openpond-sdk 0.12.3 → 0.12.5
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-77N3QEIL.js → chunk-25L5IXHA.js} +5 -5
- package/dist/chunk-25L5IXHA.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-ZJ3H3VS7.js → chunk-6GQ3GIAP.js} +7 -5
- package/dist/chunk-6GQ3GIAP.js.map +1 -0
- package/dist/{chunk-36WJTS5V.js → chunk-6KKM2XRR.js} +75 -15
- package/dist/chunk-6KKM2XRR.js.map +1 -0
- package/dist/{chunk-BSK3VLM3.js → chunk-6MFFKAU6.js} +3 -3
- package/dist/chunk-6MFFKAU6.js.map +1 -0
- package/dist/{chunk-UPS52OPZ.js → chunk-6Y67YHVH.js} +8 -8
- package/dist/chunk-6Y67YHVH.js.map +1 -0
- package/dist/{chunk-G6Y6CQYV.js → chunk-AUNNXJQE.js} +4 -4
- package/dist/chunk-AUNNXJQE.js.map +1 -0
- package/dist/{chunk-XYQD4LDI.js → chunk-BGNPKYBJ.js} +3 -3
- package/dist/chunk-BGNPKYBJ.js.map +1 -0
- package/dist/{chunk-K5WRVSOL.js → chunk-BJ5MKLWH.js} +8 -8
- package/dist/chunk-BJ5MKLWH.js.map +1 -0
- package/dist/{chunk-VY5AFE7A.js → chunk-BYNVSWIQ.js} +4 -4
- package/dist/chunk-BYNVSWIQ.js.map +1 -0
- package/dist/{chunk-TCHAWJMB.js → chunk-CAMLCKSF.js} +3 -3
- package/dist/chunk-CAMLCKSF.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-CTVKI3BV.js → chunk-DBKCSQVH.js} +3 -3
- package/dist/chunk-DBKCSQVH.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-EJZ7HOIS.js +8709 -0
- package/dist/chunk-EJZ7HOIS.js.map +1 -0
- package/dist/{chunk-QEIT5VXJ.js → chunk-F4GITADW.js} +4 -4
- package/dist/chunk-F4GITADW.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-FY7NHCEH.js → chunk-FLNWE5ZA.js} +5 -5
- package/dist/chunk-FLNWE5ZA.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-YUKV7RCD.js → chunk-GWGDWZVN.js} +5 -5
- package/dist/chunk-GWGDWZVN.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-CEZEVILU.js → chunk-HJ2AIAKX.js} +1 -1
- package/dist/chunk-HJ2AIAKX.js.map +1 -0
- package/dist/{chunk-BFR3VC25.js → chunk-HOOHJCE7.js} +18 -18
- package/dist/chunk-HOOHJCE7.js.map +1 -0
- package/dist/{chunk-TSMGMA6R.js → chunk-KKL6OICH.js} +3 -3
- package/dist/chunk-KKL6OICH.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-QTJQIHDZ.js → chunk-MJELLVQP.js} +4 -3
- package/dist/chunk-MJELLVQP.js.map +1 -0
- package/dist/{chunk-CK3ONRRC.js → chunk-O54KGALH.js} +3 -3
- package/dist/chunk-O54KGALH.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-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-HKWX624G.js → chunk-QFBSNSUL.js} +5 -5
- package/dist/chunk-QFBSNSUL.js.map +1 -0
- package/dist/{chunk-MLPPSCDQ.js → chunk-RE426Q7Y.js} +4 -4
- package/dist/chunk-RE426Q7Y.js.map +1 -0
- package/dist/{chunk-G53MA5XX.js → chunk-T3MFWGMX.js} +3 -3
- package/dist/chunk-T3MFWGMX.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-6Q4I33CH.js → chunk-VZPNRPXL.js} +5 -5
- package/dist/chunk-VZPNRPXL.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-TMUAKJWX.js → chunk-XACJ63U4.js} +14 -5
- package/dist/chunk-XACJ63U4.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-CWXM57QR.js → chunk-ZFYI4P6C.js} +5 -5
- package/dist/chunk-ZFYI4P6C.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 +22 -21
- 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 +26 -25
- 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 +154 -62
- 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 +19 -18
- package/dist/taskset-drafts.js.map +1 -7
- package/dist/taskset-packages.js +25 -22
- package/dist/taskset-packages.js.map +1 -7
- package/dist/training-bundle.js +44 -19
- 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 +11 -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/hosted-chat.d.ts +28 -0
- package/dist/types/packages/cloud/src/hosted-chat.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 +375 -351
- package/dist/types/packages/sdk/src/candidate-evaluation-client.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/candidate-evaluations.d.ts +80 -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 +2102 -2014
- package/dist/types/packages/sdk/src/experiment-client.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/experiment-contracts.d.ts +165 -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 +587 -516
- 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 +38 -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 +303 -243
- 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 +33 -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 +69 -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-bound-taskset-authoring.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 +342 -322
- 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 +179 -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 +865 -833
- package/dist/types/packages/sdk/src/model-taskset-runs.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/post-training.d.ts +444 -375
- 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 +745 -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 +4600 -4576
- 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 +14 -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 +2 -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 +75 -71
- package/dist/types/packages/sdk/src/training-qualification-preparation.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/training.d.ts +1442 -1416
- 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-36WJTS5V.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-77N3QEIL.js.map +0 -7
- package/dist/chunk-AG5DDL7T.js.map +0 -7
- package/dist/chunk-BFR3VC25.js.map +0 -7
- package/dist/chunk-BSK3VLM3.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-CTVKI3BV.js.map +0 -7
- package/dist/chunk-CWXM57QR.js.map +0 -7
- package/dist/chunk-DUKW3R7O.js.map +0 -7
- package/dist/chunk-FHCSDXIA.js +0 -1417
- package/dist/chunk-FHCSDXIA.js.map +0 -7
- package/dist/chunk-FY7NHCEH.js.map +0 -7
- package/dist/chunk-G53MA5XX.js.map +0 -7
- package/dist/chunk-G6Y6CQYV.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-K5WRVSOL.js.map +0 -7
- package/dist/chunk-KAE7AMYE.js.map +0 -7
- package/dist/chunk-MEKWESBF.js.map +0 -7
- package/dist/chunk-MLPPSCDQ.js.map +0 -7
- package/dist/chunk-O4GE6VTR.js.map +0 -7
- package/dist/chunk-P27DQ65Z.js.map +0 -7
- package/dist/chunk-QEIT5VXJ.js.map +0 -7
- package/dist/chunk-QTJQIHDZ.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-TMUAKJWX.js.map +0 -7
- package/dist/chunk-TSMGMA6R.js.map +0 -7
- package/dist/chunk-UPS52OPZ.js.map +0 -7
- package/dist/chunk-VY5AFE7A.js.map +0 -7
- package/dist/chunk-XYQD4LDI.js.map +0 -7
- package/dist/chunk-YCQVVFE5.js.map +0 -7
- package/dist/chunk-YUKV7RCD.js.map +0 -7
- package/dist/chunk-YVUIHH6Z.js.map +0 -7
- package/dist/chunk-ZJ3H3VS7.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;
|
|
@@ -247,28 +286,6 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
247
286
|
};
|
|
248
287
|
modelConfigurationHash: string;
|
|
249
288
|
packageHash: string;
|
|
250
|
-
maxOutputTokens?: number | undefined;
|
|
251
|
-
temperature?: number | undefined;
|
|
252
|
-
topP?: number | undefined;
|
|
253
|
-
candidate?: {
|
|
254
|
-
artifactId: string;
|
|
255
|
-
contentHash: string;
|
|
256
|
-
preparationId: string;
|
|
257
|
-
dispatchId: string;
|
|
258
|
-
baseProfileId: string;
|
|
259
|
-
workerImageDigest: string;
|
|
260
|
-
} | undefined;
|
|
261
|
-
sourceCandidate?: {
|
|
262
|
-
candidateId: string;
|
|
263
|
-
frozenRevision: number;
|
|
264
|
-
frozenHash: string;
|
|
265
|
-
harnessRelease: {
|
|
266
|
-
id: string;
|
|
267
|
-
contentHash: string;
|
|
268
|
-
};
|
|
269
|
-
protectedClosureHash: string;
|
|
270
|
-
} | undefined;
|
|
271
|
-
reasoningEffort?: "high" | "low" | "medium" | "xhigh" | "off" | "max" | undefined;
|
|
272
289
|
externalDatasetBinding?: {
|
|
273
290
|
schemaVersion: "openpond.profileExternalDatasetBinding.v1";
|
|
274
291
|
dataset: {
|
|
@@ -292,7 +309,35 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
292
309
|
};
|
|
293
310
|
declaredProfileCatalogHash: string | null;
|
|
294
311
|
protectedProfileClosureHash: string;
|
|
295
|
-
|
|
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";
|
|
296
341
|
population: {
|
|
297
342
|
taskId: string;
|
|
298
343
|
seed: string;
|
|
@@ -306,16 +351,16 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
306
351
|
};
|
|
307
352
|
feedbackKey: string;
|
|
308
353
|
configurationHash: string;
|
|
309
|
-
output: "boolean" | "
|
|
354
|
+
output: "boolean" | "category" | "score";
|
|
310
355
|
categories: string[];
|
|
311
356
|
}[];
|
|
312
357
|
fieldMappingsHash: string;
|
|
313
358
|
fieldMappings: {
|
|
314
359
|
graderId: string;
|
|
315
360
|
fields: {
|
|
316
|
-
destination: "
|
|
361
|
+
destination: "evaluatorContext" | "expectedOutput" | "input" | "output";
|
|
317
362
|
path: string[];
|
|
318
|
-
source: "
|
|
363
|
+
source: "evaluatorContext" | "expectedOutput" | "input" | "output";
|
|
319
364
|
sourcePath: string[];
|
|
320
365
|
}[];
|
|
321
366
|
}[];
|
|
@@ -326,37 +371,7 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
326
371
|
requireComplete: boolean;
|
|
327
372
|
};
|
|
328
373
|
contentHash: string;
|
|
329
|
-
recordedOrigin?: {
|
|
330
|
-
execution: {
|
|
331
|
-
id: string;
|
|
332
|
-
contentHash: string;
|
|
333
|
-
};
|
|
334
|
-
originalEvidence: {
|
|
335
|
-
id: string;
|
|
336
|
-
contentHash: string;
|
|
337
|
-
};
|
|
338
|
-
members: {
|
|
339
|
-
caseId: string;
|
|
340
|
-
seed: string;
|
|
341
|
-
fixtureId: string;
|
|
342
|
-
sourceId: string;
|
|
343
|
-
snapshotHash: string;
|
|
344
|
-
boundaryRevisionHash: string;
|
|
345
|
-
}[];
|
|
346
|
-
} | undefined;
|
|
347
|
-
gradingSource?: {
|
|
348
|
-
pass: {
|
|
349
|
-
id: string;
|
|
350
|
-
contentHash: string;
|
|
351
|
-
};
|
|
352
|
-
evidence: {
|
|
353
|
-
id: string;
|
|
354
|
-
contentHash: string;
|
|
355
|
-
};
|
|
356
|
-
} | undefined;
|
|
357
374
|
} | undefined;
|
|
358
|
-
} | {
|
|
359
|
-
kind: "fixture";
|
|
360
375
|
};
|
|
361
376
|
population: {
|
|
362
377
|
receiptId: string;
|
|
@@ -364,13 +379,6 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
364
379
|
seed: string;
|
|
365
380
|
fixtureId: string | null;
|
|
366
381
|
}[];
|
|
367
|
-
name?: string | undefined;
|
|
368
|
-
project?: {
|
|
369
|
-
id: string;
|
|
370
|
-
revision: number;
|
|
371
|
-
contentHash: string;
|
|
372
|
-
targetId: string | null;
|
|
373
|
-
} | undefined;
|
|
374
382
|
};
|
|
375
383
|
manifest: {
|
|
376
384
|
schemaVersion: "openpond.tasksetRunManifest.v1";
|
|
@@ -398,61 +406,11 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
398
406
|
contentHash: string;
|
|
399
407
|
};
|
|
400
408
|
};
|
|
401
|
-
|
|
402
|
-
|
|
403
|
-
|
|
404
|
-
|
|
405
|
-
|
|
406
|
-
revision: string | null;
|
|
407
|
-
artifactHash: string | null;
|
|
408
|
-
tokenizerRevision: string | null;
|
|
409
|
-
chatTemplateHash: string | null;
|
|
410
|
-
};
|
|
411
|
-
configurationHash: string;
|
|
412
|
-
} | {
|
|
413
|
-
kind: "fixture";
|
|
414
|
-
};
|
|
415
|
-
gradingRole: "evaluation";
|
|
416
|
-
metricPolicy: {
|
|
417
|
-
schemaVersion: "openpond.tasksetMetricPolicy.v1";
|
|
418
|
-
primaryMetric: string;
|
|
419
|
-
aggregation: "custom" | "mean_score" | "pass_rate" | "weighted_mean";
|
|
420
|
-
missingReward: "zero" | "exclude";
|
|
421
|
-
customAggregator: {
|
|
422
|
-
module: string;
|
|
423
|
-
exportName: string;
|
|
424
|
-
contentHash: string;
|
|
425
|
-
timeoutMs: number;
|
|
426
|
-
networkPolicy: "none";
|
|
427
|
-
} | null;
|
|
428
|
-
taskWeights?: Record<string, number> | undefined;
|
|
429
|
-
};
|
|
430
|
-
population: {
|
|
431
|
-
receiptId: string;
|
|
432
|
-
taskId: string;
|
|
433
|
-
seed: string;
|
|
434
|
-
fixtureId: string | null;
|
|
435
|
-
}[];
|
|
436
|
-
runtimeTarget: {
|
|
437
|
-
adapterId: string;
|
|
438
|
-
placement: "local" | "remote" | "colocated" | "provider_native";
|
|
439
|
-
runtimeVersion: string;
|
|
440
|
-
capabilityReceipt: string;
|
|
441
|
-
};
|
|
442
|
-
limits: {
|
|
443
|
-
maxTurns: number;
|
|
444
|
-
timeoutMs: number;
|
|
445
|
-
maxOutputBytes: number;
|
|
446
|
-
maximumSpendUsd: number | null;
|
|
447
|
-
};
|
|
448
|
-
createdAt: string;
|
|
449
|
-
metadata: Record<string, unknown>;
|
|
450
|
-
contentHash: string;
|
|
451
|
-
profileEvaluation?: {
|
|
452
|
-
profileId: string;
|
|
453
|
-
sourceRevision: string;
|
|
454
|
-
harnessRelease: {
|
|
455
|
-
id: string;
|
|
409
|
+
profileEvaluation?: {
|
|
410
|
+
profileId: string;
|
|
411
|
+
sourceRevision: string;
|
|
412
|
+
harnessRelease: {
|
|
413
|
+
id: string;
|
|
456
414
|
contentHash: string;
|
|
457
415
|
};
|
|
458
416
|
catalogHash: string;
|
|
@@ -494,7 +452,35 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
494
452
|
};
|
|
495
453
|
declaredProfileCatalogHash: string | null;
|
|
496
454
|
protectedProfileClosureHash: string;
|
|
497
|
-
|
|
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";
|
|
498
484
|
population: {
|
|
499
485
|
taskId: string;
|
|
500
486
|
seed: string;
|
|
@@ -508,16 +494,16 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
508
494
|
};
|
|
509
495
|
feedbackKey: string;
|
|
510
496
|
configurationHash: string;
|
|
511
|
-
output: "boolean" | "
|
|
497
|
+
output: "boolean" | "category" | "score";
|
|
512
498
|
categories: string[];
|
|
513
499
|
}[];
|
|
514
500
|
fieldMappingsHash: string;
|
|
515
501
|
fieldMappings: {
|
|
516
502
|
graderId: string;
|
|
517
503
|
fields: {
|
|
518
|
-
destination: "
|
|
504
|
+
destination: "evaluatorContext" | "expectedOutput" | "input" | "output";
|
|
519
505
|
path: string[];
|
|
520
|
-
source: "
|
|
506
|
+
source: "evaluatorContext" | "expectedOutput" | "input" | "output";
|
|
521
507
|
sourcePath: string[];
|
|
522
508
|
}[];
|
|
523
509
|
}[];
|
|
@@ -528,34 +514,6 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
528
514
|
requireComplete: boolean;
|
|
529
515
|
};
|
|
530
516
|
contentHash: string;
|
|
531
|
-
recordedOrigin?: {
|
|
532
|
-
execution: {
|
|
533
|
-
id: string;
|
|
534
|
-
contentHash: string;
|
|
535
|
-
};
|
|
536
|
-
originalEvidence: {
|
|
537
|
-
id: string;
|
|
538
|
-
contentHash: string;
|
|
539
|
-
};
|
|
540
|
-
members: {
|
|
541
|
-
caseId: string;
|
|
542
|
-
seed: string;
|
|
543
|
-
fixtureId: string;
|
|
544
|
-
sourceId: string;
|
|
545
|
-
snapshotHash: string;
|
|
546
|
-
boundaryRevisionHash: string;
|
|
547
|
-
}[];
|
|
548
|
-
} | undefined;
|
|
549
|
-
gradingSource?: {
|
|
550
|
-
pass: {
|
|
551
|
-
id: string;
|
|
552
|
-
contentHash: string;
|
|
553
|
-
};
|
|
554
|
-
evidence: {
|
|
555
|
-
id: string;
|
|
556
|
-
contentHash: string;
|
|
557
|
-
};
|
|
558
|
-
} | undefined;
|
|
559
517
|
} | undefined;
|
|
560
518
|
sourceCandidate?: {
|
|
561
519
|
candidateId: string;
|
|
@@ -568,6 +526,56 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
568
526
|
protectedClosureHash: string;
|
|
569
527
|
} | undefined;
|
|
570
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;
|
|
571
579
|
};
|
|
572
580
|
policySnapshot: {
|
|
573
581
|
modelId: string;
|
|
@@ -585,15 +593,22 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
585
593
|
revision: number;
|
|
586
594
|
teamId: string;
|
|
587
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;
|
|
588
603
|
operationId: string;
|
|
589
604
|
taskset: {
|
|
590
605
|
id: string;
|
|
591
606
|
contentHash: string;
|
|
592
607
|
revision: number;
|
|
593
608
|
};
|
|
594
|
-
policyKind: "
|
|
609
|
+
policyKind: "fixture" | "hosted_chat" | "hosted_harness";
|
|
595
610
|
manifestHash: string;
|
|
596
|
-
status: "cancelled" | "
|
|
611
|
+
status: "cancelled" | "cancelling" | "completed" | "failed" | "queued" | "running";
|
|
597
612
|
totalCount: number;
|
|
598
613
|
counts: {
|
|
599
614
|
pending: number;
|
|
@@ -613,6 +628,12 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
613
628
|
code: string;
|
|
614
629
|
message: string;
|
|
615
630
|
} | null;
|
|
631
|
+
};
|
|
632
|
+
request: {
|
|
633
|
+
schemaVersion: "openpond.modelTasksetRunRequest.v1";
|
|
634
|
+
operationId: string;
|
|
635
|
+
teamId: string;
|
|
636
|
+
modelProjectId: string | null;
|
|
616
637
|
name?: string | undefined;
|
|
617
638
|
project?: {
|
|
618
639
|
id: string;
|
|
@@ -620,12 +641,6 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
620
641
|
contentHash: string;
|
|
621
642
|
targetId: string | null;
|
|
622
643
|
} | undefined;
|
|
623
|
-
};
|
|
624
|
-
request: {
|
|
625
|
-
schemaVersion: "openpond.modelTasksetRunRequest.v1";
|
|
626
|
-
operationId: string;
|
|
627
|
-
teamId: string;
|
|
628
|
-
modelProjectId: string | null;
|
|
629
644
|
taskset: {
|
|
630
645
|
id: string;
|
|
631
646
|
contentHash: string;
|
|
@@ -634,20 +649,7 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
634
649
|
policy: {
|
|
635
650
|
kind: "hosted_chat";
|
|
636
651
|
modelId: string;
|
|
637
|
-
maxOutputTokens: number;
|
|
638
|
-
temperature: number;
|
|
639
|
-
topP: number;
|
|
640
652
|
localRuntime?: {
|
|
641
|
-
providerId: "claude-code";
|
|
642
|
-
configurationHash: string;
|
|
643
|
-
executableHash: string;
|
|
644
|
-
version: string;
|
|
645
|
-
capabilityHash: string;
|
|
646
|
-
maximumRequestCostUsd: number;
|
|
647
|
-
requestTimeoutMs: number;
|
|
648
|
-
maximumTurns: number;
|
|
649
|
-
holdOpen: boolean;
|
|
650
|
-
} | {
|
|
651
653
|
providerId: "codex";
|
|
652
654
|
configurationHash: string;
|
|
653
655
|
accountHash: string;
|
|
@@ -659,6 +661,16 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
659
661
|
configurationHash: string;
|
|
660
662
|
maximumRequestCostUsd: number;
|
|
661
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;
|
|
662
674
|
} | undefined;
|
|
663
675
|
candidate?: {
|
|
664
676
|
artifactId: string;
|
|
@@ -680,12 +692,47 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
680
692
|
sourcePackageHash: string;
|
|
681
693
|
} | undefined;
|
|
682
694
|
messages?: {
|
|
683
|
-
role: "
|
|
695
|
+
role: "assistant" | "system" | "user";
|
|
684
696
|
content: string;
|
|
685
697
|
}[] | undefined;
|
|
698
|
+
maxOutputTokens: number;
|
|
699
|
+
temperature: number;
|
|
700
|
+
topP: number;
|
|
701
|
+
} | {
|
|
702
|
+
kind: "fixture";
|
|
686
703
|
} | {
|
|
687
704
|
kind: "hosted_harness";
|
|
688
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;
|
|
689
736
|
profileRepositoryId: string;
|
|
690
737
|
source: {
|
|
691
738
|
profileId: string;
|
|
@@ -733,7 +780,35 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
733
780
|
};
|
|
734
781
|
declaredProfileCatalogHash: string | null;
|
|
735
782
|
protectedProfileClosureHash: string;
|
|
736
|
-
|
|
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";
|
|
737
812
|
population: {
|
|
738
813
|
taskId: string;
|
|
739
814
|
seed: string;
|
|
@@ -747,16 +822,16 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
747
822
|
};
|
|
748
823
|
feedbackKey: string;
|
|
749
824
|
configurationHash: string;
|
|
750
|
-
output: "boolean" | "
|
|
825
|
+
output: "boolean" | "category" | "score";
|
|
751
826
|
categories: string[];
|
|
752
827
|
}[];
|
|
753
828
|
fieldMappingsHash: string;
|
|
754
829
|
fieldMappings: {
|
|
755
830
|
graderId: string;
|
|
756
831
|
fields: {
|
|
757
|
-
destination: "
|
|
832
|
+
destination: "evaluatorContext" | "expectedOutput" | "input" | "output";
|
|
758
833
|
path: string[];
|
|
759
|
-
source: "
|
|
834
|
+
source: "evaluatorContext" | "expectedOutput" | "input" | "output";
|
|
760
835
|
sourcePath: string[];
|
|
761
836
|
}[];
|
|
762
837
|
}[];
|
|
@@ -767,34 +842,6 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
767
842
|
requireComplete: boolean;
|
|
768
843
|
};
|
|
769
844
|
contentHash: string;
|
|
770
|
-
recordedOrigin?: {
|
|
771
|
-
execution: {
|
|
772
|
-
id: string;
|
|
773
|
-
contentHash: string;
|
|
774
|
-
};
|
|
775
|
-
originalEvidence: {
|
|
776
|
-
id: string;
|
|
777
|
-
contentHash: string;
|
|
778
|
-
};
|
|
779
|
-
members: {
|
|
780
|
-
caseId: string;
|
|
781
|
-
seed: string;
|
|
782
|
-
fixtureId: string;
|
|
783
|
-
sourceId: string;
|
|
784
|
-
snapshotHash: string;
|
|
785
|
-
boundaryRevisionHash: string;
|
|
786
|
-
}[];
|
|
787
|
-
} | undefined;
|
|
788
|
-
gradingSource?: {
|
|
789
|
-
pass: {
|
|
790
|
-
id: string;
|
|
791
|
-
contentHash: string;
|
|
792
|
-
};
|
|
793
|
-
evidence: {
|
|
794
|
-
id: string;
|
|
795
|
-
contentHash: string;
|
|
796
|
-
};
|
|
797
|
-
} | undefined;
|
|
798
845
|
} | undefined;
|
|
799
846
|
sourceCandidate?: {
|
|
800
847
|
candidateId: string;
|
|
@@ -809,28 +856,6 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
809
856
|
};
|
|
810
857
|
modelConfigurationHash: string;
|
|
811
858
|
packageHash: string;
|
|
812
|
-
maxOutputTokens?: number | undefined;
|
|
813
|
-
temperature?: number | undefined;
|
|
814
|
-
topP?: number | undefined;
|
|
815
|
-
candidate?: {
|
|
816
|
-
artifactId: string;
|
|
817
|
-
contentHash: string;
|
|
818
|
-
preparationId: string;
|
|
819
|
-
dispatchId: string;
|
|
820
|
-
baseProfileId: string;
|
|
821
|
-
workerImageDigest: string;
|
|
822
|
-
} | undefined;
|
|
823
|
-
sourceCandidate?: {
|
|
824
|
-
candidateId: string;
|
|
825
|
-
frozenRevision: number;
|
|
826
|
-
frozenHash: string;
|
|
827
|
-
harnessRelease: {
|
|
828
|
-
id: string;
|
|
829
|
-
contentHash: string;
|
|
830
|
-
};
|
|
831
|
-
protectedClosureHash: string;
|
|
832
|
-
} | undefined;
|
|
833
|
-
reasoningEffort?: "high" | "low" | "medium" | "xhigh" | "off" | "max" | undefined;
|
|
834
859
|
externalDatasetBinding?: {
|
|
835
860
|
schemaVersion: "openpond.profileExternalDatasetBinding.v1";
|
|
836
861
|
dataset: {
|
|
@@ -854,7 +879,35 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
854
879
|
};
|
|
855
880
|
declaredProfileCatalogHash: string | null;
|
|
856
881
|
protectedProfileClosureHash: string;
|
|
857
|
-
|
|
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";
|
|
858
911
|
population: {
|
|
859
912
|
taskId: string;
|
|
860
913
|
seed: string;
|
|
@@ -868,16 +921,16 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
868
921
|
};
|
|
869
922
|
feedbackKey: string;
|
|
870
923
|
configurationHash: string;
|
|
871
|
-
output: "boolean" | "
|
|
924
|
+
output: "boolean" | "category" | "score";
|
|
872
925
|
categories: string[];
|
|
873
926
|
}[];
|
|
874
927
|
fieldMappingsHash: string;
|
|
875
928
|
fieldMappings: {
|
|
876
929
|
graderId: string;
|
|
877
930
|
fields: {
|
|
878
|
-
destination: "
|
|
931
|
+
destination: "evaluatorContext" | "expectedOutput" | "input" | "output";
|
|
879
932
|
path: string[];
|
|
880
|
-
source: "
|
|
933
|
+
source: "evaluatorContext" | "expectedOutput" | "input" | "output";
|
|
881
934
|
sourcePath: string[];
|
|
882
935
|
}[];
|
|
883
936
|
}[];
|
|
@@ -888,37 +941,7 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
888
941
|
requireComplete: boolean;
|
|
889
942
|
};
|
|
890
943
|
contentHash: string;
|
|
891
|
-
recordedOrigin?: {
|
|
892
|
-
execution: {
|
|
893
|
-
id: string;
|
|
894
|
-
contentHash: string;
|
|
895
|
-
};
|
|
896
|
-
originalEvidence: {
|
|
897
|
-
id: string;
|
|
898
|
-
contentHash: string;
|
|
899
|
-
};
|
|
900
|
-
members: {
|
|
901
|
-
caseId: string;
|
|
902
|
-
seed: string;
|
|
903
|
-
fixtureId: string;
|
|
904
|
-
sourceId: string;
|
|
905
|
-
snapshotHash: string;
|
|
906
|
-
boundaryRevisionHash: string;
|
|
907
|
-
}[];
|
|
908
|
-
} | undefined;
|
|
909
|
-
gradingSource?: {
|
|
910
|
-
pass: {
|
|
911
|
-
id: string;
|
|
912
|
-
contentHash: string;
|
|
913
|
-
};
|
|
914
|
-
evidence: {
|
|
915
|
-
id: string;
|
|
916
|
-
contentHash: string;
|
|
917
|
-
};
|
|
918
|
-
} | undefined;
|
|
919
944
|
} | undefined;
|
|
920
|
-
} | {
|
|
921
|
-
kind: "fixture";
|
|
922
945
|
};
|
|
923
946
|
population: {
|
|
924
947
|
receiptId: string;
|
|
@@ -926,13 +949,6 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
926
949
|
seed: string;
|
|
927
950
|
fixtureId: string | null;
|
|
928
951
|
}[];
|
|
929
|
-
name?: string | undefined;
|
|
930
|
-
project?: {
|
|
931
|
-
id: string;
|
|
932
|
-
revision: number;
|
|
933
|
-
contentHash: string;
|
|
934
|
-
targetId: string | null;
|
|
935
|
-
} | undefined;
|
|
936
952
|
};
|
|
937
953
|
manifest: {
|
|
938
954
|
schemaVersion: "openpond.tasksetRunManifest.v1";
|
|
@@ -960,56 +976,6 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
960
976
|
contentHash: string;
|
|
961
977
|
};
|
|
962
978
|
};
|
|
963
|
-
policy: {
|
|
964
|
-
kind: "model";
|
|
965
|
-
model: {
|
|
966
|
-
provider: string;
|
|
967
|
-
model: string;
|
|
968
|
-
revision: string | null;
|
|
969
|
-
artifactHash: string | null;
|
|
970
|
-
tokenizerRevision: string | null;
|
|
971
|
-
chatTemplateHash: string | null;
|
|
972
|
-
};
|
|
973
|
-
configurationHash: string;
|
|
974
|
-
} | {
|
|
975
|
-
kind: "fixture";
|
|
976
|
-
};
|
|
977
|
-
gradingRole: "evaluation";
|
|
978
|
-
metricPolicy: {
|
|
979
|
-
schemaVersion: "openpond.tasksetMetricPolicy.v1";
|
|
980
|
-
primaryMetric: string;
|
|
981
|
-
aggregation: "custom" | "mean_score" | "pass_rate" | "weighted_mean";
|
|
982
|
-
missingReward: "zero" | "exclude";
|
|
983
|
-
customAggregator: {
|
|
984
|
-
module: string;
|
|
985
|
-
exportName: string;
|
|
986
|
-
contentHash: string;
|
|
987
|
-
timeoutMs: number;
|
|
988
|
-
networkPolicy: "none";
|
|
989
|
-
} | null;
|
|
990
|
-
taskWeights?: Record<string, number> | undefined;
|
|
991
|
-
};
|
|
992
|
-
population: {
|
|
993
|
-
receiptId: string;
|
|
994
|
-
taskId: string;
|
|
995
|
-
seed: string;
|
|
996
|
-
fixtureId: string | null;
|
|
997
|
-
}[];
|
|
998
|
-
runtimeTarget: {
|
|
999
|
-
adapterId: string;
|
|
1000
|
-
placement: "local" | "remote" | "colocated" | "provider_native";
|
|
1001
|
-
runtimeVersion: string;
|
|
1002
|
-
capabilityReceipt: string;
|
|
1003
|
-
};
|
|
1004
|
-
limits: {
|
|
1005
|
-
maxTurns: number;
|
|
1006
|
-
timeoutMs: number;
|
|
1007
|
-
maxOutputBytes: number;
|
|
1008
|
-
maximumSpendUsd: number | null;
|
|
1009
|
-
};
|
|
1010
|
-
createdAt: string;
|
|
1011
|
-
metadata: Record<string, unknown>;
|
|
1012
|
-
contentHash: string;
|
|
1013
979
|
profileEvaluation?: {
|
|
1014
980
|
profileId: string;
|
|
1015
981
|
sourceRevision: string;
|
|
@@ -1056,7 +1022,35 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
1056
1022
|
};
|
|
1057
1023
|
declaredProfileCatalogHash: string | null;
|
|
1058
1024
|
protectedProfileClosureHash: string;
|
|
1059
|
-
|
|
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";
|
|
1060
1054
|
population: {
|
|
1061
1055
|
taskId: string;
|
|
1062
1056
|
seed: string;
|
|
@@ -1070,16 +1064,16 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
1070
1064
|
};
|
|
1071
1065
|
feedbackKey: string;
|
|
1072
1066
|
configurationHash: string;
|
|
1073
|
-
output: "boolean" | "
|
|
1067
|
+
output: "boolean" | "category" | "score";
|
|
1074
1068
|
categories: string[];
|
|
1075
1069
|
}[];
|
|
1076
1070
|
fieldMappingsHash: string;
|
|
1077
1071
|
fieldMappings: {
|
|
1078
1072
|
graderId: string;
|
|
1079
1073
|
fields: {
|
|
1080
|
-
destination: "
|
|
1074
|
+
destination: "evaluatorContext" | "expectedOutput" | "input" | "output";
|
|
1081
1075
|
path: string[];
|
|
1082
|
-
source: "
|
|
1076
|
+
source: "evaluatorContext" | "expectedOutput" | "input" | "output";
|
|
1083
1077
|
sourcePath: string[];
|
|
1084
1078
|
}[];
|
|
1085
1079
|
}[];
|
|
@@ -1090,34 +1084,6 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
1090
1084
|
requireComplete: boolean;
|
|
1091
1085
|
};
|
|
1092
1086
|
contentHash: string;
|
|
1093
|
-
recordedOrigin?: {
|
|
1094
|
-
execution: {
|
|
1095
|
-
id: string;
|
|
1096
|
-
contentHash: string;
|
|
1097
|
-
};
|
|
1098
|
-
originalEvidence: {
|
|
1099
|
-
id: string;
|
|
1100
|
-
contentHash: string;
|
|
1101
|
-
};
|
|
1102
|
-
members: {
|
|
1103
|
-
caseId: string;
|
|
1104
|
-
seed: string;
|
|
1105
|
-
fixtureId: string;
|
|
1106
|
-
sourceId: string;
|
|
1107
|
-
snapshotHash: string;
|
|
1108
|
-
boundaryRevisionHash: string;
|
|
1109
|
-
}[];
|
|
1110
|
-
} | undefined;
|
|
1111
|
-
gradingSource?: {
|
|
1112
|
-
pass: {
|
|
1113
|
-
id: string;
|
|
1114
|
-
contentHash: string;
|
|
1115
|
-
};
|
|
1116
|
-
evidence: {
|
|
1117
|
-
id: string;
|
|
1118
|
-
contentHash: string;
|
|
1119
|
-
};
|
|
1120
|
-
} | undefined;
|
|
1121
1087
|
} | undefined;
|
|
1122
1088
|
sourceCandidate?: {
|
|
1123
1089
|
candidateId: string;
|
|
@@ -1130,6 +1096,56 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
1130
1096
|
protectedClosureHash: string;
|
|
1131
1097
|
} | undefined;
|
|
1132
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;
|
|
1133
1149
|
};
|
|
1134
1150
|
policySnapshot: {
|
|
1135
1151
|
modelId: string;
|
|
@@ -1147,15 +1163,22 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
1147
1163
|
revision: number;
|
|
1148
1164
|
teamId: string;
|
|
1149
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;
|
|
1150
1173
|
operationId: string;
|
|
1151
1174
|
taskset: {
|
|
1152
1175
|
id: string;
|
|
1153
1176
|
contentHash: string;
|
|
1154
1177
|
revision: number;
|
|
1155
1178
|
};
|
|
1156
|
-
policyKind: "
|
|
1179
|
+
policyKind: "fixture" | "hosted_chat" | "hosted_harness";
|
|
1157
1180
|
manifestHash: string;
|
|
1158
|
-
status: "cancelled" | "
|
|
1181
|
+
status: "cancelled" | "cancelling" | "completed" | "failed" | "queued" | "running";
|
|
1159
1182
|
totalCount: number;
|
|
1160
1183
|
counts: {
|
|
1161
1184
|
pending: number;
|
|
@@ -1175,41 +1198,28 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
1175
1198
|
code: string;
|
|
1176
1199
|
message: string;
|
|
1177
1200
|
} | null;
|
|
1178
|
-
name?: string | undefined;
|
|
1179
|
-
project?: {
|
|
1180
|
-
id: string;
|
|
1181
|
-
revision: number;
|
|
1182
|
-
contentHash: string;
|
|
1183
|
-
targetId: string | null;
|
|
1184
|
-
} | undefined;
|
|
1185
1201
|
};
|
|
1186
1202
|
request: {
|
|
1187
1203
|
schemaVersion: "openpond.modelTasksetRunRequest.v1";
|
|
1188
1204
|
operationId: string;
|
|
1189
1205
|
teamId: string;
|
|
1190
1206
|
modelProjectId: string | null;
|
|
1191
|
-
|
|
1207
|
+
name?: string | undefined;
|
|
1208
|
+
project?: {
|
|
1192
1209
|
id: string;
|
|
1193
|
-
|
|
1210
|
+
revision: number;
|
|
1211
|
+
contentHash: string;
|
|
1212
|
+
targetId: string | null;
|
|
1213
|
+
} | undefined;
|
|
1214
|
+
taskset: {
|
|
1215
|
+
id: string;
|
|
1216
|
+
contentHash: string;
|
|
1194
1217
|
revision: number;
|
|
1195
1218
|
};
|
|
1196
1219
|
policy: {
|
|
1197
1220
|
kind: "hosted_chat";
|
|
1198
1221
|
modelId: string;
|
|
1199
|
-
maxOutputTokens: number;
|
|
1200
|
-
temperature: number;
|
|
1201
|
-
topP: number;
|
|
1202
1222
|
localRuntime?: {
|
|
1203
|
-
providerId: "claude-code";
|
|
1204
|
-
configurationHash: string;
|
|
1205
|
-
executableHash: string;
|
|
1206
|
-
version: string;
|
|
1207
|
-
capabilityHash: string;
|
|
1208
|
-
maximumRequestCostUsd: number;
|
|
1209
|
-
requestTimeoutMs: number;
|
|
1210
|
-
maximumTurns: number;
|
|
1211
|
-
holdOpen: boolean;
|
|
1212
|
-
} | {
|
|
1213
1223
|
providerId: "codex";
|
|
1214
1224
|
configurationHash: string;
|
|
1215
1225
|
accountHash: string;
|
|
@@ -1221,6 +1231,16 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
1221
1231
|
configurationHash: string;
|
|
1222
1232
|
maximumRequestCostUsd: number;
|
|
1223
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;
|
|
1224
1244
|
} | undefined;
|
|
1225
1245
|
candidate?: {
|
|
1226
1246
|
artifactId: string;
|
|
@@ -1242,12 +1262,47 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
1242
1262
|
sourcePackageHash: string;
|
|
1243
1263
|
} | undefined;
|
|
1244
1264
|
messages?: {
|
|
1245
|
-
role: "
|
|
1265
|
+
role: "assistant" | "system" | "user";
|
|
1246
1266
|
content: string;
|
|
1247
1267
|
}[] | undefined;
|
|
1268
|
+
maxOutputTokens: number;
|
|
1269
|
+
temperature: number;
|
|
1270
|
+
topP: number;
|
|
1271
|
+
} | {
|
|
1272
|
+
kind: "fixture";
|
|
1248
1273
|
} | {
|
|
1249
1274
|
kind: "hosted_harness";
|
|
1250
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;
|
|
1251
1306
|
profileRepositoryId: string;
|
|
1252
1307
|
source: {
|
|
1253
1308
|
profileId: string;
|
|
@@ -1295,7 +1350,35 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
1295
1350
|
};
|
|
1296
1351
|
declaredProfileCatalogHash: string | null;
|
|
1297
1352
|
protectedProfileClosureHash: string;
|
|
1298
|
-
|
|
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";
|
|
1299
1382
|
population: {
|
|
1300
1383
|
taskId: string;
|
|
1301
1384
|
seed: string;
|
|
@@ -1309,16 +1392,16 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
1309
1392
|
};
|
|
1310
1393
|
feedbackKey: string;
|
|
1311
1394
|
configurationHash: string;
|
|
1312
|
-
output: "boolean" | "
|
|
1395
|
+
output: "boolean" | "category" | "score";
|
|
1313
1396
|
categories: string[];
|
|
1314
1397
|
}[];
|
|
1315
1398
|
fieldMappingsHash: string;
|
|
1316
1399
|
fieldMappings: {
|
|
1317
1400
|
graderId: string;
|
|
1318
1401
|
fields: {
|
|
1319
|
-
destination: "
|
|
1402
|
+
destination: "evaluatorContext" | "expectedOutput" | "input" | "output";
|
|
1320
1403
|
path: string[];
|
|
1321
|
-
source: "
|
|
1404
|
+
source: "evaluatorContext" | "expectedOutput" | "input" | "output";
|
|
1322
1405
|
sourcePath: string[];
|
|
1323
1406
|
}[];
|
|
1324
1407
|
}[];
|
|
@@ -1329,34 +1412,6 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
1329
1412
|
requireComplete: boolean;
|
|
1330
1413
|
};
|
|
1331
1414
|
contentHash: string;
|
|
1332
|
-
recordedOrigin?: {
|
|
1333
|
-
execution: {
|
|
1334
|
-
id: string;
|
|
1335
|
-
contentHash: string;
|
|
1336
|
-
};
|
|
1337
|
-
originalEvidence: {
|
|
1338
|
-
id: string;
|
|
1339
|
-
contentHash: string;
|
|
1340
|
-
};
|
|
1341
|
-
members: {
|
|
1342
|
-
caseId: string;
|
|
1343
|
-
seed: string;
|
|
1344
|
-
fixtureId: string;
|
|
1345
|
-
sourceId: string;
|
|
1346
|
-
snapshotHash: string;
|
|
1347
|
-
boundaryRevisionHash: string;
|
|
1348
|
-
}[];
|
|
1349
|
-
} | undefined;
|
|
1350
|
-
gradingSource?: {
|
|
1351
|
-
pass: {
|
|
1352
|
-
id: string;
|
|
1353
|
-
contentHash: string;
|
|
1354
|
-
};
|
|
1355
|
-
evidence: {
|
|
1356
|
-
id: string;
|
|
1357
|
-
contentHash: string;
|
|
1358
|
-
};
|
|
1359
|
-
} | undefined;
|
|
1360
1415
|
} | undefined;
|
|
1361
1416
|
sourceCandidate?: {
|
|
1362
1417
|
candidateId: string;
|
|
@@ -1371,28 +1426,6 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
1371
1426
|
};
|
|
1372
1427
|
modelConfigurationHash: string;
|
|
1373
1428
|
packageHash: string;
|
|
1374
|
-
maxOutputTokens?: number | undefined;
|
|
1375
|
-
temperature?: number | undefined;
|
|
1376
|
-
topP?: number | undefined;
|
|
1377
|
-
candidate?: {
|
|
1378
|
-
artifactId: string;
|
|
1379
|
-
contentHash: string;
|
|
1380
|
-
preparationId: string;
|
|
1381
|
-
dispatchId: string;
|
|
1382
|
-
baseProfileId: string;
|
|
1383
|
-
workerImageDigest: string;
|
|
1384
|
-
} | undefined;
|
|
1385
|
-
sourceCandidate?: {
|
|
1386
|
-
candidateId: string;
|
|
1387
|
-
frozenRevision: number;
|
|
1388
|
-
frozenHash: string;
|
|
1389
|
-
harnessRelease: {
|
|
1390
|
-
id: string;
|
|
1391
|
-
contentHash: string;
|
|
1392
|
-
};
|
|
1393
|
-
protectedClosureHash: string;
|
|
1394
|
-
} | undefined;
|
|
1395
|
-
reasoningEffort?: "high" | "low" | "medium" | "xhigh" | "off" | "max" | undefined;
|
|
1396
1429
|
externalDatasetBinding?: {
|
|
1397
1430
|
schemaVersion: "openpond.profileExternalDatasetBinding.v1";
|
|
1398
1431
|
dataset: {
|
|
@@ -1416,7 +1449,35 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
1416
1449
|
};
|
|
1417
1450
|
declaredProfileCatalogHash: string | null;
|
|
1418
1451
|
protectedProfileClosureHash: string;
|
|
1419
|
-
|
|
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";
|
|
1420
1481
|
population: {
|
|
1421
1482
|
taskId: string;
|
|
1422
1483
|
seed: string;
|
|
@@ -1430,16 +1491,16 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
1430
1491
|
};
|
|
1431
1492
|
feedbackKey: string;
|
|
1432
1493
|
configurationHash: string;
|
|
1433
|
-
output: "boolean" | "
|
|
1494
|
+
output: "boolean" | "category" | "score";
|
|
1434
1495
|
categories: string[];
|
|
1435
1496
|
}[];
|
|
1436
1497
|
fieldMappingsHash: string;
|
|
1437
1498
|
fieldMappings: {
|
|
1438
1499
|
graderId: string;
|
|
1439
1500
|
fields: {
|
|
1440
|
-
destination: "
|
|
1501
|
+
destination: "evaluatorContext" | "expectedOutput" | "input" | "output";
|
|
1441
1502
|
path: string[];
|
|
1442
|
-
source: "
|
|
1503
|
+
source: "evaluatorContext" | "expectedOutput" | "input" | "output";
|
|
1443
1504
|
sourcePath: string[];
|
|
1444
1505
|
}[];
|
|
1445
1506
|
}[];
|
|
@@ -1450,37 +1511,7 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
1450
1511
|
requireComplete: boolean;
|
|
1451
1512
|
};
|
|
1452
1513
|
contentHash: string;
|
|
1453
|
-
recordedOrigin?: {
|
|
1454
|
-
execution: {
|
|
1455
|
-
id: string;
|
|
1456
|
-
contentHash: string;
|
|
1457
|
-
};
|
|
1458
|
-
originalEvidence: {
|
|
1459
|
-
id: string;
|
|
1460
|
-
contentHash: string;
|
|
1461
|
-
};
|
|
1462
|
-
members: {
|
|
1463
|
-
caseId: string;
|
|
1464
|
-
seed: string;
|
|
1465
|
-
fixtureId: string;
|
|
1466
|
-
sourceId: string;
|
|
1467
|
-
snapshotHash: string;
|
|
1468
|
-
boundaryRevisionHash: string;
|
|
1469
|
-
}[];
|
|
1470
|
-
} | undefined;
|
|
1471
|
-
gradingSource?: {
|
|
1472
|
-
pass: {
|
|
1473
|
-
id: string;
|
|
1474
|
-
contentHash: string;
|
|
1475
|
-
};
|
|
1476
|
-
evidence: {
|
|
1477
|
-
id: string;
|
|
1478
|
-
contentHash: string;
|
|
1479
|
-
};
|
|
1480
|
-
} | undefined;
|
|
1481
1514
|
} | undefined;
|
|
1482
|
-
} | {
|
|
1483
|
-
kind: "fixture";
|
|
1484
1515
|
};
|
|
1485
1516
|
population: {
|
|
1486
1517
|
receiptId: string;
|
|
@@ -1488,13 +1519,6 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
1488
1519
|
seed: string;
|
|
1489
1520
|
fixtureId: string | null;
|
|
1490
1521
|
}[];
|
|
1491
|
-
name?: string | undefined;
|
|
1492
|
-
project?: {
|
|
1493
|
-
id: string;
|
|
1494
|
-
revision: number;
|
|
1495
|
-
contentHash: string;
|
|
1496
|
-
targetId: string | null;
|
|
1497
|
-
} | undefined;
|
|
1498
1522
|
};
|
|
1499
1523
|
manifest: {
|
|
1500
1524
|
schemaVersion: "openpond.tasksetRunManifest.v1";
|
|
@@ -1506,72 +1530,22 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
1506
1530
|
packageHash: string;
|
|
1507
1531
|
execution: {
|
|
1508
1532
|
kind: "taskset";
|
|
1509
|
-
environmentRelease: {
|
|
1510
|
-
id: string;
|
|
1511
|
-
contentHash: string;
|
|
1512
|
-
};
|
|
1513
|
-
verifierSetRelease: {
|
|
1514
|
-
id: string;
|
|
1515
|
-
contentHash: string;
|
|
1516
|
-
};
|
|
1517
|
-
policyHash: string;
|
|
1518
|
-
} | {
|
|
1519
|
-
kind: "harness";
|
|
1520
|
-
harnessRelease: {
|
|
1521
|
-
id: string;
|
|
1522
|
-
contentHash: string;
|
|
1523
|
-
};
|
|
1524
|
-
};
|
|
1525
|
-
policy: {
|
|
1526
|
-
kind: "model";
|
|
1527
|
-
model: {
|
|
1528
|
-
provider: string;
|
|
1529
|
-
model: string;
|
|
1530
|
-
revision: string | null;
|
|
1531
|
-
artifactHash: string | null;
|
|
1532
|
-
tokenizerRevision: string | null;
|
|
1533
|
-
chatTemplateHash: string | null;
|
|
1534
|
-
};
|
|
1535
|
-
configurationHash: string;
|
|
1536
|
-
} | {
|
|
1537
|
-
kind: "fixture";
|
|
1538
|
-
};
|
|
1539
|
-
gradingRole: "evaluation";
|
|
1540
|
-
metricPolicy: {
|
|
1541
|
-
schemaVersion: "openpond.tasksetMetricPolicy.v1";
|
|
1542
|
-
primaryMetric: string;
|
|
1543
|
-
aggregation: "custom" | "mean_score" | "pass_rate" | "weighted_mean";
|
|
1544
|
-
missingReward: "zero" | "exclude";
|
|
1545
|
-
customAggregator: {
|
|
1546
|
-
module: string;
|
|
1547
|
-
exportName: string;
|
|
1548
|
-
contentHash: string;
|
|
1549
|
-
timeoutMs: number;
|
|
1550
|
-
networkPolicy: "none";
|
|
1551
|
-
} | null;
|
|
1552
|
-
taskWeights?: Record<string, number> | undefined;
|
|
1553
|
-
};
|
|
1554
|
-
population: {
|
|
1555
|
-
receiptId: string;
|
|
1556
|
-
taskId: string;
|
|
1557
|
-
seed: string;
|
|
1558
|
-
fixtureId: string | null;
|
|
1559
|
-
}[];
|
|
1560
|
-
runtimeTarget: {
|
|
1561
|
-
adapterId: string;
|
|
1562
|
-
placement: "local" | "remote" | "colocated" | "provider_native";
|
|
1563
|
-
runtimeVersion: string;
|
|
1564
|
-
capabilityReceipt: string;
|
|
1565
|
-
};
|
|
1566
|
-
limits: {
|
|
1567
|
-
maxTurns: number;
|
|
1568
|
-
timeoutMs: number;
|
|
1569
|
-
maxOutputBytes: number;
|
|
1570
|
-
maximumSpendUsd: number | null;
|
|
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
|
+
};
|
|
1571
1548
|
};
|
|
1572
|
-
createdAt: string;
|
|
1573
|
-
metadata: Record<string, unknown>;
|
|
1574
|
-
contentHash: string;
|
|
1575
1549
|
profileEvaluation?: {
|
|
1576
1550
|
profileId: string;
|
|
1577
1551
|
sourceRevision: string;
|
|
@@ -1618,7 +1592,35 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
1618
1592
|
};
|
|
1619
1593
|
declaredProfileCatalogHash: string | null;
|
|
1620
1594
|
protectedProfileClosureHash: string;
|
|
1621
|
-
|
|
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";
|
|
1622
1624
|
population: {
|
|
1623
1625
|
taskId: string;
|
|
1624
1626
|
seed: string;
|
|
@@ -1632,16 +1634,16 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
1632
1634
|
};
|
|
1633
1635
|
feedbackKey: string;
|
|
1634
1636
|
configurationHash: string;
|
|
1635
|
-
output: "boolean" | "
|
|
1637
|
+
output: "boolean" | "category" | "score";
|
|
1636
1638
|
categories: string[];
|
|
1637
1639
|
}[];
|
|
1638
1640
|
fieldMappingsHash: string;
|
|
1639
1641
|
fieldMappings: {
|
|
1640
1642
|
graderId: string;
|
|
1641
1643
|
fields: {
|
|
1642
|
-
destination: "
|
|
1644
|
+
destination: "evaluatorContext" | "expectedOutput" | "input" | "output";
|
|
1643
1645
|
path: string[];
|
|
1644
|
-
source: "
|
|
1646
|
+
source: "evaluatorContext" | "expectedOutput" | "input" | "output";
|
|
1645
1647
|
sourcePath: string[];
|
|
1646
1648
|
}[];
|
|
1647
1649
|
}[];
|
|
@@ -1652,34 +1654,6 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
1652
1654
|
requireComplete: boolean;
|
|
1653
1655
|
};
|
|
1654
1656
|
contentHash: string;
|
|
1655
|
-
recordedOrigin?: {
|
|
1656
|
-
execution: {
|
|
1657
|
-
id: string;
|
|
1658
|
-
contentHash: string;
|
|
1659
|
-
};
|
|
1660
|
-
originalEvidence: {
|
|
1661
|
-
id: string;
|
|
1662
|
-
contentHash: string;
|
|
1663
|
-
};
|
|
1664
|
-
members: {
|
|
1665
|
-
caseId: string;
|
|
1666
|
-
seed: string;
|
|
1667
|
-
fixtureId: string;
|
|
1668
|
-
sourceId: string;
|
|
1669
|
-
snapshotHash: string;
|
|
1670
|
-
boundaryRevisionHash: string;
|
|
1671
|
-
}[];
|
|
1672
|
-
} | undefined;
|
|
1673
|
-
gradingSource?: {
|
|
1674
|
-
pass: {
|
|
1675
|
-
id: string;
|
|
1676
|
-
contentHash: string;
|
|
1677
|
-
};
|
|
1678
|
-
evidence: {
|
|
1679
|
-
id: string;
|
|
1680
|
-
contentHash: string;
|
|
1681
|
-
};
|
|
1682
|
-
} | undefined;
|
|
1683
1657
|
} | undefined;
|
|
1684
1658
|
sourceCandidate?: {
|
|
1685
1659
|
candidateId: string;
|
|
@@ -1692,6 +1666,56 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
1692
1666
|
protectedClosureHash: string;
|
|
1693
1667
|
} | undefined;
|
|
1694
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;
|
|
1695
1719
|
};
|
|
1696
1720
|
policySnapshot: {
|
|
1697
1721
|
modelId: string;
|
|
@@ -1711,15 +1735,22 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
1711
1735
|
revision: number;
|
|
1712
1736
|
teamId: string;
|
|
1713
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;
|
|
1714
1745
|
operationId: string;
|
|
1715
1746
|
taskset: {
|
|
1716
1747
|
id: string;
|
|
1717
1748
|
contentHash: string;
|
|
1718
1749
|
revision: number;
|
|
1719
1750
|
};
|
|
1720
|
-
policyKind: "
|
|
1751
|
+
policyKind: "fixture" | "hosted_chat" | "hosted_harness";
|
|
1721
1752
|
manifestHash: string;
|
|
1722
|
-
status: "cancelled" | "
|
|
1753
|
+
status: "cancelled" | "cancelling" | "completed" | "failed" | "queued" | "running";
|
|
1723
1754
|
totalCount: number;
|
|
1724
1755
|
counts: {
|
|
1725
1756
|
pending: number;
|
|
@@ -1739,6 +1770,12 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
1739
1770
|
code: string;
|
|
1740
1771
|
message: string;
|
|
1741
1772
|
} | null;
|
|
1773
|
+
};
|
|
1774
|
+
request: {
|
|
1775
|
+
schemaVersion: "openpond.modelTasksetRunRequest.v1";
|
|
1776
|
+
operationId: string;
|
|
1777
|
+
teamId: string;
|
|
1778
|
+
modelProjectId: string | null;
|
|
1742
1779
|
name?: string | undefined;
|
|
1743
1780
|
project?: {
|
|
1744
1781
|
id: string;
|
|
@@ -1746,12 +1783,6 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
1746
1783
|
contentHash: string;
|
|
1747
1784
|
targetId: string | null;
|
|
1748
1785
|
} | undefined;
|
|
1749
|
-
};
|
|
1750
|
-
request: {
|
|
1751
|
-
schemaVersion: "openpond.modelTasksetRunRequest.v1";
|
|
1752
|
-
operationId: string;
|
|
1753
|
-
teamId: string;
|
|
1754
|
-
modelProjectId: string | null;
|
|
1755
1786
|
taskset: {
|
|
1756
1787
|
id: string;
|
|
1757
1788
|
contentHash: string;
|
|
@@ -1760,20 +1791,7 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
1760
1791
|
policy: {
|
|
1761
1792
|
kind: "hosted_chat";
|
|
1762
1793
|
modelId: string;
|
|
1763
|
-
maxOutputTokens: number;
|
|
1764
|
-
temperature: number;
|
|
1765
|
-
topP: number;
|
|
1766
1794
|
localRuntime?: {
|
|
1767
|
-
providerId: "claude-code";
|
|
1768
|
-
configurationHash: string;
|
|
1769
|
-
executableHash: string;
|
|
1770
|
-
version: string;
|
|
1771
|
-
capabilityHash: string;
|
|
1772
|
-
maximumRequestCostUsd: number;
|
|
1773
|
-
requestTimeoutMs: number;
|
|
1774
|
-
maximumTurns: number;
|
|
1775
|
-
holdOpen: boolean;
|
|
1776
|
-
} | {
|
|
1777
1795
|
providerId: "codex";
|
|
1778
1796
|
configurationHash: string;
|
|
1779
1797
|
accountHash: string;
|
|
@@ -1785,6 +1803,16 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
1785
1803
|
configurationHash: string;
|
|
1786
1804
|
maximumRequestCostUsd: number;
|
|
1787
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;
|
|
1788
1816
|
} | undefined;
|
|
1789
1817
|
candidate?: {
|
|
1790
1818
|
artifactId: string;
|
|
@@ -1806,12 +1834,47 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
1806
1834
|
sourcePackageHash: string;
|
|
1807
1835
|
} | undefined;
|
|
1808
1836
|
messages?: {
|
|
1809
|
-
role: "
|
|
1837
|
+
role: "assistant" | "system" | "user";
|
|
1810
1838
|
content: string;
|
|
1811
1839
|
}[] | undefined;
|
|
1840
|
+
maxOutputTokens: number;
|
|
1841
|
+
temperature: number;
|
|
1842
|
+
topP: number;
|
|
1843
|
+
} | {
|
|
1844
|
+
kind: "fixture";
|
|
1812
1845
|
} | {
|
|
1813
1846
|
kind: "hosted_harness";
|
|
1814
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;
|
|
1815
1878
|
profileRepositoryId: string;
|
|
1816
1879
|
source: {
|
|
1817
1880
|
profileId: string;
|
|
@@ -1859,40 +1922,6 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
1859
1922
|
};
|
|
1860
1923
|
declaredProfileCatalogHash: string | null;
|
|
1861
1924
|
protectedProfileClosureHash: string;
|
|
1862
|
-
split: "train" | "validation" | "test" | "frozen_eval";
|
|
1863
|
-
population: {
|
|
1864
|
-
taskId: string;
|
|
1865
|
-
seed: string;
|
|
1866
|
-
fixtureId: string | null;
|
|
1867
|
-
}[];
|
|
1868
|
-
evaluators: {
|
|
1869
|
-
release: {
|
|
1870
|
-
id: string;
|
|
1871
|
-
contentHash: string;
|
|
1872
|
-
revision: string | number;
|
|
1873
|
-
};
|
|
1874
|
-
feedbackKey: string;
|
|
1875
|
-
configurationHash: string;
|
|
1876
|
-
output: "boolean" | "score" | "category";
|
|
1877
|
-
categories: string[];
|
|
1878
|
-
}[];
|
|
1879
|
-
fieldMappingsHash: string;
|
|
1880
|
-
fieldMappings: {
|
|
1881
|
-
graderId: string;
|
|
1882
|
-
fields: {
|
|
1883
|
-
destination: "input" | "output" | "expectedOutput" | "evaluatorContext";
|
|
1884
|
-
path: string[];
|
|
1885
|
-
source: "input" | "output" | "expectedOutput" | "evaluatorContext";
|
|
1886
|
-
sourcePath: string[];
|
|
1887
|
-
}[];
|
|
1888
|
-
}[];
|
|
1889
|
-
environmentHash: string;
|
|
1890
|
-
privateDatasetClosureHash: string;
|
|
1891
|
-
criterion: {
|
|
1892
|
-
minimumPassRate: number;
|
|
1893
|
-
requireComplete: boolean;
|
|
1894
|
-
};
|
|
1895
|
-
contentHash: string;
|
|
1896
1925
|
recordedOrigin?: {
|
|
1897
1926
|
execution: {
|
|
1898
1927
|
id: string;
|
|
@@ -1921,6 +1950,40 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
1921
1950
|
contentHash: string;
|
|
1922
1951
|
};
|
|
1923
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;
|
|
1924
1987
|
} | undefined;
|
|
1925
1988
|
sourceCandidate?: {
|
|
1926
1989
|
candidateId: string;
|
|
@@ -1935,28 +1998,6 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
1935
1998
|
};
|
|
1936
1999
|
modelConfigurationHash: string;
|
|
1937
2000
|
packageHash: string;
|
|
1938
|
-
maxOutputTokens?: number | undefined;
|
|
1939
|
-
temperature?: number | undefined;
|
|
1940
|
-
topP?: number | undefined;
|
|
1941
|
-
candidate?: {
|
|
1942
|
-
artifactId: string;
|
|
1943
|
-
contentHash: string;
|
|
1944
|
-
preparationId: string;
|
|
1945
|
-
dispatchId: string;
|
|
1946
|
-
baseProfileId: string;
|
|
1947
|
-
workerImageDigest: string;
|
|
1948
|
-
} | undefined;
|
|
1949
|
-
sourceCandidate?: {
|
|
1950
|
-
candidateId: string;
|
|
1951
|
-
frozenRevision: number;
|
|
1952
|
-
frozenHash: string;
|
|
1953
|
-
harnessRelease: {
|
|
1954
|
-
id: string;
|
|
1955
|
-
contentHash: string;
|
|
1956
|
-
};
|
|
1957
|
-
protectedClosureHash: string;
|
|
1958
|
-
} | undefined;
|
|
1959
|
-
reasoningEffort?: "high" | "low" | "medium" | "xhigh" | "off" | "max" | undefined;
|
|
1960
2001
|
externalDatasetBinding?: {
|
|
1961
2002
|
schemaVersion: "openpond.profileExternalDatasetBinding.v1";
|
|
1962
2003
|
dataset: {
|
|
@@ -1980,7 +2021,35 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
1980
2021
|
};
|
|
1981
2022
|
declaredProfileCatalogHash: string | null;
|
|
1982
2023
|
protectedProfileClosureHash: string;
|
|
1983
|
-
|
|
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";
|
|
1984
2053
|
population: {
|
|
1985
2054
|
taskId: string;
|
|
1986
2055
|
seed: string;
|
|
@@ -1994,16 +2063,16 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
1994
2063
|
};
|
|
1995
2064
|
feedbackKey: string;
|
|
1996
2065
|
configurationHash: string;
|
|
1997
|
-
output: "boolean" | "
|
|
2066
|
+
output: "boolean" | "category" | "score";
|
|
1998
2067
|
categories: string[];
|
|
1999
2068
|
}[];
|
|
2000
2069
|
fieldMappingsHash: string;
|
|
2001
2070
|
fieldMappings: {
|
|
2002
2071
|
graderId: string;
|
|
2003
2072
|
fields: {
|
|
2004
|
-
destination: "
|
|
2073
|
+
destination: "evaluatorContext" | "expectedOutput" | "input" | "output";
|
|
2005
2074
|
path: string[];
|
|
2006
|
-
source: "
|
|
2075
|
+
source: "evaluatorContext" | "expectedOutput" | "input" | "output";
|
|
2007
2076
|
sourcePath: string[];
|
|
2008
2077
|
}[];
|
|
2009
2078
|
}[];
|
|
@@ -2014,37 +2083,7 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
2014
2083
|
requireComplete: boolean;
|
|
2015
2084
|
};
|
|
2016
2085
|
contentHash: string;
|
|
2017
|
-
recordedOrigin?: {
|
|
2018
|
-
execution: {
|
|
2019
|
-
id: string;
|
|
2020
|
-
contentHash: string;
|
|
2021
|
-
};
|
|
2022
|
-
originalEvidence: {
|
|
2023
|
-
id: string;
|
|
2024
|
-
contentHash: string;
|
|
2025
|
-
};
|
|
2026
|
-
members: {
|
|
2027
|
-
caseId: string;
|
|
2028
|
-
seed: string;
|
|
2029
|
-
fixtureId: string;
|
|
2030
|
-
sourceId: string;
|
|
2031
|
-
snapshotHash: string;
|
|
2032
|
-
boundaryRevisionHash: string;
|
|
2033
|
-
}[];
|
|
2034
|
-
} | undefined;
|
|
2035
|
-
gradingSource?: {
|
|
2036
|
-
pass: {
|
|
2037
|
-
id: string;
|
|
2038
|
-
contentHash: string;
|
|
2039
|
-
};
|
|
2040
|
-
evidence: {
|
|
2041
|
-
id: string;
|
|
2042
|
-
contentHash: string;
|
|
2043
|
-
};
|
|
2044
|
-
} | undefined;
|
|
2045
2086
|
} | undefined;
|
|
2046
|
-
} | {
|
|
2047
|
-
kind: "fixture";
|
|
2048
2087
|
};
|
|
2049
2088
|
population: {
|
|
2050
2089
|
receiptId: string;
|
|
@@ -2052,13 +2091,6 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
2052
2091
|
seed: string;
|
|
2053
2092
|
fixtureId: string | null;
|
|
2054
2093
|
}[];
|
|
2055
|
-
name?: string | undefined;
|
|
2056
|
-
project?: {
|
|
2057
|
-
id: string;
|
|
2058
|
-
revision: number;
|
|
2059
|
-
contentHash: string;
|
|
2060
|
-
targetId: string | null;
|
|
2061
|
-
} | undefined;
|
|
2062
2094
|
};
|
|
2063
2095
|
manifest: {
|
|
2064
2096
|
schemaVersion: "openpond.tasksetRunManifest.v1";
|
|
@@ -2086,56 +2118,6 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
2086
2118
|
contentHash: string;
|
|
2087
2119
|
};
|
|
2088
2120
|
};
|
|
2089
|
-
policy: {
|
|
2090
|
-
kind: "model";
|
|
2091
|
-
model: {
|
|
2092
|
-
provider: string;
|
|
2093
|
-
model: string;
|
|
2094
|
-
revision: string | null;
|
|
2095
|
-
artifactHash: string | null;
|
|
2096
|
-
tokenizerRevision: string | null;
|
|
2097
|
-
chatTemplateHash: string | null;
|
|
2098
|
-
};
|
|
2099
|
-
configurationHash: string;
|
|
2100
|
-
} | {
|
|
2101
|
-
kind: "fixture";
|
|
2102
|
-
};
|
|
2103
|
-
gradingRole: "evaluation";
|
|
2104
|
-
metricPolicy: {
|
|
2105
|
-
schemaVersion: "openpond.tasksetMetricPolicy.v1";
|
|
2106
|
-
primaryMetric: string;
|
|
2107
|
-
aggregation: "custom" | "mean_score" | "pass_rate" | "weighted_mean";
|
|
2108
|
-
missingReward: "zero" | "exclude";
|
|
2109
|
-
customAggregator: {
|
|
2110
|
-
module: string;
|
|
2111
|
-
exportName: string;
|
|
2112
|
-
contentHash: string;
|
|
2113
|
-
timeoutMs: number;
|
|
2114
|
-
networkPolicy: "none";
|
|
2115
|
-
} | null;
|
|
2116
|
-
taskWeights?: Record<string, number> | undefined;
|
|
2117
|
-
};
|
|
2118
|
-
population: {
|
|
2119
|
-
receiptId: string;
|
|
2120
|
-
taskId: string;
|
|
2121
|
-
seed: string;
|
|
2122
|
-
fixtureId: string | null;
|
|
2123
|
-
}[];
|
|
2124
|
-
runtimeTarget: {
|
|
2125
|
-
adapterId: string;
|
|
2126
|
-
placement: "local" | "remote" | "colocated" | "provider_native";
|
|
2127
|
-
runtimeVersion: string;
|
|
2128
|
-
capabilityReceipt: string;
|
|
2129
|
-
};
|
|
2130
|
-
limits: {
|
|
2131
|
-
maxTurns: number;
|
|
2132
|
-
timeoutMs: number;
|
|
2133
|
-
maxOutputBytes: number;
|
|
2134
|
-
maximumSpendUsd: number | null;
|
|
2135
|
-
};
|
|
2136
|
-
createdAt: string;
|
|
2137
|
-
metadata: Record<string, unknown>;
|
|
2138
|
-
contentHash: string;
|
|
2139
2121
|
profileEvaluation?: {
|
|
2140
2122
|
profileId: string;
|
|
2141
2123
|
sourceRevision: string;
|
|
@@ -2182,7 +2164,35 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
2182
2164
|
};
|
|
2183
2165
|
declaredProfileCatalogHash: string | null;
|
|
2184
2166
|
protectedProfileClosureHash: string;
|
|
2185
|
-
|
|
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";
|
|
2186
2196
|
population: {
|
|
2187
2197
|
taskId: string;
|
|
2188
2198
|
seed: string;
|
|
@@ -2196,16 +2206,16 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
2196
2206
|
};
|
|
2197
2207
|
feedbackKey: string;
|
|
2198
2208
|
configurationHash: string;
|
|
2199
|
-
output: "boolean" | "
|
|
2209
|
+
output: "boolean" | "category" | "score";
|
|
2200
2210
|
categories: string[];
|
|
2201
2211
|
}[];
|
|
2202
2212
|
fieldMappingsHash: string;
|
|
2203
2213
|
fieldMappings: {
|
|
2204
2214
|
graderId: string;
|
|
2205
2215
|
fields: {
|
|
2206
|
-
destination: "
|
|
2216
|
+
destination: "evaluatorContext" | "expectedOutput" | "input" | "output";
|
|
2207
2217
|
path: string[];
|
|
2208
|
-
source: "
|
|
2218
|
+
source: "evaluatorContext" | "expectedOutput" | "input" | "output";
|
|
2209
2219
|
sourcePath: string[];
|
|
2210
2220
|
}[];
|
|
2211
2221
|
}[];
|
|
@@ -2216,34 +2226,6 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
2216
2226
|
requireComplete: boolean;
|
|
2217
2227
|
};
|
|
2218
2228
|
contentHash: string;
|
|
2219
|
-
recordedOrigin?: {
|
|
2220
|
-
execution: {
|
|
2221
|
-
id: string;
|
|
2222
|
-
contentHash: string;
|
|
2223
|
-
};
|
|
2224
|
-
originalEvidence: {
|
|
2225
|
-
id: string;
|
|
2226
|
-
contentHash: string;
|
|
2227
|
-
};
|
|
2228
|
-
members: {
|
|
2229
|
-
caseId: string;
|
|
2230
|
-
seed: string;
|
|
2231
|
-
fixtureId: string;
|
|
2232
|
-
sourceId: string;
|
|
2233
|
-
snapshotHash: string;
|
|
2234
|
-
boundaryRevisionHash: string;
|
|
2235
|
-
}[];
|
|
2236
|
-
} | undefined;
|
|
2237
|
-
gradingSource?: {
|
|
2238
|
-
pass: {
|
|
2239
|
-
id: string;
|
|
2240
|
-
contentHash: string;
|
|
2241
|
-
};
|
|
2242
|
-
evidence: {
|
|
2243
|
-
id: string;
|
|
2244
|
-
contentHash: string;
|
|
2245
|
-
};
|
|
2246
|
-
} | undefined;
|
|
2247
2229
|
} | undefined;
|
|
2248
2230
|
sourceCandidate?: {
|
|
2249
2231
|
candidateId: string;
|
|
@@ -2256,6 +2238,56 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
2256
2238
|
protectedClosureHash: string;
|
|
2257
2239
|
} | undefined;
|
|
2258
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;
|
|
2259
2291
|
};
|
|
2260
2292
|
policySnapshot: {
|
|
2261
2293
|
modelId: string;
|
|
@@ -2274,7 +2306,7 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
2274
2306
|
taskId: string;
|
|
2275
2307
|
seed: string;
|
|
2276
2308
|
terminal: boolean;
|
|
2277
|
-
failureClass: "
|
|
2309
|
+
failureClass: "cancelled" | "environment_failure" | "grader_failure" | "infrastructure_failure" | "policy_failure" | "timeout" | null;
|
|
2278
2310
|
outputHash: string | null;
|
|
2279
2311
|
traceHash: string;
|
|
2280
2312
|
artifactRefs: {
|
|
@@ -2311,7 +2343,8 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
2311
2343
|
schemaVersion: "openpond.tasksetMetricPolicy.v1";
|
|
2312
2344
|
primaryMetric: string;
|
|
2313
2345
|
aggregation: "custom" | "mean_score" | "pass_rate" | "weighted_mean";
|
|
2314
|
-
missingReward: "
|
|
2346
|
+
missingReward: "exclude" | "zero";
|
|
2347
|
+
taskWeights?: Record<string, number> | undefined;
|
|
2315
2348
|
customAggregator: {
|
|
2316
2349
|
module: string;
|
|
2317
2350
|
exportName: string;
|
|
@@ -2319,7 +2352,6 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
2319
2352
|
timeoutMs: number;
|
|
2320
2353
|
networkPolicy: "none";
|
|
2321
2354
|
} | null;
|
|
2322
|
-
taskWeights?: Record<string, number> | undefined;
|
|
2323
2355
|
};
|
|
2324
2356
|
policyHash: string;
|
|
2325
2357
|
receiptRefs: {
|
|
@@ -2343,15 +2375,22 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
2343
2375
|
revision: number;
|
|
2344
2376
|
teamId: string;
|
|
2345
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;
|
|
2346
2385
|
operationId: string;
|
|
2347
2386
|
taskset: {
|
|
2348
2387
|
id: string;
|
|
2349
2388
|
contentHash: string;
|
|
2350
2389
|
revision: number;
|
|
2351
2390
|
};
|
|
2352
|
-
policyKind: "
|
|
2391
|
+
policyKind: "fixture" | "hosted_chat" | "hosted_harness";
|
|
2353
2392
|
manifestHash: string;
|
|
2354
|
-
status: "cancelled" | "
|
|
2393
|
+
status: "cancelled" | "cancelling" | "completed" | "failed" | "queued" | "running";
|
|
2355
2394
|
totalCount: number;
|
|
2356
2395
|
counts: {
|
|
2357
2396
|
pending: number;
|
|
@@ -2371,13 +2410,6 @@ export declare class OpenPondModelTasksetRunsClient {
|
|
|
2371
2410
|
code: string;
|
|
2372
2411
|
message: string;
|
|
2373
2412
|
} | null;
|
|
2374
|
-
name?: string | undefined;
|
|
2375
|
-
project?: {
|
|
2376
|
-
id: string;
|
|
2377
|
-
revision: number;
|
|
2378
|
-
contentHash: string;
|
|
2379
|
-
targetId: string | null;
|
|
2380
|
-
} | undefined;
|
|
2381
2413
|
}[];
|
|
2382
2414
|
nextCursor: string | null;
|
|
2383
2415
|
}>;
|