
originalEvidence: {
|
|
214
|
-
id: string;
|
|
215
|
-
contentHash: string;
|
|
216
|
-
};
|
|
217
|
-
members: {
|
|
218
|
-
caseId: string;
|
|
219
|
-
seed: string;
|
|
220
|
-
fixtureId: string;
|
|
221
|
-
sourceId: string;
|
|
222
|
-
snapshotHash: string;
|
|
223
|
-
boundaryRevisionHash: string;
|
|
224
|
-
}[];
|
|
225
|
-
} | undefined;
|
|
226
|
-
gradingSource?: {
|
|
227
|
-
pass: {
|
|
228
|
-
id: string;
|
|
229
|
-
contentHash: string;
|
|
230
|
-
};
|
|
231
|
-
evidence: {
|
|
232
|
-
id: string;
|
|
233
|
-
contentHash: string;
|
|
234
|
-
};
|
|
235
|
-
} | undefined;
|
|
236
275
|
} | undefined;
|
|
237
276
|
sourceCandidate?: {
|
|
238
277
|
candidateId: string;
|
|
@@ -247,25 +286,6 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
247
286
|
};
|
|
248
287
|
modelConfigurationHash: string;
|
|
249
288
|
packageHash: string;
|
|
250
|
-
candidate?: {
|
|
251
|
-
artifactId: string;
|
|
252
|
-
contentHash: string;
|
|
253
|
-
preparationId: string;
|
|
254
|
-
dispatchId: string;
|
|
255
|
-
baseProfileId: string;
|
|
256
|
-
workerImageDigest: string;
|
|
257
|
-
} | undefined;
|
|
258
|
-
sourceCandidate?: {
|
|
259
|
-
candidateId: string;
|
|
260
|
-
frozenRevision: number;
|
|
261
|
-
frozenHash: string;
|
|
262
|
-
harnessRelease: {
|
|
263
|
-
id: string;
|
|
264
|
-
contentHash: string;
|
|
265
|
-
};
|
|
266
|
-
protectedClosureHash: string;
|
|
267
|
-
} | undefined;
|
|
268
|
-
reasoningEffort?: "high" | "low" | "medium" | "xhigh" | "off" | "max" | undefined;
|
|
269
289
|
externalDatasetBinding?: {
|
|
270
290
|
schemaVersion: "openpond.profileExternalDatasetBinding.v1";
|
|
271
291
|
dataset: {
|
|
@@ -289,7 +309,35 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
289
309
|
};
|
|
290
310
|
declaredProfileCatalogHash: string | null;
|
|
291
311
|
protectedProfileClosureHash: string;
|
|
292
|
-
|
|
312
|
+
recordedOrigin?: {
|
|
313
|
+
execution: {
|
|
314
|
+
id: string;
|
|
315
|
+
contentHash: string;
|
|
316
|
+
};
|
|
317
|
+
originalEvidence: {
|
|
318
|
+
id: string;
|
|
319
|
+
contentHash: string;
|
|
320
|
+
};
|
|
321
|
+
members: {
|
|
322
|
+
caseId: string;
|
|
323
|
+
seed: string;
|
|
324
|
+
fixtureId: string;
|
|
325
|
+
sourceId: string;
|
|
326
|
+
snapshotHash: string;
|
|
327
|
+
boundaryRevisionHash: string;
|
|
328
|
+
}[];
|
|
329
|
+
} | undefined;
|
|
330
|
+
gradingSource?: {
|
|
331
|
+
pass: {
|
|
332
|
+
id: string;
|
|
333
|
+
contentHash: string;
|
|
334
|
+
};
|
|
335
|
+
evidence: {
|
|
336
|
+
id: string;
|
|
337
|
+
contentHash: string;
|
|
338
|
+
};
|
|
339
|
+
} | undefined;
|
|
340
|
+
split: "frozen_eval" | "test" | "train" | "validation";
|
|
293
341
|
population: {
|
|
294
342
|
taskId: string;
|
|
295
343
|
seed: string;
|
|
@@ -303,16 +351,16 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
303
351
|
};
|
|
304
352
|
feedbackKey: string;
|
|
305
353
|
configurationHash: string;
|
|
306
|
-
output: "boolean" | "
|
|
354
|
+
output: "boolean" | "category" | "score";
|
|
307
355
|
categories: string[];
|
|
308
356
|
}[];
|
|
309
357
|
fieldMappingsHash: string;
|
|
310
358
|
fieldMappings: {
|
|
311
359
|
graderId: string;
|
|
312
360
|
fields: {
|
|
313
|
-
destination: "
|
|
361
|
+
destination: "evaluatorContext" | "expectedOutput" | "input" | "output";
|
|
314
362
|
path: string[];
|
|
315
|
-
source: "
|
|
363
|
+
source: "evaluatorContext" | "expectedOutput" | "input" | "output";
|
|
316
364
|
sourcePath: string[];
|
|
317
365
|
}[];
|
|
318
366
|
}[];
|
|
@@ -323,37 +371,7 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
323
371
|
requireComplete: boolean;
|
|
324
372
|
};
|
|
325
373
|
contentHash: string;
|
|
326
|
-
recordedOrigin?: {
|
|
327
|
-
execution: {
|
|
328
|
-
id: string;
|
|
329
|
-
contentHash: string;
|
|
330
|
-
};
|
|
331
|
-
originalEvidence: {
|
|
332
|
-
id: string;
|
|
333
|
-
contentHash: string;
|
|
334
|
-
};
|
|
335
|
-
members: {
|
|
336
|
-
caseId: string;
|
|
337
|
-
seed: string;
|
|
338
|
-
fixtureId: string;
|
|
339
|
-
sourceId: string;
|
|
340
|
-
snapshotHash: string;
|
|
341
|
-
boundaryRevisionHash: string;
|
|
342
|
-
}[];
|
|
343
|
-
} | undefined;
|
|
344
|
-
gradingSource?: {
|
|
345
|
-
pass: {
|
|
346
|
-
id: string;
|
|
347
|
-
contentHash: string;
|
|
348
|
-
};
|
|
349
|
-
evidence: {
|
|
350
|
-
id: string;
|
|
351
|
-
contentHash: string;
|
|
352
|
-
};
|
|
353
|
-
} | undefined;
|
|
354
374
|
} | undefined;
|
|
355
|
-
} | {
|
|
356
|
-
kind: "fixture";
|
|
357
375
|
};
|
|
358
376
|
population: {
|
|
359
377
|
receiptId: string;
|
|
@@ -361,13 +379,6 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
361
379
|
seed: string;
|
|
362
380
|
fixtureId: string | null;
|
|
363
381
|
}[];
|
|
364
|
-
name?: string | undefined;
|
|
365
|
-
project?: {
|
|
366
|
-
id: string;
|
|
367
|
-
revision: number;
|
|
368
|
-
contentHash: string;
|
|
369
|
-
targetId: string | null;
|
|
370
|
-
} | undefined;
|
|
371
382
|
};
|
|
372
383
|
manifest: {
|
|
373
384
|
schemaVersion: "openpond.tasksetRunManifest.v1";
|
|
@@ -395,61 +406,11 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
395
406
|
contentHash: string;
|
|
396
407
|
};
|
|
397
408
|
};
|
|
398
|
-
|
|
399
|
-
|
|
400
|
-
|
|
401
|
-
|
|
402
|
-
|
|
403
|
-
revision: string | null;
|
|
404
|
-
artifactHash: string | null;
|
|
405
|
-
tokenizerRevision: string | null;
|
|
406
|
-
chatTemplateHash: string | null;
|
|
407
|
-
};
|
|
408
|
-
configurationHash: string;
|
|
409
|
-
} | {
|
|
410
|
-
kind: "fixture";
|
|
411
|
-
};
|
|
412
|
-
gradingRole: "evaluation";
|
|
413
|
-
metricPolicy: {
|
|
414
|
-
schemaVersion: "openpond.tasksetMetricPolicy.v1";
|
|
415
|
-
primaryMetric: string;
|
|
416
|
-
aggregation: "custom" | "mean_score" | "pass_rate" | "weighted_mean";
|
|
417
|
-
missingReward: "zero" | "exclude";
|
|
418
|
-
customAggregator: {
|
|
419
|
-
module: string;
|
|
420
|
-
exportName: string;
|
|
421
|
-
contentHash: string;
|
|
422
|
-
timeoutMs: number;
|
|
423
|
-
networkPolicy: "none";
|
|
424
|
-
} | null;
|
|
425
|
-
taskWeights?: Record<string, number> | undefined;
|
|
426
|
-
};
|
|
427
|
-
population: {
|
|
428
|
-
receiptId: string;
|
|
429
|
-
taskId: string;
|
|
430
|
-
seed: string;
|
|
431
|
-
fixtureId: string | null;
|
|
432
|
-
}[];
|
|
433
|
-
runtimeTarget: {
|
|
434
|
-
adapterId: string;
|
|
435
|
-
placement: "local" | "remote" | "colocated" | "provider_native";
|
|
436
|
-
runtimeVersion: string;
|
|
437
|
-
capabilityReceipt: string;
|
|
438
|
-
};
|
|
439
|
-
limits: {
|
|
440
|
-
maxTurns: number;
|
|
441
|
-
timeoutMs: number;
|
|
442
|
-
maxOutputBytes: number;
|
|
443
|
-
maximumSpendUsd: number | null;
|
|
444
|
-
};
|
|
445
|
-
createdAt: string;
|
|
446
|
-
metadata: Record<string, unknown>;
|
|
447
|
-
contentHash: string;
|
|
448
|
-
profileEvaluation?: {
|
|
449
|
-
profileId: string;
|
|
450
|
-
sourceRevision: string;
|
|
451
|
-
harnessRelease: {
|
|
452
|
-
id: string;
|
|
409
|
+
profileEvaluation?: {
|
|
410
|
+
profileId: string;
|
|
411
|
+
sourceRevision: string;
|
|
412
|
+
harnessRelease: {
|
|
413
|
+
id: string;
|
|
453
414
|
contentHash: string;
|
|
454
415
|
};
|
|
455
416
|
catalogHash: string;
|
|
@@ -491,7 +452,35 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
491
452
|
};
|
|
492
453
|
declaredProfileCatalogHash: string | null;
|
|
493
454
|
protectedProfileClosureHash: string;
|
|
494
|
-
|
|
455
|
+
recordedOrigin?: {
|
|
456
|
+
execution: {
|
|
457
|
+
id: string;
|
|
458
|
+
contentHash: string;
|
|
459
|
+
};
|
|
460
|
+
originalEvidence: {
|
|
461
|
+
id: string;
|
|
462
|
+
contentHash: string;
|
|
463
|
+
};
|
|
464
|
+
members: {
|
|
465
|
+
caseId: string;
|
|
466
|
+
seed: string;
|
|
467
|
+
fixtureId: string;
|
|
468
|
+
sourceId: string;
|
|
469
|
+
snapshotHash: string;
|
|
470
|
+
boundaryRevisionHash: string;
|
|
471
|
+
}[];
|
|
472
|
+
} | undefined;
|
|
473
|
+
gradingSource?: {
|
|
474
|
+
pass: {
|
|
475
|
+
id: string;
|
|
476
|
+
contentHash: string;
|
|
477
|
+
};
|
|
478
|
+
evidence: {
|
|
479
|
+
id: string;
|
|
480
|
+
contentHash: string;
|
|
481
|
+
};
|
|
482
|
+
} | undefined;
|
|
483
|
+
split: "frozen_eval" | "test" | "train" | "validation";
|
|
495
484
|
population: {
|
|
496
485
|
taskId: string;
|
|
497
486
|
seed: string;
|
|
@@ -505,16 +494,16 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
505
494
|
};
|
|
506
495
|
feedbackKey: string;
|
|
507
496
|
configurationHash: string;
|
|
508
|
-
output: "boolean" | "
|
|
497
|
+
output: "boolean" | "category" | "score";
|
|
509
498
|
categories: string[];
|
|
510
499
|
}[];
|
|
511
500
|
fieldMappingsHash: string;
|
|
512
501
|
fieldMappings: {
|
|
513
502
|
graderId: string;
|
|
514
503
|
fields: {
|
|
515
|
-
destination: "
|
|
504
|
+
destination: "evaluatorContext" | "expectedOutput" | "input" | "output";
|
|
516
505
|
path: string[];
|
|
517
|
-
source: "
|
|
506
|
+
source: "evaluatorContext" | "expectedOutput" | "input" | "output";
|
|
518
507
|
sourcePath: string[];
|
|
519
508
|
}[];
|
|
520
509
|
}[];
|
|
@@ -525,34 +514,6 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
525
514
|
requireComplete: boolean;
|
|
526
515
|
};
|
|
527
516
|
contentHash: string;
|
|
528
|
-
recordedOrigin?: {
|
|
529
|
-
execution: {
|
|
530
|
-
id: string;
|
|
531
|
-
contentHash: string;
|
|
532
|
-
};
|
|
533
|
-
originalEvidence: {
|
|
534
|
-
id: string;
|
|
535
|
-
contentHash: string;
|
|
536
|
-
};
|
|
537
|
-
members: {
|
|
538
|
-
caseId: string;
|
|
539
|
-
seed: string;
|
|
540
|
-
fixtureId: string;
|
|
541
|
-
sourceId: string;
|
|
542
|
-
snapshotHash: string;
|
|
543
|
-
boundaryRevisionHash: string;
|
|
544
|
-
}[];
|
|
545
|
-
} | undefined;
|
|
546
|
-
gradingSource?: {
|
|
547
|
-
pass: {
|
|
548
|
-
id: string;
|
|
549
|
-
contentHash: string;
|
|
550
|
-
};
|
|
551
|
-
evidence: {
|
|
552
|
-
id: string;
|
|
553
|
-
contentHash: string;
|
|
554
|
-
};
|
|
555
|
-
} | undefined;
|
|
556
517
|
} | undefined;
|
|
557
518
|
sourceCandidate?: {
|
|
558
519
|
candidateId: string;
|
|
@@ -565,6 +526,56 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
565
526
|
protectedClosureHash: string;
|
|
566
527
|
} | undefined;
|
|
567
528
|
} | undefined;
|
|
529
|
+
policy: {
|
|
530
|
+
kind: "model";
|
|
531
|
+
model: {
|
|
532
|
+
provider: string;
|
|
533
|
+
model: string;
|
|
534
|
+
revision: string | null;
|
|
535
|
+
artifactHash: string | null;
|
|
536
|
+
tokenizerRevision: string | null;
|
|
537
|
+
chatTemplateHash: string | null;
|
|
538
|
+
};
|
|
539
|
+
configurationHash: string;
|
|
540
|
+
} | {
|
|
541
|
+
kind: "fixture";
|
|
542
|
+
};
|
|
543
|
+
gradingRole: "evaluation";
|
|
544
|
+
metricPolicy: {
|
|
545
|
+
schemaVersion: "openpond.tasksetMetricPolicy.v1";
|
|
546
|
+
primaryMetric: string;
|
|
547
|
+
aggregation: "custom" | "mean_score" | "pass_rate" | "weighted_mean";
|
|
548
|
+
missingReward: "exclude" | "zero";
|
|
549
|
+
taskWeights?: Record<string, number> | undefined;
|
|
550
|
+
customAggregator: {
|
|
551
|
+
module: string;
|
|
552
|
+
exportName: string;
|
|
553
|
+
contentHash: string;
|
|
554
|
+
timeoutMs: number;
|
|
555
|
+
networkPolicy: "none";
|
|
556
|
+
} | null;
|
|
557
|
+
};
|
|
558
|
+
population: {
|
|
559
|
+
receiptId: string;
|
|
560
|
+
taskId: string;
|
|
561
|
+
seed: string;
|
|
562
|
+
fixtureId: string | null;
|
|
563
|
+
}[];
|
|
564
|
+
runtimeTarget: {
|
|
565
|
+
adapterId: string;
|
|
566
|
+
placement: "colocated" | "local" | "provider_native" | "remote";
|
|
567
|
+
runtimeVersion: string;
|
|
568
|
+
capabilityReceipt: string;
|
|
569
|
+
};
|
|
570
|
+
limits: {
|
|
571
|
+
maxTurns: number;
|
|
572
|
+
timeoutMs: number;
|
|
573
|
+
maxOutputBytes: number;
|
|
574
|
+
maximumSpendUsd: number | null;
|
|
575
|
+
};
|
|
576
|
+
createdAt: string;
|
|
577
|
+
metadata: Record<string, unknown>;
|
|
578
|
+
contentHash: string;
|
|
568
579
|
};
|
|
569
580
|
policySnapshot: {
|
|
570
581
|
modelId: string;
|
|
@@ -582,15 +593,22 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
582
593
|
revision: number;
|
|
583
594
|
teamId: string;
|
|
584
595
|
modelProjectId: string | null;
|
|
596
|
+
name?: string | undefined;
|
|
597
|
+
project?: {
|
|
598
|
+
id: string;
|
|
599
|
+
revision: number;
|
|
600
|
+
contentHash: string;
|
|
601
|
+
targetId: string | null;
|
|
602
|
+
} | undefined;
|
|
585
603
|
operationId: string;
|
|
586
604
|
taskset: {
|
|
587
605
|
id: string;
|
|
588
606
|
contentHash: string;
|
|
589
607
|
revision: number;
|
|
590
608
|
};
|
|
591
|
-
policyKind: "
|
|
609
|
+
policyKind: "fixture" | "hosted_chat" | "hosted_harness";
|
|
592
610
|
manifestHash: string;
|
|
593
|
-
status: "cancelled" | "
|
|
611
|
+
status: "cancelled" | "cancelling" | "completed" | "failed" | "queued" | "running";
|
|
594
612
|
totalCount: number;
|
|
595
613
|
counts: {
|
|
596
614
|
pending: number;
|
|
@@ -610,6 +628,12 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
610
628
|
code: string;
|
|
611
629
|
message: string;
|
|
612
630
|
} | null;
|
|
631
|
+
};
|
|
632
|
+
request: {
|
|
633
|
+
schemaVersion: "openpond.modelTasksetRunRequest.v1";
|
|
634
|
+
operationId: string;
|
|
635
|
+
teamId: string;
|
|
636
|
+
modelProjectId: string | null;
|
|
613
637
|
name?: string | undefined;
|
|
614
638
|
project?: {
|
|
615
639
|
id: string;
|
|
@@ -617,12 +641,6 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
617
641
|
contentHash: string;
|
|
618
642
|
targetId: string | null;
|
|
619
643
|
} | undefined;
|
|
620
|
-
};
|
|
621
|
-
request: {
|
|
622
|
-
schemaVersion: "openpond.modelTasksetRunRequest.v1";
|
|
623
|
-
operationId: string;
|
|
624
|
-
teamId: string;
|
|
625
|
-
modelProjectId: string | null;
|
|
626
644
|
taskset: {
|
|
627
645
|
id: string;
|
|
628
646
|
contentHash: string;
|
|
@@ -631,20 +649,7 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
631
649
|
policy: {
|
|
632
650
|
kind: "hosted_chat";
|
|
633
651
|
modelId: string;
|
|
634
|
-
maxOutputTokens: number;
|
|
635
|
-
temperature: number;
|
|
636
|
-
topP: number;
|
|
637
652
|
localRuntime?: {
|
|
638
|
-
providerId: "claude-code";
|
|
639
|
-
configurationHash: string;
|
|
640
|
-
executableHash: string;
|
|
641
|
-
version: string;
|
|
642
|
-
capabilityHash: string;
|
|
643
|
-
maximumRequestCostUsd: number;
|
|
644
|
-
requestTimeoutMs: number;
|
|
645
|
-
maximumTurns: number;
|
|
646
|
-
holdOpen: boolean;
|
|
647
|
-
} | {
|
|
648
653
|
providerId: "codex";
|
|
649
654
|
configurationHash: string;
|
|
650
655
|
accountHash: string;
|
|
@@ -656,6 +661,16 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
656
661
|
configurationHash: string;
|
|
657
662
|
maximumRequestCostUsd: number;
|
|
658
663
|
requestTimeoutMs: number;
|
|
664
|
+
} | {
|
|
665
|
+
providerId: "claude-code";
|
|
666
|
+
configurationHash: string;
|
|
667
|
+
executableHash: string;
|
|
668
|
+
version: string;
|
|
669
|
+
capabilityHash: string;
|
|
670
|
+
maximumRequestCostUsd: number;
|
|
671
|
+
requestTimeoutMs: number;
|
|
672
|
+
maximumTurns: number;
|
|
673
|
+
holdOpen: boolean;
|
|
659
674
|
} | undefined;
|
|
660
675
|
candidate?: {
|
|
661
676
|
artifactId: string;
|
|
@@ -677,12 +692,47 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
677
692
|
sourcePackageHash: string;
|
|
678
693
|
} | undefined;
|
|
679
694
|
messages?: {
|
|
680
|
-
role: "
|
|
695
|
+
role: "assistant" | "system" | "user";
|
|
681
696
|
content: string;
|
|
682
697
|
}[] | undefined;
|
|
698
|
+
maxOutputTokens: number;
|
|
699
|
+
temperature: number;
|
|
700
|
+
topP: number;
|
|
701
|
+
} | {
|
|
702
|
+
kind: "fixture";
|
|
683
703
|
} | {
|
|
684
704
|
kind: "hosted_harness";
|
|
685
705
|
modelId: string;
|
|
706
|
+
maxOutputTokens?: number | undefined;
|
|
707
|
+
temperature?: number | undefined;
|
|
708
|
+
topP?: number | undefined;
|
|
709
|
+
localRuntime?: {
|
|
710
|
+
providerId: "codex";
|
|
711
|
+
configurationHash: string;
|
|
712
|
+
accountHash: string;
|
|
713
|
+
reasoningEffort: "high" | "low" | "medium" | "xhigh";
|
|
714
|
+
maximumRequestCostUsd: number;
|
|
715
|
+
requestTimeoutMs: number;
|
|
716
|
+
} | undefined;
|
|
717
|
+
candidate?: {
|
|
718
|
+
artifactId: string;
|
|
719
|
+
contentHash: string;
|
|
720
|
+
preparationId: string;
|
|
721
|
+
dispatchId: string;
|
|
722
|
+
baseProfileId: string;
|
|
723
|
+
workerImageDigest: string;
|
|
724
|
+
} | undefined;
|
|
725
|
+
sourceCandidate?: {
|
|
726
|
+
candidateId: string;
|
|
727
|
+
frozenRevision: number;
|
|
728
|
+
frozenHash: string;
|
|
729
|
+
harnessRelease: {
|
|
730
|
+
id: string;
|
|
731
|
+
contentHash: string;
|
|
732
|
+
};
|
|
733
|
+
protectedClosureHash: string;
|
|
734
|
+
} | undefined;
|
|
735
|
+
reasoningEffort?: "high" | "low" | "max" | "medium" | "off" | "xhigh" | undefined;
|
|
686
736
|
profileRepositoryId: string;
|
|
687
737
|
source: {
|
|
688
738
|
profileId: string;
|
|
@@ -730,7 +780,35 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
730
780
|
};
|
|
731
781
|
declaredProfileCatalogHash: string | null;
|
|
732
782
|
protectedProfileClosureHash: string;
|
|
733
|
-
|
|
783
|
+
recordedOrigin?: {
|
|
784
|
+
execution: {
|
|
785
|
+
id: string;
|
|
786
|
+
contentHash: string;
|
|
787
|
+
};
|
|
788
|
+
originalEvidence: {
|
|
789
|
+
id: string;
|
|
790
|
+
contentHash: string;
|
|
791
|
+
};
|
|
792
|
+
members: {
|
|
793
|
+
caseId: string;
|
|
794
|
+
seed: string;
|
|
795
|
+
fixtureId: string;
|
|
796
|
+
sourceId: string;
|
|
797
|
+
snapshotHash: string;
|
|
798
|
+
boundaryRevisionHash: string;
|
|
799
|
+
}[];
|
|
800
|
+
} | undefined;
|
|
801
|
+
gradingSource?: {
|
|
802
|
+
pass: {
|
|
803
|
+
id: string;
|
|
804
|
+
contentHash: string;
|
|
805
|
+
};
|
|
806
|
+
evidence: {
|
|
807
|
+
id: string;
|
|
808
|
+
contentHash: string;
|
|
809
|
+
};
|
|
810
|
+
} | undefined;
|
|
811
|
+
split: "frozen_eval" | "test" | "train" | "validation";
|
|
734
812
|
population: {
|
|
735
813
|
taskId: string;
|
|
736
814
|
seed: string;
|
|
@@ -744,16 +822,16 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
744
822
|
};
|
|
745
823
|
feedbackKey: string;
|
|
746
824
|
configurationHash: string;
|
|
747
|
-
output: "boolean" | "
|
|
825
|
+
output: "boolean" | "category" | "score";
|
|
748
826
|
categories: string[];
|
|
749
827
|
}[];
|
|
750
828
|
fieldMappingsHash: string;
|
|
751
829
|
fieldMappings: {
|
|
752
830
|
graderId: string;
|
|
753
831
|
fields: {
|
|
754
|
-
destination: "
|
|
832
|
+
destination: "evaluatorContext" | "expectedOutput" | "input" | "output";
|
|
755
833
|
path: string[];
|
|
756
|
-
source: "
|
|
834
|
+
source: "evaluatorContext" | "expectedOutput" | "input" | "output";
|
|
757
835
|
sourcePath: string[];
|
|
758
836
|
}[];
|
|
759
837
|
}[];
|
|
@@ -764,34 +842,6 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
764
842
|
requireComplete: boolean;
|
|
765
843
|
};
|
|
766
844
|
contentHash: string;
|
|
767
|
-
recordedOrigin?: {
|
|
768
|
-
execution: {
|
|
769
|
-
id: string;
|
|
770
|
-
contentHash: string;
|
|
771
|
-
};
|
|
772
|
-
originalEvidence: {
|
|
773
|
-
id: string;
|
|
774
|
-
contentHash: string;
|
|
775
|
-
};
|
|
776
|
-
members: {
|
|
777
|
-
caseId: string;
|
|
778
|
-
seed: string;
|
|
779
|
-
fixtureId: string;
|
|
780
|
-
sourceId: string;
|
|
781
|
-
snapshotHash: string;
|
|
782
|
-
boundaryRevisionHash: string;
|
|
783
|
-
}[];
|
|
784
|
-
} | undefined;
|
|
785
|
-
gradingSource?: {
|
|
786
|
-
pass: {
|
|
787
|
-
id: string;
|
|
788
|
-
contentHash: string;
|
|
789
|
-
};
|
|
790
|
-
evidence: {
|
|
791
|
-
id: string;
|
|
792
|
-
contentHash: string;
|
|
793
|
-
};
|
|
794
|
-
} | undefined;
|
|
795
845
|
} | undefined;
|
|
796
846
|
sourceCandidate?: {
|
|
797
847
|
candidateId: string;
|
|
@@ -806,25 +856,6 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
806
856
|
};
|
|
807
857
|
modelConfigurationHash: string;
|
|
808
858
|
packageHash: string;
|
|
809
|
-
candidate?: {
|
|
810
|
-
artifactId: string;
|
|
811
|
-
contentHash: string;
|
|
812
|
-
preparationId: string;
|
|
813
|
-
dispatchId: string;
|
|
814
|
-
baseProfileId: string;
|
|
815
|
-
workerImageDigest: string;
|
|
816
|
-
} | undefined;
|
|
817
|
-
sourceCandidate?: {
|
|
818
|
-
candidateId: string;
|
|
819
|
-
frozenRevision: number;
|
|
820
|
-
frozenHash: string;
|
|
821
|
-
harnessRelease: {
|
|
822
|
-
id: string;
|
|
823
|
-
contentHash: string;
|
|
824
|
-
};
|
|
825
|
-
protectedClosureHash: string;
|
|
826
|
-
} | undefined;
|
|
827
|
-
reasoningEffort?: "high" | "low" | "medium" | "xhigh" | "off" | "max" | undefined;
|
|
828
859
|
externalDatasetBinding?: {
|
|
829
860
|
schemaVersion: "openpond.profileExternalDatasetBinding.v1";
|
|
830
861
|
dataset: {
|
|
@@ -848,7 +879,35 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
848
879
|
};
|
|
849
880
|
declaredProfileCatalogHash: string | null;
|
|
850
881
|
protectedProfileClosureHash: string;
|
|
851
|
-
|
|
882
|
+
recordedOrigin?: {
|
|
883
|
+
execution: {
|
|
884
|
+
id: string;
|
|
885
|
+
contentHash: string;
|
|
886
|
+
};
|
|
887
|
+
originalEvidence: {
|
|
888
|
+
id: string;
|
|
889
|
+
contentHash: string;
|
|
890
|
+
};
|
|
891
|
+
members: {
|
|
892
|
+
caseId: string;
|
|
893
|
+
seed: string;
|
|
894
|
+
fixtureId: string;
|
|
895
|
+
sourceId: string;
|
|
896
|
+
snapshotHash: string;
|
|
897
|
+
boundaryRevisionHash: string;
|
|
898
|
+
}[];
|
|
899
|
+
} | undefined;
|
|
900
|
+
gradingSource?: {
|
|
901
|
+
pass: {
|
|
902
|
+
id: string;
|
|
903
|
+
contentHash: string;
|
|
904
|
+
};
|
|
905
|
+
evidence: {
|
|
906
|
+
id: string;
|
|
907
|
+
contentHash: string;
|
|
908
|
+
};
|
|
909
|
+
} | undefined;
|
|
910
|
+
split: "frozen_eval" | "test" | "train" | "validation";
|
|
852
911
|
population: {
|
|
853
912
|
taskId: string;
|
|
854
913
|
seed: string;
|
|
@@ -862,16 +921,16 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
862
921
|
};
|
|
863
922
|
feedbackKey: string;
|
|
864
923
|
configurationHash: string;
|
|
865
|
-
output: "boolean" | "
|
|
924
|
+
output: "boolean" | "category" | "score";
|
|
866
925
|
categories: string[];
|
|
867
926
|
}[];
|
|
868
927
|
fieldMappingsHash: string;
|
|
869
928
|
fieldMappings: {
|
|
870
929
|
graderId: string;
|
|
871
930
|
fields: {
|
|
872
|
-
destination: "
|
|
931
|
+
destination: "evaluatorContext" | "expectedOutput" | "input" | "output";
|
|
873
932
|
path: string[];
|
|
874
|
-
source: "
|
|
933
|
+
source: "evaluatorContext" | "expectedOutput" | "input" | "output";
|
|
875
934
|
sourcePath: string[];
|
|
876
935
|
}[];
|
|
877
936
|
}[];
|
|
@@ -882,37 +941,7 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
882
941
|
requireComplete: boolean;
|
|
883
942
|
};
|
|
884
943
|
contentHash: string;
|
|
885
|
-
recordedOrigin?: {
|
|
886
|
-
execution: {
|
|
887
|
-
id: string;
|
|
888
|
-
contentHash: string;
|
|
889
|
-
};
|
|
890
|
-
originalEvidence: {
|
|
891
|
-
id: string;
|
|
892
|
-
contentHash: string;
|
|
893
|
-
};
|
|
894
|
-
members: {
|
|
895
|
-
caseId: string;
|
|
896
|
-
seed: string;
|
|
897
|
-
fixtureId: string;
|
|
898
|
-
sourceId: string;
|
|
899
|
-
snapshotHash: string;
|
|
900
|
-
boundaryRevisionHash: string;
|
|
901
|
-
}[];
|
|
902
|
-
} | undefined;
|
|
903
|
-
gradingSource?: {
|
|
904
|
-
pass: {
|
|
905
|
-
id: string;
|
|
906
|
-
contentHash: string;
|
|
907
|
-
};
|
|
908
|
-
evidence: {
|
|
909
|
-
id: string;
|
|
910
|
-
contentHash: string;
|
|
911
|
-
};
|
|
912
|
-
} | undefined;
|
|
913
944
|
} | undefined;
|
|
914
|
-
} | {
|
|
915
|
-
kind: "fixture";
|
|
916
945
|
};
|
|
917
946
|
population: {
|
|
918
947
|
receiptId: string;
|
|
@@ -920,13 +949,6 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
920
949
|
seed: string;
|
|
921
950
|
fixtureId: string | null;
|
|
922
951
|
}[];
|
|
923
|
-
name?: string | undefined;
|
|
924
|
-
project?: {
|
|
925
|
-
id: string;
|
|
926
|
-
revision: number;
|
|
927
|
-
contentHash: string;
|
|
928
|
-
targetId: string | null;
|
|
929
|
-
} | undefined;
|
|
930
952
|
};
|
|
931
953
|
manifest: {
|
|
932
954
|
schemaVersion: "openpond.tasksetRunManifest.v1";
|
|
@@ -954,56 +976,6 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
954
976
|
contentHash: string;
|
|
955
977
|
};
|
|
956
978
|
};
|
|
957
|
-
policy: {
|
|
958
|
-
kind: "model";
|
|
959
|
-
model: {
|
|
960
|
-
provider: string;
|
|
961
|
-
model: string;
|
|
962
|
-
revision: string | null;
|
|
963
|
-
artifactHash: string | null;
|
|
964
|
-
tokenizerRevision: string | null;
|
|
965
|
-
chatTemplateHash: string | null;
|
|
966
|
-
};
|
|
967
|
-
configurationHash: string;
|
|
968
|
-
} | {
|
|
969
|
-
kind: "fixture";
|
|
970
|
-
};
|
|
971
|
-
gradingRole: "evaluation";
|
|
972
|
-
metricPolicy: {
|
|
973
|
-
schemaVersion: "openpond.tasksetMetricPolicy.v1";
|
|
974
|
-
primaryMetric: string;
|
|
975
|
-
aggregation: "custom" | "mean_score" | "pass_rate" | "weighted_mean";
|
|
976
|
-
missingReward: "zero" | "exclude";
|
|
977
|
-
customAggregator: {
|
|
978
|
-
module: string;
|
|
979
|
-
exportName: string;
|
|
980
|
-
contentHash: string;
|
|
981
|
-
timeoutMs: number;
|
|
982
|
-
networkPolicy: "none";
|
|
983
|
-
} | null;
|
|
984
|
-
taskWeights?: Record<string, number> | undefined;
|
|
985
|
-
};
|
|
986
|
-
population: {
|
|
987
|
-
receiptId: string;
|
|
988
|
-
taskId: string;
|
|
989
|
-
seed: string;
|
|
990
|
-
fixtureId: string | null;
|
|
991
|
-
}[];
|
|
992
|
-
runtimeTarget: {
|
|
993
|
-
adapterId: string;
|
|
994
|
-
placement: "local" | "remote" | "colocated" | "provider_native";
|
|
995
|
-
runtimeVersion: string;
|
|
996
|
-
capabilityReceipt: string;
|
|
997
|
-
};
|
|
998
|
-
limits: {
|
|
999
|
-
maxTurns: number;
|
|
1000
|
-
timeoutMs: number;
|
|
1001
|
-
maxOutputBytes: number;
|
|
1002
|
-
maximumSpendUsd: number | null;
|
|
1003
|
-
};
|
|
1004
|
-
createdAt: string;
|
|
1005
|
-
metadata: Record<string, unknown>;
|
|
1006
|
-
contentHash: string;
|
|
1007
979
|
profileEvaluation?: {
|
|
1008
980
|
profileId: string;
|
|
1009
981
|
sourceRevision: string;
|
|
@@ -1050,7 +1022,35 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
1050
1022
|
};
|
|
1051
1023
|
declaredProfileCatalogHash: string | null;
|
|
1052
1024
|
protectedProfileClosureHash: string;
|
|
1053
|
-
|
|
1025
|
+
recordedOrigin?: {
|
|
1026
|
+
execution: {
|
|
1027
|
+
id: string;
|
|
1028
|
+
contentHash: string;
|
|
1029
|
+
};
|
|
1030
|
+
originalEvidence: {
|
|
1031
|
+
id: string;
|
|
1032
|
+
contentHash: string;
|
|
1033
|
+
};
|
|
1034
|
+
members: {
|
|
1035
|
+
caseId: string;
|
|
1036
|
+
seed: string;
|
|
1037
|
+
fixtureId: string;
|
|
1038
|
+
sourceId: string;
|
|
1039
|
+
snapshotHash: string;
|
|
1040
|
+
boundaryRevisionHash: string;
|
|
1041
|
+
}[];
|
|
1042
|
+
} | undefined;
|
|
1043
|
+
gradingSource?: {
|
|
1044
|
+
pass: {
|
|
1045
|
+
id: string;
|
|
1046
|
+
contentHash: string;
|
|
1047
|
+
};
|
|
1048
|
+
evidence: {
|
|
1049
|
+
id: string;
|
|
1050
|
+
contentHash: string;
|
|
1051
|
+
};
|
|
1052
|
+
} | undefined;
|
|
1053
|
+
split: "frozen_eval" | "test" | "train" | "validation";
|
|
1054
1054
|
population: {
|
|
1055
1055
|
taskId: string;
|
|
1056
1056
|
seed: string;
|
|
@@ -1064,16 +1064,16 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
1064
1064
|
};
|
|
1065
1065
|
feedbackKey: string;
|
|
1066
1066
|
configurationHash: string;
|
|
1067
|
-
output: "boolean" | "
|
|
1067
|
+
output: "boolean" | "category" | "score";
|
|
1068
1068
|
categories: string[];
|
|
1069
1069
|
}[];
|
|
1070
1070
|
fieldMappingsHash: string;
|
|
1071
1071
|
fieldMappings: {
|
|
1072
1072
|
graderId: string;
|
|
1073
1073
|
fields: {
|
|
1074
|
-
destination: "
|
|
1074
|
+
destination: "evaluatorContext" | "expectedOutput" | "input" | "output";
|
|
1075
1075
|
path: string[];
|
|
1076
|
-
source: "
|
|
1076
|
+
source: "evaluatorContext" | "expectedOutput" | "input" | "output";
|
|
1077
1077
|
sourcePath: string[];
|
|
1078
1078
|
}[];
|
|
1079
1079
|
}[];
|
|
@@ -1084,34 +1084,6 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
1084
1084
|
requireComplete: boolean;
|
|
1085
1085
|
};
|
|
1086
1086
|
contentHash: string;
|
|
1087
|
-
recordedOrigin?: {
|
|
1088
|
-
execution: {
|
|
1089
|
-
id: string;
|
|
1090
|
-
contentHash: string;
|
|
1091
|
-
};
|
|
1092
|
-
originalEvidence: {
|
|
1093
|
-
id: string;
|
|
1094
|
-
contentHash: string;
|
|
1095
|
-
};
|
|
1096
|
-
members: {
|
|
1097
|
-
caseId: string;
|
|
1098
|
-
seed: string;
|
|
1099
|
-
fixtureId: string;
|
|
1100
|
-
sourceId: string;
|
|
1101
|
-
snapshotHash: string;
|
|
1102
|
-
boundaryRevisionHash: string;
|
|
1103
|
-
}[];
|
|
1104
|
-
} | undefined;
|
|
1105
|
-
gradingSource?: {
|
|
1106
|
-
pass: {
|
|
1107
|
-
id: string;
|
|
1108
|
-
contentHash: string;
|
|
1109
|
-
};
|
|
1110
|
-
evidence: {
|
|
1111
|
-
id: string;
|
|
1112
|
-
contentHash: string;
|
|
1113
|
-
};
|
|
1114
|
-
} | undefined;
|
|
1115
1087
|
} | undefined;
|
|
1116
1088
|
sourceCandidate?: {
|
|
1117
1089
|
candidateId: string;
|
|
@@ -1124,6 +1096,56 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
1124
1096
|
protectedClosureHash: string;
|
|
1125
1097
|
} | undefined;
|
|
1126
1098
|
} | undefined;
|
|
1099
|
+
policy: {
|
|
1100
|
+
kind: "model";
|
|
1101
|
+
model: {
|
|
1102
|
+
provider: string;
|
|
1103
|
+
model: string;
|
|
1104
|
+
revision: string | null;
|
|
1105
|
+
artifactHash: string | null;
|
|
1106
|
+
tokenizerRevision: string | null;
|
|
1107
|
+
chatTemplateHash: string | null;
|
|
1108
|
+
};
|
|
1109
|
+
configurationHash: string;
|
|
1110
|
+
} | {
|
|
1111
|
+
kind: "fixture";
|
|
1112
|
+
};
|
|
1113
|
+
gradingRole: "evaluation";
|
|
1114
|
+
metricPolicy: {
|
|
1115
|
+
schemaVersion: "openpond.tasksetMetricPolicy.v1";
|
|
1116
|
+
primaryMetric: string;
|
|
1117
|
+
aggregation: "custom" | "mean_score" | "pass_rate" | "weighted_mean";
|
|
1118
|
+
missingReward: "exclude" | "zero";
|
|
1119
|
+
taskWeights?: Record<string, number> | undefined;
|
|
1120
|
+
customAggregator: {
|
|
1121
|
+
module: string;
|
|
1122
|
+
exportName: string;
|
|
1123
|
+
contentHash: string;
|
|
1124
|
+
timeoutMs: number;
|
|
1125
|
+
networkPolicy: "none";
|
|
1126
|
+
} | null;
|
|
1127
|
+
};
|
|
1128
|
+
population: {
|
|
1129
|
+
receiptId: string;
|
|
1130
|
+
taskId: string;
|
|
1131
|
+
seed: string;
|
|
1132
|
+
fixtureId: string | null;
|
|
1133
|
+
}[];
|
|
1134
|
+
runtimeTarget: {
|
|
1135
|
+
adapterId: string;
|
|
1136
|
+
placement: "colocated" | "local" | "provider_native" | "remote";
|
|
1137
|
+
runtimeVersion: string;
|
|
1138
|
+
capabilityReceipt: string;
|
|
1139
|
+
};
|
|
1140
|
+
limits: {
|
|
1141
|
+
maxTurns: number;
|
|
1142
|
+
timeoutMs: number;
|
|
1143
|
+
maxOutputBytes: number;
|
|
1144
|
+
maximumSpendUsd: number | null;
|
|
1145
|
+
};
|
|
1146
|
+
createdAt: string;
|
|
1147
|
+
metadata: Record<string, unknown>;
|
|
1148
|
+
contentHash: string;
|
|
1127
1149
|
};
|
|
1128
1150
|
policySnapshot: {
|
|
1129
1151
|
modelId: string;
|
|
@@ -1141,15 +1163,22 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
1141
1163
|
revision: number;
|
|
1142
1164
|
teamId: string;
|
|
1143
1165
|
modelProjectId: string | null;
|
|
1166
|
+
name?: string | undefined;
|
|
1167
|
+
project?: {
|
|
1168
|
+
id: string;
|
|
1169
|
+
revision: number;
|
|
1170
|
+
contentHash: string;
|
|
1171
|
+
targetId: string | null;
|
|
1172
|
+
} | undefined;
|
|
1144
1173
|
operationId: string;
|
|
1145
1174
|
taskset: {
|
|
1146
1175
|
id: string;
|
|
1147
1176
|
contentHash: string;
|
|
1148
1177
|
revision: number;
|
|
1149
1178
|
};
|
|
1150
|
-
policyKind: "
|
|
1179
|
+
policyKind: "fixture" | "hosted_chat" | "hosted_harness";
|
|
1151
1180
|
manifestHash: string;
|
|
1152
|
-
status: "cancelled" | "
|
|
1181
|
+
status: "cancelled" | "cancelling" | "completed" | "failed" | "queued" | "running";
|
|
1153
1182
|
totalCount: number;
|
|
1154
1183
|
counts: {
|
|
1155
1184
|
pending: number;
|
|
@@ -1169,6 +1198,12 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
1169
1198
|
code: string;
|
|
1170
1199
|
message: string;
|
|
1171
1200
|
} | null;
|
|
1201
|
+
};
|
|
1202
|
+
request: {
|
|
1203
|
+
schemaVersion: "openpond.modelTasksetRunRequest.v1";
|
|
1204
|
+
operationId: string;
|
|
1205
|
+
teamId: string;
|
|
1206
|
+
modelProjectId: string | null;
|
|
1172
1207
|
name?: string | undefined;
|
|
1173
1208
|
project?: {
|
|
1174
1209
|
id: string;
|
|
@@ -1176,12 +1211,6 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
1176
1211
|
contentHash: string;
|
|
1177
1212
|
targetId: string | null;
|
|
1178
1213
|
} | undefined;
|
|
1179
|
-
};
|
|
1180
|
-
request: {
|
|
1181
|
-
schemaVersion: "openpond.modelTasksetRunRequest.v1";
|
|
1182
|
-
operationId: string;
|
|
1183
|
-
teamId: string;
|
|
1184
|
-
modelProjectId: string | null;
|
|
1185
1214
|
taskset: {
|
|
1186
1215
|
id: string;
|
|
1187
1216
|
contentHash: string;
|
|
@@ -1190,20 +1219,7 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
1190
1219
|
policy: {
|
|
1191
1220
|
kind: "hosted_chat";
|
|
1192
1221
|
modelId: string;
|
|
1193
|
-
maxOutputTokens: number;
|
|
1194
|
-
temperature: number;
|
|
1195
|
-
topP: number;
|
|
1196
1222
|
localRuntime?: {
|
|
1197
|
-
providerId: "claude-code";
|
|
1198
|
-
configurationHash: string;
|
|
1199
|
-
executableHash: string;
|
|
1200
|
-
version: string;
|
|
1201
|
-
capabilityHash: string;
|
|
1202
|
-
maximumRequestCostUsd: number;
|
|
1203
|
-
requestTimeoutMs: number;
|
|
1204
|
-
maximumTurns: number;
|
|
1205
|
-
holdOpen: boolean;
|
|
1206
|
-
} | {
|
|
1207
1223
|
providerId: "codex";
|
|
1208
1224
|
configurationHash: string;
|
|
1209
1225
|
accountHash: string;
|
|
@@ -1215,6 +1231,16 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
1215
1231
|
configurationHash: string;
|
|
1216
1232
|
maximumRequestCostUsd: number;
|
|
1217
1233
|
requestTimeoutMs: number;
|
|
1234
|
+
} | {
|
|
1235
|
+
providerId: "claude-code";
|
|
1236
|
+
configurationHash: string;
|
|
1237
|
+
executableHash: string;
|
|
1238
|
+
version: string;
|
|
1239
|
+
capabilityHash: string;
|
|
1240
|
+
maximumRequestCostUsd: number;
|
|
1241
|
+
requestTimeoutMs: number;
|
|
1242
|
+
maximumTurns: number;
|
|
1243
|
+
holdOpen: boolean;
|
|
1218
1244
|
} | undefined;
|
|
1219
1245
|
candidate?: {
|
|
1220
1246
|
artifactId: string;
|
|
@@ -1236,12 +1262,47 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
1236
1262
|
sourcePackageHash: string;
|
|
1237
1263
|
} | undefined;
|
|
1238
1264
|
messages?: {
|
|
1239
|
-
role: "
|
|
1265
|
+
role: "assistant" | "system" | "user";
|
|
1240
1266
|
content: string;
|
|
1241
1267
|
}[] | undefined;
|
|
1268
|
+
maxOutputTokens: number;
|
|
1269
|
+
temperature: number;
|
|
1270
|
+
topP: number;
|
|
1271
|
+
} | {
|
|
1272
|
+
kind: "fixture";
|
|
1242
1273
|
} | {
|
|
1243
1274
|
kind: "hosted_harness";
|
|
1244
1275
|
modelId: string;
|
|
1276
|
+
maxOutputTokens?: number | undefined;
|
|
1277
|
+
temperature?: number | undefined;
|
|
1278
|
+
topP?: number | undefined;
|
|
1279
|
+
localRuntime?: {
|
|
1280
|
+
providerId: "codex";
|
|
1281
|
+
configurationHash: string;
|
|
1282
|
+
accountHash: string;
|
|
1283
|
+
reasoningEffort: "high" | "low" | "medium" | "xhigh";
|
|
1284
|
+
maximumRequestCostUsd: number;
|
|
1285
|
+
requestTimeoutMs: number;
|
|
1286
|
+
} | undefined;
|
|
1287
|
+
candidate?: {
|
|
1288
|
+
artifactId: string;
|
|
1289
|
+
contentHash: string;
|
|
1290
|
+
preparationId: string;
|
|
1291
|
+
dispatchId: string;
|
|
1292
|
+
baseProfileId: string;
|
|
1293
|
+
workerImageDigest: string;
|
|
1294
|
+
} | undefined;
|
|
1295
|
+
sourceCandidate?: {
|
|
1296
|
+
candidateId: string;
|
|
1297
|
+
frozenRevision: number;
|
|
1298
|
+
frozenHash: string;
|
|
1299
|
+
harnessRelease: {
|
|
1300
|
+
id: string;
|
|
1301
|
+
contentHash: string;
|
|
1302
|
+
};
|
|
1303
|
+
protectedClosureHash: string;
|
|
1304
|
+
} | undefined;
|
|
1305
|
+
reasoningEffort?: "high" | "low" | "max" | "medium" | "off" | "xhigh" | undefined;
|
|
1245
1306
|
profileRepositoryId: string;
|
|
1246
1307
|
source: {
|
|
1247
1308
|
profileId: string;
|
|
@@ -1289,7 +1350,35 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
1289
1350
|
};
|
|
1290
1351
|
declaredProfileCatalogHash: string | null;
|
|
1291
1352
|
protectedProfileClosureHash: string;
|
|
1292
|
-
|
|
1353
|
+
recordedOrigin?: {
|
|
1354
|
+
execution: {
|
|
1355
|
+
id: string;
|
|
1356
|
+
contentHash: string;
|
|
1357
|
+
};
|
|
1358
|
+
originalEvidence: {
|
|
1359
|
+
id: string;
|
|
1360
|
+
contentHash: string;
|
|
1361
|
+
};
|
|
1362
|
+
members: {
|
|
1363
|
+
caseId: string;
|
|
1364
|
+
seed: string;
|
|
1365
|
+
fixtureId: string;
|
|
1366
|
+
sourceId: string;
|
|
1367
|
+
snapshotHash: string;
|
|
1368
|
+
boundaryRevisionHash: string;
|
|
1369
|
+
}[];
|
|
1370
|
+
} | undefined;
|
|
1371
|
+
gradingSource?: {
|
|
1372
|
+
pass: {
|
|
1373
|
+
id: string;
|
|
1374
|
+
contentHash: string;
|
|
1375
|
+
};
|
|
1376
|
+
evidence: {
|
|
1377
|
+
id: string;
|
|
1378
|
+
contentHash: string;
|
|
1379
|
+
};
|
|
1380
|
+
} | undefined;
|
|
1381
|
+
split: "frozen_eval" | "test" | "train" | "validation";
|
|
1293
1382
|
population: {
|
|
1294
1383
|
taskId: string;
|
|
1295
1384
|
seed: string;
|
|
@@ -1303,16 +1392,16 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
1303
1392
|
};
|
|
1304
1393
|
feedbackKey: string;
|
|
1305
1394
|
configurationHash: string;
|
|
1306
|
-
output: "boolean" | "
|
|
1395
|
+
output: "boolean" | "category" | "score";
|
|
1307
1396
|
categories: string[];
|
|
1308
1397
|
}[];
|
|
1309
1398
|
fieldMappingsHash: string;
|
|
1310
1399
|
fieldMappings: {
|
|
1311
1400
|
graderId: string;
|
|
1312
1401
|
fields: {
|
|
1313
|
-
destination: "
|
|
1402
|
+
destination: "evaluatorContext" | "expectedOutput" | "input" | "output";
|
|
1314
1403
|
path: string[];
|
|
1315
|
-
source: "
|
|
1404
|
+
source: "evaluatorContext" | "expectedOutput" | "input" | "output";
|
|
1316
1405
|
sourcePath: string[];
|
|
1317
1406
|
}[];
|
|
1318
1407
|
}[];
|
|
@@ -1323,34 +1412,6 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
1323
1412
|
requireComplete: boolean;
|
|
1324
1413
|
};
|
|
1325
1414
|
contentHash: string;
|
|
1326
|
-
recordedOrigin?: {
|
|
1327
|
-
execution: {
|
|
1328
|
-
id: string;
|
|
1329
|
-
contentHash: string;
|
|
1330
|
-
};
|
|
1331
|
-
originalEvidence: {
|
|
1332
|
-
id: string;
|
|
1333
|
-
contentHash: string;
|
|
1334
|
-
};
|
|
1335
|
-
members: {
|
|
1336
|
-
caseId: string;
|
|
1337
|
-
seed: string;
|
|
1338
|
-
fixtureId: string;
|
|
1339
|
-
sourceId: string;
|
|
1340
|
-
snapshotHash: string;
|
|
1341
|
-
boundaryRevisionHash: string;
|
|
1342
|
-
}[];
|
|
1343
|
-
} | undefined;
|
|
1344
|
-
gradingSource?: {
|
|
1345
|
-
pass: {
|
|
1346
|
-
id: string;
|
|
1347
|
-
contentHash: string;
|
|
1348
|
-
};
|
|
1349
|
-
evidence: {
|
|
1350
|
-
id: string;
|
|
1351
|
-
contentHash: string;
|
|
1352
|
-
};
|
|
1353
|
-
} | undefined;
|
|
1354
1415
|
} | undefined;
|
|
1355
1416
|
sourceCandidate?: {
|
|
1356
1417
|
candidateId: string;
|
|
@@ -1365,25 +1426,6 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
1365
1426
|
};
|
|
1366
1427
|
modelConfigurationHash: string;
|
|
1367
1428
|
packageHash: string;
|
|
1368
|
-
candidate?: {
|
|
1369
|
-
artifactId: string;
|
|
1370
|
-
contentHash: string;
|
|
1371
|
-
preparationId: string;
|
|
1372
|
-
dispatchId: string;
|
|
1373
|
-
baseProfileId: string;
|
|
1374
|
-
workerImageDigest: string;
|
|
1375
|
-
} | undefined;
|
|
1376
|
-
sourceCandidate?: {
|
|
1377
|
-
candidateId: string;
|
|
1378
|
-
frozenRevision: number;
|
|
1379
|
-
frozenHash: string;
|
|
1380
|
-
harnessRelease: {
|
|
1381
|
-
id: string;
|
|
1382
|
-
contentHash: string;
|
|
1383
|
-
};
|
|
1384
|
-
protectedClosureHash: string;
|
|
1385
|
-
} | undefined;
|
|
1386
|
-
reasoningEffort?: "high" | "low" | "medium" | "xhigh" | "off" | "max" | undefined;
|
|
1387
1429
|
externalDatasetBinding?: {
|
|
1388
1430
|
schemaVersion: "openpond.profileExternalDatasetBinding.v1";
|
|
1389
1431
|
dataset: {
|
|
@@ -1407,7 +1449,35 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
1407
1449
|
};
|
|
1408
1450
|
declaredProfileCatalogHash: string | null;
|
|
1409
1451
|
protectedProfileClosureHash: string;
|
|
1410
|
-
|
|
1452
|
+
recordedOrigin?: {
|
|
1453
|
+
execution: {
|
|
1454
|
+
id: string;
|
|
1455
|
+
contentHash: string;
|
|
1456
|
+
};
|
|
1457
|
+
originalEvidence: {
|
|
1458
|
+
id: string;
|
|
1459
|
+
contentHash: string;
|
|
1460
|
+
};
|
|
1461
|
+
members: {
|
|
1462
|
+
caseId: string;
|
|
1463
|
+
seed: string;
|
|
1464
|
+
fixtureId: string;
|
|
1465
|
+
sourceId: string;
|
|
1466
|
+
snapshotHash: string;
|
|
1467
|
+
boundaryRevisionHash: string;
|
|
1468
|
+
}[];
|
|
1469
|
+
} | undefined;
|
|
1470
|
+
gradingSource?: {
|
|
1471
|
+
pass: {
|
|
1472
|
+
id: string;
|
|
1473
|
+
contentHash: string;
|
|
1474
|
+
};
|
|
1475
|
+
evidence: {
|
|
1476
|
+
id: string;
|
|
1477
|
+
contentHash: string;
|
|
1478
|
+
};
|
|
1479
|
+
} | undefined;
|
|
1480
|
+
split: "frozen_eval" | "test" | "train" | "validation";
|
|
1411
1481
|
population: {
|
|
1412
1482
|
taskId: string;
|
|
1413
1483
|
seed: string;
|
|
@@ -1421,16 +1491,16 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
1421
1491
|
};
|
|
1422
1492
|
feedbackKey: string;
|
|
1423
1493
|
configurationHash: string;
|
|
1424
|
-
output: "boolean" | "
|
|
1494
|
+
output: "boolean" | "category" | "score";
|
|
1425
1495
|
categories: string[];
|
|
1426
1496
|
}[];
|
|
1427
1497
|
fieldMappingsHash: string;
|
|
1428
1498
|
fieldMappings: {
|
|
1429
1499
|
graderId: string;
|
|
1430
1500
|
fields: {
|
|
1431
|
-
destination: "
|
|
1501
|
+
destination: "evaluatorContext" | "expectedOutput" | "input" | "output";
|
|
1432
1502
|
path: string[];
|
|
1433
|
-
source: "
|
|
1503
|
+
source: "evaluatorContext" | "expectedOutput" | "input" | "output";
|
|
1434
1504
|
sourcePath: string[];
|
|
1435
1505
|
}[];
|
|
1436
1506
|
}[];
|
|
@@ -1441,37 +1511,7 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
1441
1511
|
requireComplete: boolean;
|
|
1442
1512
|
};
|
|
1443
1513
|
contentHash: string;
|
|
1444
|
-
recordedOrigin?: {
|
|
1445
|
-
execution: {
|
|
1446
|
-
id: string;
|
|
1447
|
-
contentHash: string;
|
|
1448
|
-
};
|
|
1449
|
-
originalEvidence: {
|
|
1450
|
-
id: string;
|
|
1451
|
-
contentHash: string;
|
|
1452
|
-
};
|
|
1453
|
-
members: {
|
|
1454
|
-
caseId: string;
|
|
1455
|
-
seed: string;
|
|
1456
|
-
fixtureId: string;
|
|
1457
|
-
sourceId: string;
|
|
1458
|
-
snapshotHash: string;
|
|
1459
|
-
boundaryRevisionHash: string;
|
|
1460
|
-
}[];
|
|
1461
|
-
} | undefined;
|
|
1462
|
-
gradingSource?: {
|
|
1463
|
-
pass: {
|
|
1464
|
-
id: string;
|
|
1465
|
-
contentHash: string;
|
|
1466
|
-
};
|
|
1467
|
-
evidence: {
|
|
1468
|
-
id: string;
|
|
1469
|
-
contentHash: string;
|
|
1470
|
-
};
|
|
1471
|
-
} | undefined;
|
|
1472
1514
|
} | undefined;
|
|
1473
|
-
} | {
|
|
1474
|
-
kind: "fixture";
|
|
1475
1515
|
};
|
|
1476
1516
|
population: {
|
|
1477
1517
|
receiptId: string;
|
|
@@ -1479,90 +1519,33 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
1479
1519
|
seed: string;
|
|
1480
1520
|
fixtureId: string | null;
|
|
1481
1521
|
}[];
|
|
1482
|
-
name?: string | undefined;
|
|
1483
|
-
project?: {
|
|
1484
|
-
id: string;
|
|
1485
|
-
revision: number;
|
|
1486
|
-
contentHash: string;
|
|
1487
|
-
targetId: string | null;
|
|
1488
|
-
} | undefined;
|
|
1489
1522
|
};
|
|
1490
1523
|
manifest: {
|
|
1491
1524
|
schemaVersion: "openpond.tasksetRunManifest.v1";
|
|
1492
1525
|
id: string;
|
|
1493
1526
|
tasksetRelease: {
|
|
1494
|
-
id: string;
|
|
1495
|
-
contentHash: string;
|
|
1496
|
-
};
|
|
1497
|
-
packageHash: string;
|
|
1498
|
-
execution: {
|
|
1499
|
-
kind: "taskset";
|
|
1500
|
-
environmentRelease: {
|
|
1501
|
-
id: string;
|
|
1502
|
-
contentHash: string;
|
|
1503
|
-
};
|
|
1504
|
-
verifierSetRelease: {
|
|
1505
|
-
id: string;
|
|
1506
|
-
contentHash: string;
|
|
1507
|
-
};
|
|
1508
|
-
policyHash: string;
|
|
1509
|
-
} | {
|
|
1510
|
-
kind: "harness";
|
|
1511
|
-
harnessRelease: {
|
|
1512
|
-
id: string;
|
|
1513
|
-
contentHash: string;
|
|
1514
|
-
};
|
|
1515
|
-
};
|
|
1516
|
-
policy: {
|
|
1517
|
-
kind: "model";
|
|
1518
|
-
model: {
|
|
1519
|
-
provider: string;
|
|
1520
|
-
model: string;
|
|
1521
|
-
revision: string | null;
|
|
1522
|
-
artifactHash: string | null;
|
|
1523
|
-
tokenizerRevision: string | null;
|
|
1524
|
-
chatTemplateHash: string | null;
|
|
1525
|
-
};
|
|
1526
|
-
configurationHash: string;
|
|
1527
|
-
} | {
|
|
1528
|
-
kind: "fixture";
|
|
1529
|
-
};
|
|
1530
|
-
gradingRole: "evaluation";
|
|
1531
|
-
metricPolicy: {
|
|
1532
|
-
schemaVersion: "openpond.tasksetMetricPolicy.v1";
|
|
1533
|
-
primaryMetric: string;
|
|
1534
|
-
aggregation: "custom" | "mean_score" | "pass_rate" | "weighted_mean";
|
|
1535
|
-
missingReward: "zero" | "exclude";
|
|
1536
|
-
customAggregator: {
|
|
1537
|
-
module: string;
|
|
1538
|
-
exportName: string;
|
|
1539
|
-
contentHash: string;
|
|
1540
|
-
timeoutMs: number;
|
|
1541
|
-
networkPolicy: "none";
|
|
1542
|
-
} | null;
|
|
1543
|
-
taskWeights?: Record<string, number> | undefined;
|
|
1544
|
-
};
|
|
1545
|
-
population: {
|
|
1546
|
-
receiptId: string;
|
|
1547
|
-
taskId: string;
|
|
1548
|
-
seed: string;
|
|
1549
|
-
fixtureId: string | null;
|
|
1550
|
-
}[];
|
|
1551
|
-
runtimeTarget: {
|
|
1552
|
-
adapterId: string;
|
|
1553
|
-
placement: "local" | "remote" | "colocated" | "provider_native";
|
|
1554
|
-
runtimeVersion: string;
|
|
1555
|
-
capabilityReceipt: string;
|
|
1556
|
-
};
|
|
1557
|
-
limits: {
|
|
1558
|
-
maxTurns: number;
|
|
1559
|
-
timeoutMs: number;
|
|
1560
|
-
maxOutputBytes: number;
|
|
1561
|
-
maximumSpendUsd: number | null;
|
|
1527
|
+
id: string;
|
|
1528
|
+
contentHash: string;
|
|
1529
|
+
};
|
|
1530
|
+
packageHash: string;
|
|
1531
|
+
execution: {
|
|
1532
|
+
kind: "taskset";
|
|
1533
|
+
environmentRelease: {
|
|
1534
|
+
id: string;
|
|
1535
|
+
contentHash: string;
|
|
1536
|
+
};
|
|
1537
|
+
verifierSetRelease: {
|
|
1538
|
+
id: string;
|
|
1539
|
+
contentHash: string;
|
|
1540
|
+
};
|
|
1541
|
+
policyHash: string;
|
|
1542
|
+
} | {
|
|
1543
|
+
kind: "harness";
|
|
1544
|
+
harnessRelease: {
|
|
1545
|
+
id: string;
|
|
1546
|
+
contentHash: string;
|
|
1547
|
+
};
|
|
1562
1548
|
};
|
|
1563
|
-
createdAt: string;
|
|
1564
|
-
metadata: Record<string, unknown>;
|
|
1565
|
-
contentHash: string;
|
|
1566
1549
|
profileEvaluation?: {
|
|
1567
1550
|
profileId: string;
|
|
1568
1551
|
sourceRevision: string;
|
|
@@ -1609,7 +1592,35 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
1609
1592
|
};
|
|
1610
1593
|
declaredProfileCatalogHash: string | null;
|
|
1611
1594
|
protectedProfileClosureHash: string;
|
|
1612
|
-
|
|
1595
|
+
recordedOrigin?: {
|
|
1596
|
+
execution: {
|
|
1597
|
+
id: string;
|
|
1598
|
+
contentHash: string;
|
|
1599
|
+
};
|
|
1600
|
+
originalEvidence: {
|
|
1601
|
+
id: string;
|
|
1602
|
+
contentHash: string;
|
|
1603
|
+
};
|
|
1604
|
+
members: {
|
|
1605
|
+
caseId: string;
|
|
1606
|
+
seed: string;
|
|
1607
|
+
fixtureId: string;
|
|
1608
|
+
sourceId: string;
|
|
1609
|
+
snapshotHash: string;
|
|
1610
|
+
boundaryRevisionHash: string;
|
|
1611
|
+
}[];
|
|
1612
|
+
} | undefined;
|
|
1613
|
+
gradingSource?: {
|
|
1614
|
+
pass: {
|
|
1615
|
+
id: string;
|
|
1616
|
+
contentHash: string;
|
|
1617
|
+
};
|
|
1618
|
+
evidence: {
|
|
1619
|
+
id: string;
|
|
1620
|
+
contentHash: string;
|
|
1621
|
+
};
|
|
1622
|
+
} | undefined;
|
|
1623
|
+
split: "frozen_eval" | "test" | "train" | "validation";
|
|
1613
1624
|
population: {
|
|
1614
1625
|
taskId: string;
|
|
1615
1626
|
seed: string;
|
|
@@ -1623,16 +1634,16 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
1623
1634
|
};
|
|
1624
1635
|
feedbackKey: string;
|
|
1625
1636
|
configurationHash: string;
|
|
1626
|
-
output: "boolean" | "
|
|
1637
|
+
output: "boolean" | "category" | "score";
|
|
1627
1638
|
categories: string[];
|
|
1628
1639
|
}[];
|
|
1629
1640
|
fieldMappingsHash: string;
|
|
1630
1641
|
fieldMappings: {
|
|
1631
1642
|
graderId: string;
|
|
1632
1643
|
fields: {
|
|
1633
|
-
destination: "
|
|
1644
|
+
destination: "evaluatorContext" | "expectedOutput" | "input" | "output";
|
|
1634
1645
|
path: string[];
|
|
1635
|
-
source: "
|
|
1646
|
+
source: "evaluatorContext" | "expectedOutput" | "input" | "output";
|
|
1636
1647
|
sourcePath: string[];
|
|
1637
1648
|
}[];
|
|
1638
1649
|
}[];
|
|
@@ -1643,34 +1654,6 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
1643
1654
|
requireComplete: boolean;
|
|
1644
1655
|
};
|
|
1645
1656
|
contentHash: string;
|
|
1646
|
-
recordedOrigin?: {
|
|
1647
|
-
execution: {
|
|
1648
|
-
id: string;
|
|
1649
|
-
contentHash: string;
|
|
1650
|
-
};
|
|
1651
|
-
originalEvidence: {
|
|
1652
|
-
id: string;
|
|
1653
|
-
contentHash: string;
|
|
1654
|
-
};
|
|
1655
|
-
members: {
|
|
1656
|
-
caseId: string;
|
|
1657
|
-
seed: string;
|
|
1658
|
-
fixtureId: string;
|
|
1659
|
-
sourceId: string;
|
|
1660
|
-
snapshotHash: string;
|
|
1661
|
-
boundaryRevisionHash: string;
|
|
1662
|
-
}[];
|
|
1663
|
-
} | undefined;
|
|
1664
|
-
gradingSource?: {
|
|
1665
|
-
pass: {
|
|
1666
|
-
id: string;
|
|
1667
|
-
contentHash: string;
|
|
1668
|
-
};
|
|
1669
|
-
evidence: {
|
|
1670
|
-
id: string;
|
|
1671
|
-
contentHash: string;
|
|
1672
|
-
};
|
|
1673
|
-
} | undefined;
|
|
1674
1657
|
} | undefined;
|
|
1675
1658
|
sourceCandidate?: {
|
|
1676
1659
|
candidateId: string;
|
|
@@ -1683,6 +1666,56 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
1683
1666
|
protectedClosureHash: string;
|
|
1684
1667
|
} | undefined;
|
|
1685
1668
|
} | undefined;
|
|
1669
|
+
policy: {
|
|
1670
|
+
kind: "model";
|
|
1671
|
+
model: {
|
|
1672
|
+
provider: string;
|
|
1673
|
+
model: string;
|
|
1674
|
+
revision: string | null;
|
|
1675
|
+
artifactHash: string | null;
|
|
1676
|
+
tokenizerRevision: string | null;
|
|
1677
|
+
chatTemplateHash: string | null;
|
|
1678
|
+
};
|
|
1679
|
+
configurationHash: string;
|
|
1680
|
+
} | {
|
|
1681
|
+
kind: "fixture";
|
|
1682
|
+
};
|
|
1683
|
+
gradingRole: "evaluation";
|
|
1684
|
+
metricPolicy: {
|
|
1685
|
+
schemaVersion: "openpond.tasksetMetricPolicy.v1";
|
|
1686
|
+
primaryMetric: string;
|
|
1687
|
+
aggregation: "custom" | "mean_score" | "pass_rate" | "weighted_mean";
|
|
1688
|
+
missingReward: "exclude" | "zero";
|
|
1689
|
+
taskWeights?: Record<string, number> | undefined;
|
|
1690
|
+
customAggregator: {
|
|
1691
|
+
module: string;
|
|
1692
|
+
exportName: string;
|
|
1693
|
+
contentHash: string;
|
|
1694
|
+
timeoutMs: number;
|
|
1695
|
+
networkPolicy: "none";
|
|
1696
|
+
} | null;
|
|
1697
|
+
};
|
|
1698
|
+
population: {
|
|
1699
|
+
receiptId: string;
|
|
1700
|
+
taskId: string;
|
|
1701
|
+
seed: string;
|
|
1702
|
+
fixtureId: string | null;
|
|
1703
|
+
}[];
|
|
1704
|
+
runtimeTarget: {
|
|
1705
|
+
adapterId: string;
|
|
1706
|
+
placement: "colocated" | "local" | "provider_native" | "remote";
|
|
1707
|
+
runtimeVersion: string;
|
|
1708
|
+
capabilityReceipt: string;
|
|
1709
|
+
};
|
|
1710
|
+
limits: {
|
|
1711
|
+
maxTurns: number;
|
|
1712
|
+
timeoutMs: number;
|
|
1713
|
+
maxOutputBytes: number;
|
|
1714
|
+
maximumSpendUsd: number | null;
|
|
1715
|
+
};
|
|
1716
|
+
createdAt: string;
|
|
1717
|
+
metadata: Record<string, unknown>;
|
|
1718
|
+
contentHash: string;
|
|
1686
1719
|
};
|
|
1687
1720
|
policySnapshot: {
|
|
1688
1721
|
modelId: string;
|
|
@@ -1702,15 +1735,22 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
1702
1735
|
revision: number;
|
|
1703
1736
|
teamId: string;
|
|
1704
1737
|
modelProjectId: string | null;
|
|
1738
|
+
name?: string | undefined;
|
|
1739
|
+
project?: {
|
|
1740
|
+
id: string;
|
|
1741
|
+
revision: number;
|
|
1742
|
+
contentHash: string;
|
|
1743
|
+
targetId: string | null;
|
|
1744
|
+
} | undefined;
|
|
1705
1745
|
operationId: string;
|
|
1706
1746
|
taskset: {
|
|
1707
1747
|
id: string;
|
|
1708
1748
|
contentHash: string;
|
|
1709
1749
|
revision: number;
|
|
1710
1750
|
};
|
|
1711
|
-
policyKind: "
|
|
1751
|
+
policyKind: "fixture" | "hosted_chat" | "hosted_harness";
|
|
1712
1752
|
manifestHash: string;
|
|
1713
|
-
status: "cancelled" | "
|
|
1753
|
+
status: "cancelled" | "cancelling" | "completed" | "failed" | "queued" | "running";
|
|
1714
1754
|
totalCount: number;
|
|
1715
1755
|
counts: {
|
|
1716
1756
|
pending: number;
|
|
@@ -1730,6 +1770,12 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
1730
1770
|
code: string;
|
|
1731
1771
|
message: string;
|
|
1732
1772
|
} | null;
|
|
1773
|
+
};
|
|
1774
|
+
request: {
|
|
1775
|
+
schemaVersion: "openpond.modelTasksetRunRequest.v1";
|
|
1776
|
+
operationId: string;
|
|
1777
|
+
teamId: string;
|
|
1778
|
+
modelProjectId: string | null;
|
|
1733
1779
|
name?: string | undefined;
|
|
1734
1780
|
project?: {
|
|
1735
1781
|
id: string;
|
|
@@ -1737,12 +1783,6 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
1737
1783
|
contentHash: string;
|
|
1738
1784
|
targetId: string | null;
|
|
1739
1785
|
} | undefined;
|
|
1740
|
-
};
|
|
1741
|
-
request: {
|
|
1742
|
-
schemaVersion: "openpond.modelTasksetRunRequest.v1";
|
|
1743
|
-
operationId: string;
|
|
1744
|
-
teamId: string;
|
|
1745
|
-
modelProjectId: string | null;
|
|
1746
1786
|
taskset: {
|
|
1747
1787
|
id: string;
|
|
1748
1788
|
contentHash: string;
|
|
@@ -1751,20 +1791,7 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
1751
1791
|
policy: {
|
|
1752
1792
|
kind: "hosted_chat";
|
|
1753
1793
|
modelId: string;
|
|
1754
|
-
maxOutputTokens: number;
|
|
1755
|
-
temperature: number;
|
|
1756
|
-
topP: number;
|
|
1757
1794
|
localRuntime?: {
|
|
1758
|
-
providerId: "claude-code";
|
|
1759
|
-
configurationHash: string;
|
|
1760
|
-
executableHash: string;
|
|
1761
|
-
version: string;
|
|
1762
|
-
capabilityHash: string;
|
|
1763
|
-
maximumRequestCostUsd: number;
|
|
1764
|
-
requestTimeoutMs: number;
|
|
1765
|
-
maximumTurns: number;
|
|
1766
|
-
holdOpen: boolean;
|
|
1767
|
-
} | {
|
|
1768
1795
|
providerId: "codex";
|
|
1769
1796
|
configurationHash: string;
|
|
1770
1797
|
accountHash: string;
|
|
@@ -1776,6 +1803,16 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
1776
1803
|
configurationHash: string;
|
|
1777
1804
|
maximumRequestCostUsd: number;
|
|
1778
1805
|
requestTimeoutMs: number;
|
|
1806
|
+
} | {
|
|
1807
|
+
providerId: "claude-code";
|
|
1808
|
+
configurationHash: string;
|
|
1809
|
+
executableHash: string;
|
|
1810
|
+
version: string;
|
|
1811
|
+
capabilityHash: string;
|
|
1812
|
+
maximumRequestCostUsd: number;
|
|
1813
|
+
requestTimeoutMs: number;
|
|
1814
|
+
maximumTurns: number;
|
|
1815
|
+
holdOpen: boolean;
|
|
1779
1816
|
} | undefined;
|
|
1780
1817
|
candidate?: {
|
|
1781
1818
|
artifactId: string;
|
|
@@ -1797,12 +1834,47 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
1797
1834
|
sourcePackageHash: string;
|
|
1798
1835
|
} | undefined;
|
|
1799
1836
|
messages?: {
|
|
1800
|
-
role: "
|
|
1837
|
+
role: "assistant" | "system" | "user";
|
|
1801
1838
|
content: string;
|
|
1802
1839
|
}[] | undefined;
|
|
1840
|
+
maxOutputTokens: number;
|
|
1841
|
+
temperature: number;
|
|
1842
|
+
topP: number;
|
|
1843
|
+
} | {
|
|
1844
|
+
kind: "fixture";
|
|
1803
1845
|
} | {
|
|
1804
1846
|
kind: "hosted_harness";
|
|
1805
1847
|
modelId: string;
|
|
1848
|
+
maxOutputTokens?: number | undefined;
|
|
1849
|
+
temperature?: number | undefined;
|
|
1850
|
+
topP?: number | undefined;
|
|
1851
|
+
localRuntime?: {
|
|
1852
|
+
providerId: "codex";
|
|
1853
|
+
configurationHash: string;
|
|
1854
|
+
accountHash: string;
|
|
1855
|
+
reasoningEffort: "high" | "low" | "medium" | "xhigh";
|
|
1856
|
+
maximumRequestCostUsd: number;
|
|
1857
|
+
requestTimeoutMs: number;
|
|
1858
|
+
} | undefined;
|
|
1859
|
+
candidate?: {
|
|
1860
|
+
artifactId: string;
|
|
1861
|
+
contentHash: string;
|
|
1862
|
+
preparationId: string;
|
|
1863
|
+
dispatchId: string;
|
|
1864
|
+
baseProfileId: string;
|
|
1865
|
+
workerImageDigest: string;
|
|
1866
|
+
} | undefined;
|
|
1867
|
+
sourceCandidate?: {
|
|
1868
|
+
candidateId: string;
|
|
1869
|
+
frozenRevision: number;
|
|
1870
|
+
frozenHash: string;
|
|
1871
|
+
harnessRelease: {
|
|
1872
|
+
id: string;
|
|
1873
|
+
contentHash: string;
|
|
1874
|
+
};
|
|
1875
|
+
protectedClosureHash: string;
|
|
1876
|
+
} | undefined;
|
|
1877
|
+
reasoningEffort?: "high" | "low" | "max" | "medium" | "off" | "xhigh" | undefined;
|
|
1806
1878
|
profileRepositoryId: string;
|
|
1807
1879
|
source: {
|
|
1808
1880
|
profileId: string;
|
|
@@ -1850,40 +1922,6 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
1850
1922
|
};
|
|
1851
1923
|
declaredProfileCatalogHash: string | null;
|
|
1852
1924
|
protectedProfileClosureHash: string;
|
|
1853
|
-
split: "train" | "validation" | "test" | "frozen_eval";
|
|
1854
|
-
population: {
|
|
1855
|
-
taskId: string;
|
|
1856
|
-
seed: string;
|
|
1857
|
-
fixtureId: string | null;
|
|
1858
|
-
}[];
|
|
1859
|
-
evaluators: {
|
|
1860
|
-
release: {
|
|
1861
|
-
id: string;
|
|
1862
|
-
contentHash: string;
|
|
1863
|
-
revision: string | number;
|
|
1864
|
-
};
|
|
1865
|
-
feedbackKey: string;
|
|
1866
|
-
configurationHash: string;
|
|
1867
|
-
output: "boolean" | "score" | "category";
|
|
1868
|
-
categories: string[];
|
|
1869
|
-
}[];
|
|
1870
|
-
fieldMappingsHash: string;
|
|
1871
|
-
fieldMappings: {
|
|
1872
|
-
graderId: string;
|
|
1873
|
-
fields: {
|
|
1874
|
-
destination: "input" | "output" | "expectedOutput" | "evaluatorContext";
|
|
1875
|
-
path: string[];
|
|
1876
|
-
source: "input" | "output" | "expectedOutput" | "evaluatorContext";
|
|
1877
|
-
sourcePath: string[];
|
|
1878
|
-
}[];
|
|
1879
|
-
}[];
|
|
1880
|
-
environmentHash: string;
|
|
1881
|
-
privateDatasetClosureHash: string;
|
|
1882
|
-
criterion: {
|
|
1883
|
-
minimumPassRate: number;
|
|
1884
|
-
requireComplete: boolean;
|
|
1885
|
-
};
|
|
1886
|
-
contentHash: string;
|
|
1887
1925
|
recordedOrigin?: {
|
|
1888
1926
|
execution: {
|
|
1889
1927
|
id: string;
|
|
@@ -1912,6 +1950,40 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
1912
1950
|
contentHash: string;
|
|
1913
1951
|
};
|
|
1914
1952
|
} | undefined;
|
|
1953
|
+
split: "frozen_eval" | "test" | "train" | "validation";
|
|
1954
|
+
population: {
|
|
1955
|
+
taskId: string;
|
|
1956
|
+
seed: string;
|
|
1957
|
+
fixtureId: string | null;
|
|
1958
|
+
}[];
|
|
1959
|
+
evaluators: {
|
|
1960
|
+
release: {
|
|
1961
|
+
id: string;
|
|
1962
|
+
contentHash: string;
|
|
1963
|
+
revision: string | number;
|
|
1964
|
+
};
|
|
1965
|
+
feedbackKey: string;
|
|
1966
|
+
configurationHash: string;
|
|
1967
|
+
output: "boolean" | "category" | "score";
|
|
1968
|
+
categories: string[];
|
|
1969
|
+
}[];
|
|
1970
|
+
fieldMappingsHash: string;
|
|
1971
|
+
fieldMappings: {
|
|
1972
|
+
graderId: string;
|
|
1973
|
+
fields: {
|
|
1974
|
+
destination: "evaluatorContext" | "expectedOutput" | "input" | "output";
|
|
1975
|
+
path: string[];
|
|
1976
|
+
source: "evaluatorContext" | "expectedOutput" | "input" | "output";
|
|
1977
|
+
sourcePath: string[];
|
|
1978
|
+
}[];
|
|
1979
|
+
}[];
|
|
1980
|
+
environmentHash: string;
|
|
1981
|
+
privateDatasetClosureHash: string;
|
|
1982
|
+
criterion: {
|
|
1983
|
+
minimumPassRate: number;
|
|
1984
|
+
requireComplete: boolean;
|
|
1985
|
+
};
|
|
1986
|
+
contentHash: string;
|
|
1915
1987
|
} | undefined;
|
|
1916
1988
|
sourceCandidate?: {
|
|
1917
1989
|
candidateId: string;
|
|
@@ -1926,25 +1998,6 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
1926
1998
|
};
|
|
1927
1999
|
modelConfigurationHash: string;
|
|
1928
2000
|
packageHash: string;
|
|
1929
|
-
candidate?: {
|
|
1930
|
-
artifactId: string;
|
|
1931
|
-
contentHash: string;
|
|
1932
|
-
preparationId: string;
|
|
1933
|
-
dispatchId: string;
|
|
1934
|
-
baseProfileId: string;
|
|
1935
|
-
workerImageDigest: string;
|
|
1936
|
-
} | undefined;
|
|
1937
|
-
sourceCandidate?: {
|
|
1938
|
-
candidateId: string;
|
|
1939
|
-
frozenRevision: number;
|
|
1940
|
-
frozenHash: string;
|
|
1941
|
-
harnessRelease: {
|
|
1942
|
-
id: string;
|
|
1943
|
-
contentHash: string;
|
|
1944
|
-
};
|
|
1945
|
-
protectedClosureHash: string;
|
|
1946
|
-
} | undefined;
|
|
1947
|
-
reasoningEffort?: "high" | "low" | "medium" | "xhigh" | "off" | "max" | undefined;
|
|
1948
2001
|
externalDatasetBinding?: {
|
|
1949
2002
|
schemaVersion: "openpond.profileExternalDatasetBinding.v1";
|
|
1950
2003
|
dataset: {
|
|
@@ -1968,7 +2021,35 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
1968
2021
|
};
|
|
1969
2022
|
declaredProfileCatalogHash: string | null;
|
|
1970
2023
|
protectedProfileClosureHash: string;
|
|
1971
|
-
|
|
2024
|
+
recordedOrigin?: {
|
|
2025
|
+
execution: {
|
|
2026
|
+
id: string;
|
|
2027
|
+
contentHash: string;
|
|
2028
|
+
};
|
|
2029
|
+
originalEvidence: {
|
|
2030
|
+
id: string;
|
|
2031
|
+
contentHash: string;
|
|
2032
|
+
};
|
|
2033
|
+
members: {
|
|
2034
|
+
caseId: string;
|
|
2035
|
+
seed: string;
|
|
2036
|
+
fixtureId: string;
|
|
2037
|
+
sourceId: string;
|
|
2038
|
+
snapshotHash: string;
|
|
2039
|
+
boundaryRevisionHash: string;
|
|
2040
|
+
}[];
|
|
2041
|
+
} | undefined;
|
|
2042
|
+
gradingSource?: {
|
|
2043
|
+
pass: {
|
|
2044
|
+
id: string;
|
|
2045
|
+
contentHash: string;
|
|
2046
|
+
};
|
|
2047
|
+
evidence: {
|
|
2048
|
+
id: string;
|
|
2049
|
+
contentHash: string;
|
|
2050
|
+
};
|
|
2051
|
+
} | undefined;
|
|
2052
|
+
split: "frozen_eval" | "test" | "train" | "validation";
|
|
1972
2053
|
population: {
|
|
1973
2054
|
taskId: string;
|
|
1974
2055
|
seed: string;
|
|
@@ -1982,16 +2063,16 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
1982
2063
|
};
|
|
1983
2064
|
feedbackKey: string;
|
|
1984
2065
|
configurationHash: string;
|
|
1985
|
-
output: "boolean" | "
|
|
2066
|
+
output: "boolean" | "category" | "score";
|
|
1986
2067
|
categories: string[];
|
|
1987
2068
|
}[];
|
|
1988
2069
|
fieldMappingsHash: string;
|
|
1989
2070
|
fieldMappings: {
|
|
1990
2071
|
graderId: string;
|
|
1991
2072
|
fields: {
|
|
1992
|
-
destination: "
|
|
2073
|
+
destination: "evaluatorContext" | "expectedOutput" | "input" | "output";
|
|
1993
2074
|
path: string[];
|
|
1994
|
-
source: "
|
|
2075
|
+
source: "evaluatorContext" | "expectedOutput" | "input" | "output";
|
|
1995
2076
|
sourcePath: string[];
|
|
1996
2077
|
}[];
|
|
1997
2078
|
}[];
|
|
@@ -2002,37 +2083,7 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
2002
2083
|
requireComplete: boolean;
|
|
2003
2084
|
};
|
|
2004
2085
|
contentHash: string;
|
|
2005
|
-
recordedOrigin?: {
|
|
2006
|
-
execution: {
|
|
2007
|
-
id: string;
|
|
2008
|
-
contentHash: string;
|
|
2009
|
-
};
|
|
2010
|
-
originalEvidence: {
|
|
2011
|
-
id: string;
|
|
2012
|
-
contentHash: string;
|
|
2013
|
-
};
|
|
2014
|
-
members: {
|
|
2015
|
-
caseId: string;
|
|
2016
|
-
seed: string;
|
|
2017
|
-
fixtureId: string;
|
|
2018
|
-
sourceId: string;
|
|
2019
|
-
snapshotHash: string;
|
|
2020
|
-
boundaryRevisionHash: string;
|
|
2021
|
-
}[];
|
|
2022
|
-
} | undefined;
|
|
2023
|
-
gradingSource?: {
|
|
2024
|
-
pass: {
|
|
2025
|
-
id: string;
|
|
2026
|
-
contentHash: string;
|
|
2027
|
-
};
|
|
2028
|
-
evidence: {
|
|
2029
|
-
id: string;
|
|
2030
|
-
contentHash: string;
|
|
2031
|
-
};
|
|
2032
|
-
} | undefined;
|
|
2033
2086
|
} | undefined;
|
|
2034
|
-
} | {
|
|
2035
|
-
kind: "fixture";
|
|
2036
2087
|
};
|
|
2037
2088
|
population: {
|
|
2038
2089
|
receiptId: string;
|
|
@@ -2040,13 +2091,6 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
2040
2091
|
seed: string;
|
|
2041
2092
|
fixtureId: string | null;
|
|
2042
2093
|
}[];
|
|
2043
|
-
name?: string | undefined;
|
|
2044
|
-
project?: {
|
|
2045
|
-
id: string;
|
|
2046
|
-
revision: number;
|
|
2047
|
-
contentHash: string;
|
|
2048
|
-
targetId: string | null;
|
|
2049
|
-
} | undefined;
|
|
2050
2094
|
};
|
|
2051
2095
|
manifest: {
|
|
2052
2096
|
schemaVersion: "openpond.tasksetRunManifest.v1";
|
|
@@ -2074,56 +2118,6 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
2074
2118
|
contentHash: string;
|
|
2075
2119
|
};
|
|
2076
2120
|
};
|
|
2077
|
-
policy: {
|
|
2078
|
-
kind: "model";
|
|
2079
|
-
model: {
|
|
2080
|
-
provider: string;
|
|
2081
|
-
model: string;
|
|
2082
|
-
revision: string | null;
|
|
2083
|
-
artifactHash: string | null;
|
|
2084
|
-
tokenizerRevision: string | null;
|
|
2085
|
-
chatTemplateHash: string | null;
|
|
2086
|
-
};
|
|
2087
|
-
configurationHash: string;
|
|
2088
|
-
} | {
|
|
2089
|
-
kind: "fixture";
|
|
2090
|
-
};
|
|
2091
|
-
gradingRole: "evaluation";
|
|
2092
|
-
metricPolicy: {
|
|
2093
|
-
schemaVersion: "openpond.tasksetMetricPolicy.v1";
|
|
2094
|
-
primaryMetric: string;
|
|
2095
|
-
aggregation: "custom" | "mean_score" | "pass_rate" | "weighted_mean";
|
|
2096
|
-
missingReward: "zero" | "exclude";
|
|
2097
|
-
customAggregator: {
|
|
2098
|
-
module: string;
|
|
2099
|
-
exportName: string;
|
|
2100
|
-
contentHash: string;
|
|
2101
|
-
timeoutMs: number;
|
|
2102
|
-
networkPolicy: "none";
|
|
2103
|
-
} | null;
|
|
2104
|
-
taskWeights?: Record<string, number> | undefined;
|
|
2105
|
-
};
|
|
2106
|
-
population: {
|
|
2107
|
-
receiptId: string;
|
|
2108
|
-
taskId: string;
|
|
2109
|
-
seed: string;
|
|
2110
|
-
fixtureId: string | null;
|
|
2111
|
-
}[];
|
|
2112
|
-
runtimeTarget: {
|
|
2113
|
-
adapterId: string;
|
|
2114
|
-
placement: "local" | "remote" | "colocated" | "provider_native";
|
|
2115
|
-
runtimeVersion: string;
|
|
2116
|
-
capabilityReceipt: string;
|
|
2117
|
-
};
|
|
2118
|
-
limits: {
|
|
2119
|
-
maxTurns: number;
|
|
2120
|
-
timeoutMs: number;
|
|
2121
|
-
maxOutputBytes: number;
|
|
2122
|
-
maximumSpendUsd: number | null;
|
|
2123
|
-
};
|
|
2124
|
-
createdAt: string;
|
|
2125
|
-
metadata: Record<string, unknown>;
|
|
2126
|
-
contentHash: string;
|
|
2127
2121
|
profileEvaluation?: {
|
|
2128
2122
|
profileId: string;
|
|
2129
2123
|
sourceRevision: string;
|
|
@@ -2170,7 +2164,35 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
2170
2164
|
};
|
|
2171
2165
|
declaredProfileCatalogHash: string | null;
|
|
2172
2166
|
protectedProfileClosureHash: string;
|
|
2173
|
-
|
|
2167
|
+
recordedOrigin?: {
|
|
2168
|
+
execution: {
|
|
2169
|
+
id: string;
|
|
2170
|
+
contentHash: string;
|
|
2171
|
+
};
|
|
2172
|
+
originalEvidence: {
|
|
2173
|
+
id: string;
|
|
2174
|
+
contentHash: string;
|
|
2175
|
+
};
|
|
2176
|
+
members: {
|
|
2177
|
+
caseId: string;
|
|
2178
|
+
seed: string;
|
|
2179
|
+
fixtureId: string;
|
|
2180
|
+
sourceId: string;
|
|
2181
|
+
snapshotHash: string;
|
|
2182
|
+
boundaryRevisionHash: string;
|
|
2183
|
+
}[];
|
|
2184
|
+
} | undefined;
|
|
2185
|
+
gradingSource?: {
|
|
2186
|
+
pass: {
|
|
2187
|
+
id: string;
|
|
2188
|
+
contentHash: string;
|
|
2189
|
+
};
|
|
2190
|
+
evidence: {
|
|
2191
|
+
id: string;
|
|
2192
|
+
contentHash: string;
|
|
2193
|
+
};
|
|
2194
|
+
} | undefined;
|
|
2195
|
+
split: "frozen_eval" | "test" | "train" | "validation";
|
|
2174
2196
|
population: {
|
|
2175
2197
|
taskId: string;
|
|
2176
2198
|
seed: string;
|
|
@@ -2184,16 +2206,16 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
2184
2206
|
};
|
|
2185
2207
|
feedbackKey: string;
|
|
2186
2208
|
configurationHash: string;
|
|
2187
|
-
output: "boolean" | "
|
|
2209
|
+
output: "boolean" | "category" | "score";
|
|
2188
2210
|
categories: string[];
|
|
2189
2211
|
}[];
|
|
2190
2212
|
fieldMappingsHash: string;
|
|
2191
2213
|
fieldMappings: {
|
|
2192
2214
|
graderId: string;
|
|
2193
2215
|
fields: {
|
|
2194
|
-
destination: "
|
|
2216
|
+
destination: "evaluatorContext" | "expectedOutput" | "input" | "output";
|
|
2195
2217
|
path: string[];
|
|
2196
|
-
source: "
|
|
2218
|
+
source: "evaluatorContext" | "expectedOutput" | "input" | "output";
|
|
2197
2219
|
sourcePath: string[];
|
|
2198
2220
|
}[];
|
|
2199
2221
|
}[];
|
|
@@ -2204,34 +2226,6 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
2204
2226
|
requireComplete: boolean;
|
|
2205
2227
|
};
|
|
2206
2228
|
contentHash: string;
|
|
2207
|
-
recordedOrigin?: {
|
|
2208
|
-
execution: {
|
|
2209
|
-
id: string;
|
|
2210
|
-
contentHash: string;
|
|
2211
|
-
};
|
|
2212
|
-
originalEvidence: {
|
|
2213
|
-
id: string;
|
|
2214
|
-
contentHash: string;
|
|
2215
|
-
};
|
|
2216
|
-
members: {
|
|
2217
|
-
caseId: string;
|
|
2218
|
-
seed: string;
|
|
2219
|
-
fixtureId: string;
|
|
2220
|
-
sourceId: string;
|
|
2221
|
-
snapshotHash: string;
|
|
2222
|
-
boundaryRevisionHash: string;
|
|
2223
|
-
}[];
|
|
2224
|
-
} | undefined;
|
|
2225
|
-
gradingSource?: {
|
|
2226
|
-
pass: {
|
|
2227
|
-
id: string;
|
|
2228
|
-
contentHash: string;
|
|
2229
|
-
};
|
|
2230
|
-
evidence: {
|
|
2231
|
-
id: string;
|
|
2232
|
-
contentHash: string;
|
|
2233
|
-
};
|
|
2234
|
-
} | undefined;
|
|
2235
2229
|
} | undefined;
|
|
2236
2230
|
sourceCandidate?: {
|
|
2237
2231
|
candidateId: string;
|
|
@@ -2244,6 +2238,56 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
2244
2238
|
protectedClosureHash: string;
|
|
2245
2239
|
} | undefined;
|
|
2246
2240
|
} | undefined;
|
|
2241
|
+
policy: {
|
|
2242
|
+
kind: "model";
|
|
2243
|
+
model: {
|
|
2244
|
+
provider: string;
|
|
2245
|
+
model: string;
|
|
2246
|
+
revision: string | null;
|
|
2247
|
+
artifactHash: string | null;
|
|
2248
|
+
tokenizerRevision: string | null;
|
|
2249
|
+
chatTemplateHash: string | null;
|
|
2250
|
+
};
|
|
2251
|
+
configurationHash: string;
|
|
2252
|
+
} | {
|
|
2253
|
+
kind: "fixture";
|
|
2254
|
+
};
|
|
2255
|
+
gradingRole: "evaluation";
|
|
2256
|
+
metricPolicy: {
|
|
2257
|
+
schemaVersion: "openpond.tasksetMetricPolicy.v1";
|
|
2258
|
+
primaryMetric: string;
|
|
2259
|
+
aggregation: "custom" | "mean_score" | "pass_rate" | "weighted_mean";
|
|
2260
|
+
missingReward: "exclude" | "zero";
|
|
2261
|
+
taskWeights?: Record<string, number> | undefined;
|
|
2262
|
+
customAggregator: {
|
|
2263
|
+
module: string;
|
|
2264
|
+
exportName: string;
|
|
2265
|
+
contentHash: string;
|
|
2266
|
+
timeoutMs: number;
|
|
2267
|
+
networkPolicy: "none";
|
|
2268
|
+
} | null;
|
|
2269
|
+
};
|
|
2270
|
+
population: {
|
|
2271
|
+
receiptId: string;
|
|
2272
|
+
taskId: string;
|
|
2273
|
+
seed: string;
|
|
2274
|
+
fixtureId: string | null;
|
|
2275
|
+
}[];
|
|
2276
|
+
runtimeTarget: {
|
|
2277
|
+
adapterId: string;
|
|
2278
|
+
placement: "colocated" | "local" | "provider_native" | "remote";
|
|
2279
|
+
runtimeVersion: string;
|
|
2280
|
+
capabilityReceipt: string;
|
|
2281
|
+
};
|
|
2282
|
+
limits: {
|
|
2283
|
+
maxTurns: number;
|
|
2284
|
+
timeoutMs: number;
|
|
2285
|
+
maxOutputBytes: number;
|
|
2286
|
+
maximumSpendUsd: number | null;
|
|
2287
|
+
};
|
|
2288
|
+
createdAt: string;
|
|
2289
|
+
metadata: Record<string, unknown>;
|
|
2290
|
+
contentHash: string;
|
|
2247
2291
|
};
|
|
2248
2292
|
policySnapshot: {
|
|
2249
2293
|
modelId: string;
|
|
@@ -2262,7 +2306,7 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
2262
2306
|
taskId: string;
|
|
2263
2307
|
seed: string;
|
|
2264
2308
|
terminal: boolean;
|
|
2265
|
-
failureClass: "
|
|
2309
|
+
failureClass: "cancelled" | "environment_failure" | "grader_failure" | "infrastructure_failure" | "policy_failure" | "timeout" | null;
|
|
2266
2310
|
outputHash: string | null;
|
|
2267
2311
|
traceHash: string;
|
|
2268
2312
|
artifactRefs: {
|
|
@@ -2299,7 +2343,8 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
2299
2343
|
schemaVersion: "openpond.tasksetMetricPolicy.v1";
|
|
2300
2344
|
primaryMetric: string;
|
|
2301
2345
|
aggregation: "custom" | "mean_score" | "pass_rate" | "weighted_mean";
|
|
2302
|
-
missingReward: "
|
|
2346
|
+
missingReward: "exclude" | "zero";
|
|
2347
|
+
taskWeights?: Record<string, number> | undefined;
|
|
2303
2348
|
customAggregator: {
|
|
2304
2349
|
module: string;
|
|
2305
2350
|
exportName: string;
|
|
@@ -2307,7 +2352,6 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
2307
2352
|
timeoutMs: number;
|
|
2308
2353
|
networkPolicy: "none";
|
|
2309
2354
|
} | null;
|
|
2310
|
-
taskWeights?: Record<string, number> | undefined;
|
|
2311
2355
|
};
|
|
2312
2356
|
policyHash: string;
|
|
2313
2357
|
receiptRefs: {
|
|
@@ -2331,15 +2375,22 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
2331
2375
|
revision: number;
|
|
2332
2376
|
teamId: string;
|
|
2333
2377
|
modelProjectId: string | null;
|
|
2378
|
+
name?: string | undefined;
|
|
2379
|
+
project?: {
|
|
2380
|
+
id: string;
|
|
2381
|
+
revision: number;
|
|
2382
|
+
contentHash: string;
|
|
2383
|
+
targetId: string | null;
|
|
2384
|
+
} | undefined;
|
|
2334
2385
|
operationId: string;
|
|
2335
2386
|
taskset: {
|
|
2336
2387
|
id: string;
|
|
2337
2388
|
contentHash: string;
|
|
2338
2389
|
revision: number;
|
|
2339
2390
|
};
|
|
2340
|
-
policyKind: "
|
|
2391
|
+
policyKind: "fixture" | "hosted_chat" | "hosted_harness";
|
|
2341
2392
|
manifestHash: string;
|
|
2342
|
-
status: "cancelled" | "
|
|
2393
|
+
status: "cancelled" | "cancelling" | "completed" | "failed" | "queued" | "running";
|
|
2343
2394
|
totalCount: number;
|
|
2344
2395
|
counts: {
|
|
2345
2396
|
pending: number;
|
|
@@ -2359,13 +2410,6 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
2359
2410
|
code: string;
|
|
2360
2411
|
message: string;
|
|
2361
2412
|
} | null;
|
|
2362
|
-
name?: string | undefined;
|
|
2363
|
-
project?: {
|
|
2364
|
-
id: string;
|
|
2365
|
-
revision: number;
|
|
2366
|
-
contentHash: string;
|
|
2367
|
-
targetId: string | null;
|
|
2368
|
-
} | undefined;
|
|
2369
2413
|
}[];
|
|
2370
2414
|
nextCursor: string | null;
|
|
2371
2415
|
}>;
|