openpond-sdk 0.12.3 → 0.12.4
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/dist/actions-local.js +2 -1
- package/dist/actions-local.js.map +1 -7
- package/dist/actions.js +2 -1
- package/dist/actions.js.map +1 -7
- package/dist/advanced-refiner-evaluations.js +13 -12
- package/dist/advanced-refiner-evaluations.js.map +1 -7
- package/dist/candidate-evaluations.js +9 -8
- package/dist/candidate-evaluations.js.map +1 -7
- package/dist/{chunk-VY5AFE7A.js → chunk-24N6A6SF.js} +4 -4
- package/dist/chunk-24N6A6SF.js.map +1 -0
- package/dist/{chunk-YCQVVFE5.js → chunk-3B5KSH3V.js} +2 -2
- package/dist/chunk-3B5KSH3V.js.map +1 -0
- package/dist/{chunk-2DMYAJLS.js → chunk-3DWPLYFJ.js} +2 -2
- package/dist/chunk-3DWPLYFJ.js.map +1 -0
- package/dist/{chunk-CTVKI3BV.js → chunk-53LLNDXF.js} +3 -3
- package/dist/chunk-53LLNDXF.js.map +1 -0
- package/dist/{chunk-36WJTS5V.js → chunk-B5XTGEBE.js} +25 -13
- package/dist/chunk-B5XTGEBE.js.map +1 -0
- package/dist/{chunk-HKWX624G.js → chunk-BVMR5OFW.js} +5 -5
- package/dist/chunk-BVMR5OFW.js.map +1 -0
- package/dist/{chunk-3KXDH6OB.js → chunk-CL3IOOVL.js} +1 -1
- package/dist/chunk-CL3IOOVL.js.map +1 -0
- package/dist/{chunk-FY7NHCEH.js → chunk-CMWZZZGW.js} +5 -5
- package/dist/chunk-CMWZZZGW.js.map +1 -0
- package/dist/{chunk-UPS52OPZ.js → chunk-DF2QXVZ3.js} +8 -8
- package/dist/chunk-DF2QXVZ3.js.map +1 -0
- package/dist/{chunk-AG5DDL7T.js → chunk-DH6O4JFB.js} +2 -2
- package/dist/chunk-DH6O4JFB.js.map +1 -0
- package/dist/{chunk-QEIT5VXJ.js → chunk-EFUT3XKW.js} +4 -4
- package/dist/chunk-EFUT3XKW.js.map +1 -0
- package/dist/{chunk-BSK3VLM3.js → chunk-ETX25FTB.js} +3 -3
- package/dist/chunk-ETX25FTB.js.map +1 -0
- package/dist/{chunk-MEKWESBF.js → chunk-FBZTDDRC.js} +2 -2
- package/dist/chunk-FBZTDDRC.js.map +1 -0
- package/dist/{chunk-CWXM57QR.js → chunk-GAIDBPFY.js} +5 -5
- package/dist/chunk-GAIDBPFY.js.map +1 -0
- package/dist/{chunk-YVUIHH6Z.js → chunk-GPQLBVQC.js} +1 -1
- package/dist/chunk-GPQLBVQC.js.map +1 -0
- package/dist/{chunk-KAE7AMYE.js → chunk-H3367ISL.js} +2 -2
- package/dist/chunk-H3367ISL.js.map +1 -0
- package/dist/{chunk-COPVBJPL.js → chunk-H3JZ3P5T.js} +2 -1
- package/dist/chunk-H3JZ3P5T.js.map +1 -0
- package/dist/{chunk-3LP7DNLQ.js → chunk-H4NTJFEE.js} +2 -2
- package/dist/chunk-H4NTJFEE.js.map +1 -0
- package/dist/{chunk-TCHAWJMB.js → chunk-HGXVOFCL.js} +2 -2
- package/dist/chunk-HGXVOFCL.js.map +1 -0
- package/dist/{chunk-CEZEVILU.js → chunk-HJ2AIAKX.js} +1 -1
- package/dist/chunk-HJ2AIAKX.js.map +1 -0
- package/dist/{chunk-K5WRVSOL.js → chunk-I54WQBXL.js} +8 -8
- package/dist/chunk-I54WQBXL.js.map +1 -0
- package/dist/{chunk-G6Y6CQYV.js → chunk-K5LWSWV7.js} +4 -4
- package/dist/chunk-K5LWSWV7.js.map +1 -0
- package/dist/{chunk-IWT6JUZD.js → chunk-L6EJX2VQ.js} +2 -2
- package/dist/chunk-L6EJX2VQ.js.map +1 -0
- package/dist/{chunk-XYQD4LDI.js → chunk-LHX7ENKZ.js} +3 -3
- package/dist/chunk-LHX7ENKZ.js.map +1 -0
- package/dist/{chunk-BFR3VC25.js → chunk-LR5CTOC5.js} +18 -18
- package/dist/chunk-LR5CTOC5.js.map +1 -0
- package/dist/chunk-MO4HAJZG.js +8709 -0
- package/dist/chunk-MO4HAJZG.js.map +1 -0
- package/dist/{chunk-T4APHDES.js → chunk-O6FKHAWN.js} +1 -1
- package/dist/chunk-O6FKHAWN.js.map +1 -0
- package/dist/{chunk-77N3QEIL.js → chunk-OR3TMIJN.js} +5 -5
- package/dist/chunk-OR3TMIJN.js.map +1 -0
- package/dist/{chunk-P27DQ65Z.js → chunk-OVGDDKO2.js} +1 -1
- package/dist/chunk-OVGDDKO2.js.map +1 -0
- package/dist/{chunk-O4GE6VTR.js → chunk-PWW5QKZS.js} +2 -2
- package/dist/chunk-PWW5QKZS.js.map +1 -0
- package/dist/{chunk-DUKW3R7O.js → chunk-Q4TKI2MV.js} +2 -2
- package/dist/chunk-Q4TKI2MV.js.map +1 -0
- package/dist/{chunk-QTJQIHDZ.js → chunk-QRS7Y3EM.js} +4 -3
- package/dist/chunk-QRS7Y3EM.js.map +1 -0
- package/dist/{chunk-TMUAKJWX.js → chunk-RRKRA237.js} +14 -5
- package/dist/chunk-RRKRA237.js.map +1 -0
- package/dist/{chunk-2U54EOXJ.js → chunk-TK3DXH7A.js} +1 -1
- package/dist/chunk-TK3DXH7A.js.map +1 -0
- package/dist/{chunk-CK3ONRRC.js → chunk-U7XWIOSB.js} +3 -3
- package/dist/chunk-U7XWIOSB.js.map +1 -0
- package/dist/{chunk-YUKV7RCD.js → chunk-VHGC44HA.js} +5 -5
- package/dist/chunk-VHGC44HA.js.map +1 -0
- package/dist/{chunk-ZJ3H3VS7.js → chunk-VN5V2E2S.js} +7 -5
- package/dist/chunk-VN5V2E2S.js.map +1 -0
- package/dist/{chunk-G53MA5XX.js → chunk-WMIKJGXW.js} +3 -3
- package/dist/chunk-WMIKJGXW.js.map +1 -0
- package/dist/{chunk-ROBAX5TN.js → chunk-WQFU635Q.js} +8 -6
- package/dist/chunk-WQFU635Q.js.map +1 -0
- package/dist/{chunk-TSMGMA6R.js → chunk-XON75UFN.js} +3 -3
- package/dist/chunk-XON75UFN.js.map +1 -0
- package/dist/chunk-YLJ4XMA6.js +35 -0
- package/dist/chunk-YLJ4XMA6.js.map +1 -0
- package/dist/{chunk-K3JCP4FO.js → chunk-Z252BNAZ.js} +1 -1
- package/dist/chunk-Z252BNAZ.js.map +1 -0
- package/dist/{chunk-MLPPSCDQ.js → chunk-ZHQK3MOM.js} +4 -4
- package/dist/chunk-ZHQK3MOM.js.map +1 -0
- package/dist/{chunk-6Q4I33CH.js → chunk-ZIFZFP5I.js} +5 -5
- package/dist/chunk-ZIFZFP5I.js.map +1 -0
- package/dist/connected-evidence.js +22 -21
- package/dist/connected-evidence.js.map +1 -7
- package/dist/dataset-commerce.js +2 -0
- package/dist/dataset-commerce.js.map +1 -7
- package/dist/dataset-marketplace.js +2 -0
- package/dist/dataset-marketplace.js.map +1 -7
- package/dist/dataset-workspaces.js +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 +98 -50
- package/dist/index.js.map +1 -7
- package/dist/learning.js +4 -3
- package/dist/learning.js.map +1 -7
- package/dist/model-batch-review.js +2 -1
- package/dist/model-batch-review.js.map +1 -7
- package/dist/model-projects.js +3 -2
- package/dist/model-projects.js.map +1 -7
- package/dist/model-starter-attempts.js +5 -4
- package/dist/model-starter-attempts.js.map +1 -7
- package/dist/model-starter-catalog.js +15 -14
- package/dist/model-starter-catalog.js.map +1 -7
- package/dist/model-starters.js +15 -14
- package/dist/model-starters.js.map +1 -7
- package/dist/model-taskset-authoring.js +2 -1
- package/dist/model-taskset-authoring.js.map +1 -7
- package/dist/model-taskset-runs.js +6 -5
- package/dist/model-taskset-runs.js.map +1 -7
- package/dist/post-training.js +9 -8
- package/dist/post-training.js.map +1 -7
- package/dist/profile-actions.js +3 -2
- package/dist/profile-actions.js.map +1 -7
- package/dist/project-actions.js +3 -2
- package/dist/project-actions.js.map +1 -7
- package/dist/refiner.js +3 -2
- package/dist/refiner.js.map +1 -7
- package/dist/taskset-catalog.js +2 -1
- package/dist/taskset-catalog.js.map +1 -7
- package/dist/taskset-drafts.js +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/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-learning-intent.d.ts +1 -1
- package/dist/types/packages/sdk/src/model-learning-policy.d.ts +247 -247
- package/dist/types/packages/sdk/src/model-learning-policy.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/model-projects.d.ts +324 -324
- package/dist/types/packages/sdk/src/model-projects.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/model-starter-attempts.d.ts +187 -187
- package/dist/types/packages/sdk/src/model-starter-attempts.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/model-starter-catalog.d.ts +181 -179
- package/dist/types/packages/sdk/src/model-starter-catalog.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/model-starter-execution.d.ts +40 -40
- package/dist/types/packages/sdk/src/model-starters.d.ts +291 -287
- package/dist/types/packages/sdk/src/model-starters.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/model-taskset-authoring-contracts.d.ts +3 -3
- package/dist/types/packages/sdk/src/model-taskset-derivation.d.ts +114 -112
- package/dist/types/packages/sdk/src/model-taskset-derivation.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/model-taskset-resources.d.ts +18 -18
- package/dist/types/packages/sdk/src/model-taskset-runs-contracts.d.ts +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 +438 -377
- package/dist/types/packages/sdk/src/post-training.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/profile-actions.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/profile-evaluation-discovery.d.ts +150 -150
- package/dist/types/packages/sdk/src/profile-evaluation-discovery.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/profile-external-dataset-package.d.ts +254 -251
- package/dist/types/packages/sdk/src/profile-external-dataset-package.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/project-actions.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/protocol.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/run-diagnostics.d.ts +3 -3
- package/dist/types/packages/sdk/src/task-design-contracts.d.ts +50 -50
- package/dist/types/packages/sdk/src/task-inventory.d.ts +4 -4
- package/dist/types/packages/sdk/src/task-inventory.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/taskset-authored-contracts.d.ts +190 -189
- package/dist/types/packages/sdk/src/taskset-authored-contracts.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/taskset-authored-files.d.ts +149 -148
- package/dist/types/packages/sdk/src/taskset-authored-files.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/taskset-authored-portable-release.d.ts +15 -6
- package/dist/types/packages/sdk/src/taskset-authored-portable-release.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/taskset-calibration-fixtures.d.ts +64 -63
- package/dist/types/packages/sdk/src/taskset-calibration-fixtures.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/taskset-catalog.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/taskset-draft-api-contracts.d.ts +245 -243
- package/dist/types/packages/sdk/src/taskset-draft-api-contracts.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/taskset-draft-authoring.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/taskset-draft-client.d.ts +744 -738
- package/dist/types/packages/sdk/src/taskset-draft-client.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/taskset-draft-core.d.ts +133 -131
- package/dist/types/packages/sdk/src/taskset-draft-core.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/taskset-draft-dataset-artifacts.d.ts +80 -80
- package/dist/types/packages/sdk/src/taskset-draft-dataset-sources.d.ts +34 -34
- package/dist/types/packages/sdk/src/taskset-draft-document.d.ts +119 -117
- package/dist/types/packages/sdk/src/taskset-draft-document.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/taskset-draft-package-compiler.d.ts +500 -492
- package/dist/types/packages/sdk/src/taskset-draft-package-compiler.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/taskset-draft-source.d.ts +236 -234
- package/dist/types/packages/sdk/src/taskset-draft-source.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/taskset-draft-workspace.d.ts +228 -224
- package/dist/types/packages/sdk/src/taskset-draft-workspace.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/taskset-intake.d.ts +1 -1
- package/dist/types/packages/sdk/src/taskset-package-client.d.ts +4589 -4577
- package/dist/types/packages/sdk/src/taskset-package-client.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/taskset-package-contracts.d.ts +2068 -2054
- package/dist/types/packages/sdk/src/taskset-package-contracts.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/taskset-package-files.d.ts +1 -1
- package/dist/types/packages/sdk/src/taskset-package-grader-projection.d.ts +21 -21
- package/dist/types/packages/sdk/src/taskset-package-learning.d.ts +88 -86
- package/dist/types/packages/sdk/src/taskset-package-learning.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/taskset-package-projection.d.ts +387 -383
- package/dist/types/packages/sdk/src/taskset-package-projection.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/training-acceptance.d.ts +26 -26
- package/dist/types/packages/sdk/src/training-acceptance.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/training-bundle-contracts.d.ts +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
|
@@ -22,14 +22,18 @@ export declare function prepareImportedTasksetPackage(input: {
|
|
|
22
22
|
};
|
|
23
23
|
environment: {
|
|
24
24
|
protocolVersion: "openpond.environment.v1";
|
|
25
|
-
kind: "
|
|
25
|
+
kind: "agent" | "custom_program" | "text" | "work";
|
|
26
26
|
entrypoint: string;
|
|
27
27
|
stateful: boolean;
|
|
28
28
|
deterministicSeeds: boolean;
|
|
29
|
-
lifecycle: ("
|
|
30
|
-
networkPolicy: "
|
|
29
|
+
lifecycle: ("collect" | "create" | "destroy" | "reset" | "step")[];
|
|
30
|
+
networkPolicy: "declared_read_only" | "declared_scoped" | "none";
|
|
31
31
|
defaultTimeoutMs: number;
|
|
32
32
|
};
|
|
33
|
+
environmentRelease?: {
|
|
34
|
+
id: string;
|
|
35
|
+
contentHash: string;
|
|
36
|
+
} | undefined;
|
|
33
37
|
tools: {
|
|
34
38
|
name: string;
|
|
35
39
|
description: string;
|
|
@@ -42,12 +46,12 @@ export declare function prepareImportedTasksetPackage(input: {
|
|
|
42
46
|
id: string;
|
|
43
47
|
required: boolean;
|
|
44
48
|
scopes: string[];
|
|
45
|
-
portability: "
|
|
49
|
+
portability: "host_adapter" | "hosted_only" | "local_only" | "portable";
|
|
46
50
|
}[];
|
|
47
51
|
tasks: {
|
|
48
52
|
id: string;
|
|
49
53
|
clusterKey: string;
|
|
50
|
-
split: "
|
|
54
|
+
split: "frozen_eval" | "test" | "train" | "validation";
|
|
51
55
|
input: Record<string, unknown>;
|
|
52
56
|
expectedOutput: Record<string, unknown> | null;
|
|
53
57
|
policyVisibleContext: Record<string, unknown>;
|
|
@@ -58,9 +62,8 @@ export declare function prepareImportedTasksetPackage(input: {
|
|
|
58
62
|
contentHash: string;
|
|
59
63
|
sizeBytes: number;
|
|
60
64
|
mediaType: string;
|
|
61
|
-
visibility: "
|
|
65
|
+
visibility: "host_private" | "policy" | "verifier";
|
|
62
66
|
}[];
|
|
63
|
-
tags: string[];
|
|
64
67
|
requiredOutputs?: {
|
|
65
68
|
path: string;
|
|
66
69
|
mediaType: string;
|
|
@@ -70,24 +73,27 @@ export declare function prepareImportedTasksetPackage(input: {
|
|
|
70
73
|
contentHash: string;
|
|
71
74
|
sizeBytes: number;
|
|
72
75
|
mediaType: string;
|
|
73
|
-
visibility: "
|
|
76
|
+
visibility: "host_private" | "policy" | "verifier";
|
|
74
77
|
} | null;
|
|
78
|
+
submissionMode?: "structured_json" | undefined;
|
|
75
79
|
maxBytes: number | null;
|
|
76
80
|
metadata: Record<string, unknown>;
|
|
77
81
|
}[] | undefined;
|
|
82
|
+
tags: string[];
|
|
78
83
|
}[];
|
|
79
84
|
graders: ({
|
|
80
85
|
id: string;
|
|
86
|
+
taskIds?: string[] | undefined;
|
|
81
87
|
version: string;
|
|
82
88
|
weight: number;
|
|
83
89
|
hardGate: boolean;
|
|
84
90
|
rewardEligible: boolean;
|
|
85
91
|
privileged: boolean;
|
|
86
|
-
kind: "
|
|
92
|
+
kind: "artifact" | "content" | "runtime_event" | "schema" | "state";
|
|
87
93
|
config: Record<string, unknown>;
|
|
88
|
-
taskIds?: string[] | undefined;
|
|
89
94
|
} | {
|
|
90
95
|
id: string;
|
|
96
|
+
taskIds?: string[] | undefined;
|
|
91
97
|
version: string;
|
|
92
98
|
weight: number;
|
|
93
99
|
hardGate: boolean;
|
|
@@ -100,10 +106,9 @@ export declare function prepareImportedTasksetPackage(input: {
|
|
|
100
106
|
contentHash: string;
|
|
101
107
|
sizeBytes: number;
|
|
102
108
|
mediaType: string;
|
|
103
|
-
visibility: "
|
|
109
|
+
visibility: "host_private" | "policy" | "verifier";
|
|
104
110
|
};
|
|
105
|
-
calibrationStatus: "
|
|
106
|
-
taskIds?: string[] | undefined;
|
|
111
|
+
calibrationStatus: "failed" | "passed" | "pending";
|
|
107
112
|
model?: {
|
|
108
113
|
providerId: string;
|
|
109
114
|
modelId: string;
|
|
@@ -112,27 +117,28 @@ export declare function prepareImportedTasksetPackage(input: {
|
|
|
112
117
|
temperature?: number | undefined;
|
|
113
118
|
} | {
|
|
114
119
|
id: string;
|
|
120
|
+
taskIds?: string[] | undefined;
|
|
115
121
|
version: string;
|
|
116
122
|
weight: number;
|
|
117
123
|
hardGate: boolean;
|
|
118
124
|
rewardEligible: boolean;
|
|
119
125
|
privileged: boolean;
|
|
120
126
|
kind: "custom_verifier";
|
|
127
|
+
runtime?: "isolated_javascript" | "sandbox_process" | undefined;
|
|
121
128
|
verifierRef: {
|
|
122
129
|
id: string;
|
|
123
130
|
path: string;
|
|
124
131
|
contentHash: string;
|
|
125
132
|
sizeBytes: number;
|
|
126
133
|
mediaType: string;
|
|
127
|
-
visibility: "
|
|
134
|
+
visibility: "host_private" | "policy" | "verifier";
|
|
128
135
|
};
|
|
136
|
+
exportName?: string | undefined;
|
|
129
137
|
timeoutMs: number;
|
|
130
138
|
networkPolicy: "none";
|
|
131
|
-
taskIds?: string[] | undefined;
|
|
132
|
-
runtime?: "isolated_javascript" | "sandbox_process" | undefined;
|
|
133
|
-
exportName?: string | undefined;
|
|
134
139
|
} | {
|
|
135
140
|
id: string;
|
|
141
|
+
taskIds?: string[] | undefined;
|
|
136
142
|
version: string;
|
|
137
143
|
weight: number;
|
|
138
144
|
hardGate: boolean;
|
|
@@ -145,10 +151,9 @@ export declare function prepareImportedTasksetPackage(input: {
|
|
|
145
151
|
contentHash: string;
|
|
146
152
|
sizeBytes: number;
|
|
147
153
|
mediaType: string;
|
|
148
|
-
visibility: "
|
|
154
|
+
visibility: "host_private" | "policy" | "verifier";
|
|
149
155
|
};
|
|
150
156
|
reviewerRole: string;
|
|
151
|
-
taskIds?: string[] | undefined;
|
|
152
157
|
form?: {
|
|
153
158
|
schemaVersion: "openpond.humanForm.v1";
|
|
154
159
|
mode: "individual" | "pairwise";
|
|
@@ -159,77 +164,71 @@ export declare function prepareImportedTasksetPackage(input: {
|
|
|
159
164
|
instructions: string;
|
|
160
165
|
required: boolean;
|
|
161
166
|
allowAbstain: boolean;
|
|
162
|
-
kind: "score";
|
|
163
|
-
minimum: number;
|
|
164
|
-
maximum: number;
|
|
165
|
-
step: number;
|
|
166
167
|
requireRationale?: boolean | undefined;
|
|
167
168
|
anchors?: {
|
|
168
169
|
value: string | number | boolean;
|
|
169
170
|
description: string;
|
|
170
171
|
}[] | undefined;
|
|
172
|
+
kind: "score";
|
|
173
|
+
minimum: number;
|
|
174
|
+
maximum: number;
|
|
175
|
+
step: number;
|
|
171
176
|
} | {
|
|
172
177
|
id: string;
|
|
173
178
|
label: string;
|
|
174
179
|
instructions: string;
|
|
175
180
|
required: boolean;
|
|
176
181
|
allowAbstain: boolean;
|
|
177
|
-
kind: "category";
|
|
178
|
-
options: string[];
|
|
179
182
|
requireRationale?: boolean | undefined;
|
|
180
183
|
anchors?: {
|
|
181
184
|
value: string | number | boolean;
|
|
182
185
|
description: string;
|
|
183
186
|
}[] | undefined;
|
|
187
|
+
kind: "category";
|
|
188
|
+
options: string[];
|
|
184
189
|
} | {
|
|
185
190
|
id: string;
|
|
186
191
|
label: string;
|
|
187
192
|
instructions: string;
|
|
188
193
|
required: boolean;
|
|
189
194
|
allowAbstain: boolean;
|
|
190
|
-
kind: "boolean";
|
|
191
195
|
requireRationale?: boolean | undefined;
|
|
192
196
|
anchors?: {
|
|
193
197
|
value: string | number | boolean;
|
|
194
198
|
description: string;
|
|
195
199
|
}[] | undefined;
|
|
200
|
+
kind: "boolean";
|
|
196
201
|
} | {
|
|
197
202
|
id: string;
|
|
198
203
|
label: string;
|
|
199
204
|
instructions: string;
|
|
200
205
|
required: boolean;
|
|
201
206
|
allowAbstain: boolean;
|
|
202
|
-
kind: "text";
|
|
203
|
-
maximumLength: number;
|
|
204
207
|
requireRationale?: boolean | undefined;
|
|
205
208
|
anchors?: {
|
|
206
209
|
value: string | number | boolean;
|
|
207
210
|
description: string;
|
|
208
211
|
}[] | undefined;
|
|
212
|
+
kind: "text";
|
|
209
213
|
minimumLength?: number | undefined;
|
|
214
|
+
maximumLength: number;
|
|
210
215
|
} | {
|
|
211
216
|
id: string;
|
|
212
217
|
label: string;
|
|
213
218
|
instructions: string;
|
|
214
219
|
required: boolean;
|
|
215
220
|
allowAbstain: boolean;
|
|
216
|
-
kind: "preference";
|
|
217
|
-
allowTie: boolean;
|
|
218
221
|
requireRationale?: boolean | undefined;
|
|
219
222
|
anchors?: {
|
|
220
223
|
value: string | number | boolean;
|
|
221
224
|
description: string;
|
|
222
225
|
}[] | undefined;
|
|
226
|
+
kind: "preference";
|
|
223
227
|
requireStrength?: boolean | undefined;
|
|
228
|
+
allowTie: boolean;
|
|
224
229
|
})[];
|
|
225
230
|
} | undefined;
|
|
226
231
|
})[];
|
|
227
|
-
metadata: Record<string, unknown>;
|
|
228
|
-
contentHash: string;
|
|
229
|
-
environmentRelease?: {
|
|
230
|
-
id: string;
|
|
231
|
-
contentHash: string;
|
|
232
|
-
} | undefined;
|
|
233
232
|
verifierSetRelease?: {
|
|
234
233
|
id: string;
|
|
235
234
|
contentHash: string;
|
|
@@ -238,7 +237,8 @@ export declare function prepareImportedTasksetPackage(input: {
|
|
|
238
237
|
schemaVersion: "openpond.tasksetMetricPolicy.v1";
|
|
239
238
|
primaryMetric: string;
|
|
240
239
|
aggregation: "custom" | "mean_score" | "pass_rate" | "weighted_mean";
|
|
241
|
-
missingReward: "
|
|
240
|
+
missingReward: "exclude" | "zero";
|
|
241
|
+
taskWeights?: Record<string, number> | undefined;
|
|
242
242
|
customAggregator: {
|
|
243
243
|
module: string;
|
|
244
244
|
exportName: string;
|
|
@@ -246,8 +246,9 @@ export declare function prepareImportedTasksetPackage(input: {
|
|
|
246
246
|
timeoutMs: number;
|
|
247
247
|
networkPolicy: "none";
|
|
248
248
|
} | null;
|
|
249
|
-
taskWeights?: Record<string, number> | undefined;
|
|
250
249
|
} | undefined;
|
|
250
|
+
metadata: Record<string, unknown>;
|
|
251
|
+
contentHash: string;
|
|
251
252
|
};
|
|
252
253
|
environment: {
|
|
253
254
|
schemaVersion: "openpond.environmentRelease.v1";
|
|
@@ -255,12 +256,12 @@ export declare function prepareImportedTasksetPackage(input: {
|
|
|
255
256
|
revision: number;
|
|
256
257
|
contract: {
|
|
257
258
|
protocolVersion: "openpond.environment.v1";
|
|
258
|
-
kind: "
|
|
259
|
+
kind: "agent" | "custom_program" | "text" | "work";
|
|
259
260
|
entrypoint: string;
|
|
260
261
|
stateful: boolean;
|
|
261
262
|
deterministicSeeds: boolean;
|
|
262
|
-
lifecycle: ("
|
|
263
|
-
networkPolicy: "
|
|
263
|
+
lifecycle: ("collect" | "create" | "destroy" | "reset" | "step")[];
|
|
264
|
+
networkPolicy: "declared_read_only" | "declared_scoped" | "none";
|
|
264
265
|
defaultTimeoutMs: number;
|
|
265
266
|
};
|
|
266
267
|
actionSchemaRef: {
|
|
@@ -269,7 +270,7 @@ export declare function prepareImportedTasksetPackage(input: {
|
|
|
269
270
|
contentHash: string;
|
|
270
271
|
sizeBytes: number;
|
|
271
272
|
mediaType: string;
|
|
272
|
-
visibility: "
|
|
273
|
+
visibility: "host_private" | "policy" | "verifier";
|
|
273
274
|
} | null;
|
|
274
275
|
observationSchemaRef: {
|
|
275
276
|
id: string;
|
|
@@ -277,7 +278,7 @@ export declare function prepareImportedTasksetPackage(input: {
|
|
|
277
278
|
contentHash: string;
|
|
278
279
|
sizeBytes: number;
|
|
279
280
|
mediaType: string;
|
|
280
|
-
visibility: "
|
|
281
|
+
visibility: "host_private" | "policy" | "verifier";
|
|
281
282
|
} | null;
|
|
282
283
|
stateSchemaRef: {
|
|
283
284
|
id: string;
|
|
@@ -285,7 +286,7 @@ export declare function prepareImportedTasksetPackage(input: {
|
|
|
285
286
|
contentHash: string;
|
|
286
287
|
sizeBytes: number;
|
|
287
288
|
mediaType: string;
|
|
288
|
-
visibility: "
|
|
289
|
+
visibility: "host_private" | "policy" | "verifier";
|
|
289
290
|
} | null;
|
|
290
291
|
artifactCollection: {
|
|
291
292
|
maxArtifacts: number;
|
|
@@ -301,16 +302,17 @@ export declare function prepareImportedTasksetPackage(input: {
|
|
|
301
302
|
revision: number;
|
|
302
303
|
graders: ({
|
|
303
304
|
id: string;
|
|
305
|
+
taskIds?: string[] | undefined;
|
|
304
306
|
version: string;
|
|
305
307
|
weight: number;
|
|
306
308
|
hardGate: boolean;
|
|
307
309
|
rewardEligible: boolean;
|
|
308
310
|
privileged: boolean;
|
|
309
|
-
kind: "
|
|
311
|
+
kind: "artifact" | "content" | "runtime_event" | "schema" | "state";
|
|
310
312
|
config: Record<string, unknown>;
|
|
311
|
-
taskIds?: string[] | undefined;
|
|
312
313
|
} | {
|
|
313
314
|
id: string;
|
|
315
|
+
taskIds?: string[] | undefined;
|
|
314
316
|
version: string;
|
|
315
317
|
weight: number;
|
|
316
318
|
hardGate: boolean;
|
|
@@ -323,10 +325,9 @@ export declare function prepareImportedTasksetPackage(input: {
|
|
|
323
325
|
contentHash: string;
|
|
324
326
|
sizeBytes: number;
|
|
325
327
|
mediaType: string;
|
|
326
|
-
visibility: "
|
|
328
|
+
visibility: "host_private" | "policy" | "verifier";
|
|
327
329
|
};
|
|
328
|
-
calibrationStatus: "
|
|
329
|
-
taskIds?: string[] | undefined;
|
|
330
|
+
calibrationStatus: "failed" | "passed" | "pending";
|
|
330
331
|
model?: {
|
|
331
332
|
providerId: string;
|
|
332
333
|
modelId: string;
|
|
@@ -335,27 +336,28 @@ export declare function prepareImportedTasksetPackage(input: {
|
|
|
335
336
|
temperature?: number | undefined;
|
|
336
337
|
} | {
|
|
337
338
|
id: string;
|
|
339
|
+
taskIds?: string[] | undefined;
|
|
338
340
|
version: string;
|
|
339
341
|
weight: number;
|
|
340
342
|
hardGate: boolean;
|
|
341
343
|
rewardEligible: boolean;
|
|
342
344
|
privileged: boolean;
|
|
343
345
|
kind: "custom_verifier";
|
|
346
|
+
runtime?: "isolated_javascript" | "sandbox_process" | undefined;
|
|
344
347
|
verifierRef: {
|
|
345
348
|
id: string;
|
|
346
349
|
path: string;
|
|
347
350
|
contentHash: string;
|
|
348
351
|
sizeBytes: number;
|
|
349
352
|
mediaType: string;
|
|
350
|
-
visibility: "
|
|
353
|
+
visibility: "host_private" | "policy" | "verifier";
|
|
351
354
|
};
|
|
355
|
+
exportName?: string | undefined;
|
|
352
356
|
timeoutMs: number;
|
|
353
357
|
networkPolicy: "none";
|
|
354
|
-
taskIds?: string[] | undefined;
|
|
355
|
-
runtime?: "isolated_javascript" | "sandbox_process" | undefined;
|
|
356
|
-
exportName?: string | undefined;
|
|
357
358
|
} | {
|
|
358
359
|
id: string;
|
|
360
|
+
taskIds?: string[] | undefined;
|
|
359
361
|
version: string;
|
|
360
362
|
weight: number;
|
|
361
363
|
hardGate: boolean;
|
|
@@ -368,10 +370,9 @@ export declare function prepareImportedTasksetPackage(input: {
|
|
|
368
370
|
contentHash: string;
|
|
369
371
|
sizeBytes: number;
|
|
370
372
|
mediaType: string;
|
|
371
|
-
visibility: "
|
|
373
|
+
visibility: "host_private" | "policy" | "verifier";
|
|
372
374
|
};
|
|
373
375
|
reviewerRole: string;
|
|
374
|
-
taskIds?: string[] | undefined;
|
|
375
376
|
form?: {
|
|
376
377
|
schemaVersion: "openpond.humanForm.v1";
|
|
377
378
|
mode: "individual" | "pairwise";
|
|
@@ -382,73 +383,73 @@ export declare function prepareImportedTasksetPackage(input: {
|
|
|
382
383
|
instructions: string;
|
|
383
384
|
required: boolean;
|
|
384
385
|
allowAbstain: boolean;
|
|
385
|
-
kind: "score";
|
|
386
|
-
minimum: number;
|
|
387
|
-
maximum: number;
|
|
388
|
-
step: number;
|
|
389
386
|
requireRationale?: boolean | undefined;
|
|
390
387
|
anchors?: {
|
|
391
388
|
value: string | number | boolean;
|
|
392
389
|
description: string;
|
|
393
390
|
}[] | undefined;
|
|
391
|
+
kind: "score";
|
|
392
|
+
minimum: number;
|
|
393
|
+
maximum: number;
|
|
394
|
+
step: number;
|
|
394
395
|
} | {
|
|
395
396
|
id: string;
|
|
396
397
|
label: string;
|
|
397
398
|
instructions: string;
|
|
398
399
|
required: boolean;
|
|
399
400
|
allowAbstain: boolean;
|
|
400
|
-
kind: "category";
|
|
401
|
-
options: string[];
|
|
402
401
|
requireRationale?: boolean | undefined;
|
|
403
402
|
anchors?: {
|
|
404
403
|
value: string | number | boolean;
|
|
405
404
|
description: string;
|
|
406
405
|
}[] | undefined;
|
|
406
|
+
kind: "category";
|
|
407
|
+
options: string[];
|
|
407
408
|
} | {
|
|
408
409
|
id: string;
|
|
409
410
|
label: string;
|
|
410
411
|
instructions: string;
|
|
411
412
|
required: boolean;
|
|
412
413
|
allowAbstain: boolean;
|
|
413
|
-
kind: "boolean";
|
|
414
414
|
requireRationale?: boolean | undefined;
|
|
415
415
|
anchors?: {
|
|
416
416
|
value: string | number | boolean;
|
|
417
417
|
description: string;
|
|
418
418
|
}[] | undefined;
|
|
419
|
+
kind: "boolean";
|
|
419
420
|
} | {
|
|
420
421
|
id: string;
|
|
421
422
|
label: string;
|
|
422
423
|
instructions: string;
|
|
423
424
|
required: boolean;
|
|
424
425
|
allowAbstain: boolean;
|
|
425
|
-
kind: "text";
|
|
426
|
-
maximumLength: number;
|
|
427
426
|
requireRationale?: boolean | undefined;
|
|
428
427
|
anchors?: {
|
|
429
428
|
value: string | number | boolean;
|
|
430
429
|
description: string;
|
|
431
430
|
}[] | undefined;
|
|
431
|
+
kind: "text";
|
|
432
432
|
minimumLength?: number | undefined;
|
|
433
|
+
maximumLength: number;
|
|
433
434
|
} | {
|
|
434
435
|
id: string;
|
|
435
436
|
label: string;
|
|
436
437
|
instructions: string;
|
|
437
438
|
required: boolean;
|
|
438
439
|
allowAbstain: boolean;
|
|
439
|
-
kind: "preference";
|
|
440
|
-
allowTie: boolean;
|
|
441
440
|
requireRationale?: boolean | undefined;
|
|
442
441
|
anchors?: {
|
|
443
442
|
value: string | number | boolean;
|
|
444
443
|
description: string;
|
|
445
444
|
}[] | undefined;
|
|
445
|
+
kind: "preference";
|
|
446
446
|
requireStrength?: boolean | undefined;
|
|
447
|
+
allowTie: boolean;
|
|
447
448
|
})[];
|
|
448
449
|
} | undefined;
|
|
449
450
|
})[];
|
|
450
451
|
isolation: {
|
|
451
|
-
processBoundary: "
|
|
452
|
+
processBoundary: "container" | "isolated_process" | "same_process";
|
|
452
453
|
networkPolicy: "none";
|
|
453
454
|
defaultTimeoutMs: number;
|
|
454
455
|
};
|
|
@@ -466,11 +467,10 @@ export declare function prepareImportedTasksetPackage(input: {
|
|
|
466
467
|
contentHash: string;
|
|
467
468
|
sizeBytes: number;
|
|
468
469
|
mediaType: string;
|
|
469
|
-
visibility: "
|
|
470
|
+
visibility: "host_private" | "policy" | "verifier";
|
|
470
471
|
};
|
|
471
472
|
base64: string;
|
|
472
473
|
}[];
|
|
473
|
-
contentHash: string;
|
|
474
474
|
modelResources?: {
|
|
475
475
|
taskDefinition: {
|
|
476
476
|
schemaVersion: "openpond.taskDefinition.v1";
|
|
@@ -479,10 +479,26 @@ export declare function prepareImportedTasksetPackage(input: {
|
|
|
479
479
|
name: string;
|
|
480
480
|
description: string;
|
|
481
481
|
instructions: string;
|
|
482
|
-
category: "
|
|
482
|
+
category: "coding" | "custom" | "question_answering" | "structured" | "tool_workflow";
|
|
483
483
|
familyNamespace: string;
|
|
484
484
|
inputSchema: Record<string, unknown>;
|
|
485
485
|
outputSchema: Record<string, unknown>;
|
|
486
|
+
requiredOutputs?: {
|
|
487
|
+
path: string;
|
|
488
|
+
mediaType: string;
|
|
489
|
+
schemaRef: {
|
|
490
|
+
id: string;
|
|
491
|
+
path: string;
|
|
492
|
+
contentHash: string;
|
|
493
|
+
sizeBytes: number;
|
|
494
|
+
mediaType: string;
|
|
495
|
+
visibility: "host_private" | "policy" | "verifier";
|
|
496
|
+
} | null;
|
|
497
|
+
submissionMode?: "structured_json" | undefined;
|
|
498
|
+
maxBytes: number | null;
|
|
499
|
+
metadata: Record<string, unknown>;
|
|
500
|
+
}[] | undefined;
|
|
501
|
+
contextRequirements?: string[] | undefined;
|
|
486
502
|
rewardBinding: {
|
|
487
503
|
id: string;
|
|
488
504
|
contentHash: string;
|
|
@@ -501,14 +517,18 @@ export declare function prepareImportedTasksetPackage(input: {
|
|
|
501
517
|
};
|
|
502
518
|
environment: {
|
|
503
519
|
protocolVersion: "openpond.environment.v1";
|
|
504
|
-
kind: "
|
|
520
|
+
kind: "agent" | "custom_program" | "text" | "work";
|
|
505
521
|
entrypoint: string;
|
|
506
522
|
stateful: boolean;
|
|
507
523
|
deterministicSeeds: boolean;
|
|
508
|
-
lifecycle: ("
|
|
509
|
-
networkPolicy: "
|
|
524
|
+
lifecycle: ("collect" | "create" | "destroy" | "reset" | "step")[];
|
|
525
|
+
networkPolicy: "declared_read_only" | "declared_scoped" | "none";
|
|
510
526
|
defaultTimeoutMs: number;
|
|
511
527
|
};
|
|
528
|
+
environmentRelease?: {
|
|
529
|
+
id: string;
|
|
530
|
+
contentHash: string;
|
|
531
|
+
} | undefined;
|
|
512
532
|
tools: {
|
|
513
533
|
name: string;
|
|
514
534
|
description: string;
|
|
@@ -521,57 +541,65 @@ export declare function prepareImportedTasksetPackage(input: {
|
|
|
521
541
|
id: string;
|
|
522
542
|
required: boolean;
|
|
523
543
|
scopes: string[];
|
|
524
|
-
portability: "
|
|
544
|
+
portability: "host_adapter" | "hosted_only" | "local_only" | "portable";
|
|
525
545
|
}[];
|
|
526
|
-
environmentRelease?: {
|
|
527
|
-
id: string;
|
|
528
|
-
contentHash: string;
|
|
529
|
-
} | undefined;
|
|
530
546
|
verifierSetRelease?: {
|
|
531
547
|
id: string;
|
|
532
548
|
contentHash: string;
|
|
533
549
|
} | undefined;
|
|
534
550
|
};
|
|
535
551
|
contentHash: string;
|
|
536
|
-
requiredOutputs?: {
|
|
537
|
-
path: string;
|
|
538
|
-
mediaType: string;
|
|
539
|
-
schemaRef: {
|
|
540
|
-
id: string;
|
|
541
|
-
path: string;
|
|
542
|
-
contentHash: string;
|
|
543
|
-
sizeBytes: number;
|
|
544
|
-
mediaType: string;
|
|
545
|
-
visibility: "policy" | "verifier" | "host_private";
|
|
546
|
-
} | null;
|
|
547
|
-
maxBytes: number | null;
|
|
548
|
-
metadata: Record<string, unknown>;
|
|
549
|
-
}[] | undefined;
|
|
550
|
-
contextRequirements?: string[] | undefined;
|
|
551
552
|
};
|
|
552
|
-
|
|
553
|
-
schemaVersion: "openpond.
|
|
553
|
+
rewardBinding: {
|
|
554
|
+
schemaVersion: "openpond.rewardBinding.v1";
|
|
554
555
|
id: string;
|
|
555
|
-
revision:
|
|
556
|
-
|
|
556
|
+
revision: number;
|
|
557
|
+
name?: string | undefined;
|
|
558
|
+
description?: string | undefined;
|
|
559
|
+
recipeRef?: {
|
|
557
560
|
id: string;
|
|
558
|
-
path: string;
|
|
559
561
|
contentHash: string;
|
|
560
|
-
|
|
561
|
-
|
|
562
|
-
|
|
563
|
-
|
|
564
|
-
|
|
562
|
+
revision: number;
|
|
563
|
+
} | undefined;
|
|
564
|
+
sources: {
|
|
565
|
+
graderId: string;
|
|
566
|
+
taskIds?: string[] | undefined;
|
|
567
|
+
reward: {
|
|
568
|
+
id: string;
|
|
569
|
+
contentHash: string;
|
|
570
|
+
revision: number;
|
|
571
|
+
};
|
|
572
|
+
role: "evaluation" | "training";
|
|
573
|
+
normalization: {
|
|
574
|
+
kind: "identity";
|
|
575
|
+
} | {
|
|
576
|
+
kind: "linear";
|
|
577
|
+
minimum: number;
|
|
578
|
+
maximum: number;
|
|
579
|
+
direction: "higher" | "lower";
|
|
580
|
+
};
|
|
581
|
+
weight: number;
|
|
582
|
+
required: boolean;
|
|
583
|
+
hardGate: boolean;
|
|
584
|
+
privileged: boolean;
|
|
585
|
+
fixtureRefs: {
|
|
586
|
+
id: string;
|
|
587
|
+
contentHash: string;
|
|
588
|
+
}[];
|
|
589
|
+
}[];
|
|
590
|
+
aggregation: "weighted_mean";
|
|
591
|
+
unscorable: "exclude_optional_require_all_required";
|
|
565
592
|
contentHash: string;
|
|
566
|
-
}
|
|
593
|
+
};
|
|
567
594
|
rewards: {
|
|
568
595
|
schemaVersion: "openpond.rewardRelease.v1";
|
|
569
596
|
id: string;
|
|
570
597
|
revision: number;
|
|
571
598
|
name: string;
|
|
572
599
|
description: string;
|
|
600
|
+
feedbackKey?: string | undefined;
|
|
573
601
|
implementation: {
|
|
574
|
-
kind: "
|
|
602
|
+
kind: "artifact" | "content" | "runtime_event" | "schema" | "state";
|
|
575
603
|
config: Record<string, unknown>;
|
|
576
604
|
} | {
|
|
577
605
|
kind: "model_judge";
|
|
@@ -581,9 +609,9 @@ export declare function prepareImportedTasksetPackage(input: {
|
|
|
581
609
|
contentHash: string;
|
|
582
610
|
sizeBytes: number;
|
|
583
611
|
mediaType: string;
|
|
584
|
-
visibility: "
|
|
612
|
+
visibility: "host_private" | "policy" | "verifier";
|
|
585
613
|
};
|
|
586
|
-
calibrationStatus: "
|
|
614
|
+
calibrationStatus: "failed" | "passed" | "pending";
|
|
587
615
|
model?: {
|
|
588
616
|
providerId: string;
|
|
589
617
|
modelId: string;
|
|
@@ -591,19 +619,19 @@ export declare function prepareImportedTasksetPackage(input: {
|
|
|
591
619
|
} | undefined;
|
|
592
620
|
temperature?: number | undefined;
|
|
593
621
|
} | {
|
|
594
|
-
timeoutMs: number;
|
|
595
|
-
networkPolicy: "none";
|
|
596
622
|
kind: "custom_verifier";
|
|
623
|
+
runtime?: "isolated_javascript" | "sandbox_process" | undefined;
|
|
597
624
|
verifierRef: {
|
|
598
625
|
id: string;
|
|
599
626
|
path: string;
|
|
600
627
|
contentHash: string;
|
|
601
628
|
sizeBytes: number;
|
|
602
629
|
mediaType: string;
|
|
603
|
-
visibility: "
|
|
630
|
+
visibility: "host_private" | "policy" | "verifier";
|
|
604
631
|
};
|
|
605
632
|
exportName?: string | undefined;
|
|
606
|
-
|
|
633
|
+
timeoutMs: number;
|
|
634
|
+
networkPolicy: "none";
|
|
607
635
|
} | {
|
|
608
636
|
kind: "human";
|
|
609
637
|
rubricRef: {
|
|
@@ -612,7 +640,7 @@ export declare function prepareImportedTasksetPackage(input: {
|
|
|
612
640
|
contentHash: string;
|
|
613
641
|
sizeBytes: number;
|
|
614
642
|
mediaType: string;
|
|
615
|
-
visibility: "
|
|
643
|
+
visibility: "host_private" | "policy" | "verifier";
|
|
616
644
|
};
|
|
617
645
|
reviewerRole: string;
|
|
618
646
|
form?: {
|
|
@@ -625,68 +653,68 @@ export declare function prepareImportedTasksetPackage(input: {
|
|
|
625
653
|
instructions: string;
|
|
626
654
|
required: boolean;
|
|
627
655
|
allowAbstain: boolean;
|
|
628
|
-
kind: "score";
|
|
629
|
-
minimum: number;
|
|
630
|
-
maximum: number;
|
|
631
|
-
step: number;
|
|
632
656
|
requireRationale?: boolean | undefined;
|
|
633
657
|
anchors?: {
|
|
634
658
|
value: string | number | boolean;
|
|
635
659
|
description: string;
|
|
636
660
|
}[] | undefined;
|
|
661
|
+
kind: "score";
|
|
662
|
+
minimum: number;
|
|
663
|
+
maximum: number;
|
|
664
|
+
step: number;
|
|
637
665
|
} | {
|
|
638
666
|
id: string;
|
|
639
667
|
label: string;
|
|
640
668
|
instructions: string;
|
|
641
669
|
required: boolean;
|
|
642
670
|
allowAbstain: boolean;
|
|
643
|
-
kind: "category";
|
|
644
|
-
options: string[];
|
|
645
671
|
requireRationale?: boolean | undefined;
|
|
646
672
|
anchors?: {
|
|
647
673
|
value: string | number | boolean;
|
|
648
674
|
description: string;
|
|
649
675
|
}[] | undefined;
|
|
676
|
+
kind: "category";
|
|
677
|
+
options: string[];
|
|
650
678
|
} | {
|
|
651
679
|
id: string;
|
|
652
680
|
label: string;
|
|
653
681
|
instructions: string;
|
|
654
682
|
required: boolean;
|
|
655
683
|
allowAbstain: boolean;
|
|
656
|
-
kind: "boolean";
|
|
657
684
|
requireRationale?: boolean | undefined;
|
|
658
685
|
anchors?: {
|
|
659
686
|
value: string | number | boolean;
|
|
660
687
|
description: string;
|
|
661
688
|
}[] | undefined;
|
|
689
|
+
kind: "boolean";
|
|
662
690
|
} | {
|
|
663
691
|
id: string;
|
|
664
692
|
label: string;
|
|
665
693
|
instructions: string;
|
|
666
694
|
required: boolean;
|
|
667
695
|
allowAbstain: boolean;
|
|
668
|
-
kind: "text";
|
|
669
|
-
maximumLength: number;
|
|
670
696
|
requireRationale?: boolean | undefined;
|
|
671
697
|
anchors?: {
|
|
672
698
|
value: string | number | boolean;
|
|
673
699
|
description: string;
|
|
674
700
|
}[] | undefined;
|
|
701
|
+
kind: "text";
|
|
675
702
|
minimumLength?: number | undefined;
|
|
703
|
+
maximumLength: number;
|
|
676
704
|
} | {
|
|
677
705
|
id: string;
|
|
678
706
|
label: string;
|
|
679
707
|
instructions: string;
|
|
680
708
|
required: boolean;
|
|
681
709
|
allowAbstain: boolean;
|
|
682
|
-
kind: "preference";
|
|
683
|
-
allowTie: boolean;
|
|
684
710
|
requireRationale?: boolean | undefined;
|
|
685
711
|
anchors?: {
|
|
686
712
|
value: string | number | boolean;
|
|
687
713
|
description: string;
|
|
688
714
|
}[] | undefined;
|
|
715
|
+
kind: "preference";
|
|
689
716
|
requireStrength?: boolean | undefined;
|
|
717
|
+
allowTie: boolean;
|
|
690
718
|
})[];
|
|
691
719
|
} | undefined;
|
|
692
720
|
} | {
|
|
@@ -701,7 +729,7 @@ export declare function prepareImportedTasksetPackage(input: {
|
|
|
701
729
|
contentHash: string;
|
|
702
730
|
sizeBytes: number;
|
|
703
731
|
mediaType: string;
|
|
704
|
-
visibility: "
|
|
732
|
+
visibility: "host_private" | "policy" | "verifier";
|
|
705
733
|
};
|
|
706
734
|
};
|
|
707
735
|
rawScore: {
|
|
@@ -714,65 +742,38 @@ export declare function prepareImportedTasksetPackage(input: {
|
|
|
714
742
|
contentHash: string;
|
|
715
743
|
sizeBytes: number;
|
|
716
744
|
mediaType: string;
|
|
717
|
-
visibility: "
|
|
745
|
+
visibility: "host_private" | "policy" | "verifier";
|
|
718
746
|
}[];
|
|
719
|
-
contentHash: string;
|
|
720
|
-
feedbackKey?: string | undefined;
|
|
721
747
|
fixtureSetRef?: {
|
|
722
748
|
id: string;
|
|
723
749
|
path: string;
|
|
724
750
|
contentHash: string;
|
|
725
751
|
sizeBytes: number;
|
|
726
752
|
mediaType: string;
|
|
727
|
-
visibility: "
|
|
753
|
+
visibility: "host_private" | "policy" | "verifier";
|
|
728
754
|
} | undefined;
|
|
729
755
|
calibrationCheckRef?: {
|
|
730
756
|
id: string;
|
|
731
757
|
contentHash: string;
|
|
732
758
|
revision: number;
|
|
733
759
|
} | undefined;
|
|
760
|
+
contentHash: string;
|
|
734
761
|
}[];
|
|
735
|
-
|
|
736
|
-
schemaVersion: "openpond.
|
|
762
|
+
assets: {
|
|
763
|
+
schemaVersion: "openpond.learningTextAsset.v1";
|
|
737
764
|
id: string;
|
|
738
|
-
revision:
|
|
739
|
-
|
|
740
|
-
graderId: string;
|
|
741
|
-
reward: {
|
|
742
|
-
id: string;
|
|
743
|
-
contentHash: string;
|
|
744
|
-
revision: number;
|
|
745
|
-
};
|
|
746
|
-
role: "training" | "evaluation";
|
|
747
|
-
normalization: {
|
|
748
|
-
kind: "identity";
|
|
749
|
-
} | {
|
|
750
|
-
kind: "linear";
|
|
751
|
-
minimum: number;
|
|
752
|
-
maximum: number;
|
|
753
|
-
direction: "higher" | "lower";
|
|
754
|
-
};
|
|
755
|
-
weight: number;
|
|
756
|
-
required: boolean;
|
|
757
|
-
hardGate: boolean;
|
|
758
|
-
privileged: boolean;
|
|
759
|
-
fixtureRefs: {
|
|
760
|
-
id: string;
|
|
761
|
-
contentHash: string;
|
|
762
|
-
}[];
|
|
763
|
-
taskIds?: string[] | undefined;
|
|
764
|
-
}[];
|
|
765
|
-
aggregation: "weighted_mean";
|
|
766
|
-
unscorable: "exclude_optional_require_all_required";
|
|
767
|
-
contentHash: string;
|
|
768
|
-
name?: string | undefined;
|
|
769
|
-
description?: string | undefined;
|
|
770
|
-
recipeRef?: {
|
|
765
|
+
revision: 1;
|
|
766
|
+
asset: {
|
|
771
767
|
id: string;
|
|
768
|
+
path: string;
|
|
772
769
|
contentHash: string;
|
|
773
|
-
|
|
774
|
-
|
|
775
|
-
|
|
770
|
+
sizeBytes: number;
|
|
771
|
+
mediaType: string;
|
|
772
|
+
visibility: "host_private" | "policy" | "verifier";
|
|
773
|
+
};
|
|
774
|
+
text: string;
|
|
775
|
+
contentHash: string;
|
|
776
|
+
}[];
|
|
776
777
|
execution?: {
|
|
777
778
|
environment: {
|
|
778
779
|
schemaVersion: "openpond.environmentRelease.v1";
|
|
@@ -780,12 +781,12 @@ export declare function prepareImportedTasksetPackage(input: {
|
|
|
780
781
|
revision: number;
|
|
781
782
|
contract: {
|
|
782
783
|
protocolVersion: "openpond.environment.v1";
|
|
783
|
-
kind: "
|
|
784
|
+
kind: "agent" | "custom_program" | "text" | "work";
|
|
784
785
|
entrypoint: string;
|
|
785
786
|
stateful: boolean;
|
|
786
787
|
deterministicSeeds: boolean;
|
|
787
|
-
lifecycle: ("
|
|
788
|
-
networkPolicy: "
|
|
788
|
+
lifecycle: ("collect" | "create" | "destroy" | "reset" | "step")[];
|
|
789
|
+
networkPolicy: "declared_read_only" | "declared_scoped" | "none";
|
|
789
790
|
defaultTimeoutMs: number;
|
|
790
791
|
};
|
|
791
792
|
actionSchemaRef: {
|
|
@@ -794,7 +795,7 @@ export declare function prepareImportedTasksetPackage(input: {
|
|
|
794
795
|
contentHash: string;
|
|
795
796
|
sizeBytes: number;
|
|
796
797
|
mediaType: string;
|
|
797
|
-
visibility: "
|
|
798
|
+
visibility: "host_private" | "policy" | "verifier";
|
|
798
799
|
} | null;
|
|
799
800
|
observationSchemaRef: {
|
|
800
801
|
id: string;
|
|
@@ -802,7 +803,7 @@ export declare function prepareImportedTasksetPackage(input: {
|
|
|
802
803
|
contentHash: string;
|
|
803
804
|
sizeBytes: number;
|
|
804
805
|
mediaType: string;
|
|
805
|
-
visibility: "
|
|
806
|
+
visibility: "host_private" | "policy" | "verifier";
|
|
806
807
|
} | null;
|
|
807
808
|
stateSchemaRef: {
|
|
808
809
|
id: string;
|
|
@@ -810,7 +811,7 @@ export declare function prepareImportedTasksetPackage(input: {
|
|
|
810
811
|
contentHash: string;
|
|
811
812
|
sizeBytes: number;
|
|
812
813
|
mediaType: string;
|
|
813
|
-
visibility: "
|
|
814
|
+
visibility: "host_private" | "policy" | "verifier";
|
|
814
815
|
} | null;
|
|
815
816
|
artifactCollection: {
|
|
816
817
|
maxArtifacts: number;
|
|
@@ -826,16 +827,17 @@ export declare function prepareImportedTasksetPackage(input: {
|
|
|
826
827
|
revision: number;
|
|
827
828
|
graders: ({
|
|
828
829
|
id: string;
|
|
830
|
+
taskIds?: string[] | undefined;
|
|
829
831
|
version: string;
|
|
830
832
|
weight: number;
|
|
831
833
|
hardGate: boolean;
|
|
832
834
|
rewardEligible: boolean;
|
|
833
835
|
privileged: boolean;
|
|
834
|
-
kind: "
|
|
836
|
+
kind: "artifact" | "content" | "runtime_event" | "schema" | "state";
|
|
835
837
|
config: Record<string, unknown>;
|
|
836
|
-
taskIds?: string[] | undefined;
|
|
837
838
|
} | {
|
|
838
839
|
id: string;
|
|
840
|
+
taskIds?: string[] | undefined;
|
|
839
841
|
version: string;
|
|
840
842
|
weight: number;
|
|
841
843
|
hardGate: boolean;
|
|
@@ -848,10 +850,9 @@ export declare function prepareImportedTasksetPackage(input: {
|
|
|
848
850
|
contentHash: string;
|
|
849
851
|
sizeBytes: number;
|
|
850
852
|
mediaType: string;
|
|
851
|
-
visibility: "
|
|
853
|
+
visibility: "host_private" | "policy" | "verifier";
|
|
852
854
|
};
|
|
853
|
-
calibrationStatus: "
|
|
854
|
-
taskIds?: string[] | undefined;
|
|
855
|
+
calibrationStatus: "failed" | "passed" | "pending";
|
|
855
856
|
model?: {
|
|
856
857
|
providerId: string;
|
|
857
858
|
modelId: string;
|
|
@@ -860,27 +861,28 @@ export declare function prepareImportedTasksetPackage(input: {
|
|
|
860
861
|
temperature?: number | undefined;
|
|
861
862
|
} | {
|
|
862
863
|
id: string;
|
|
864
|
+
taskIds?: string[] | undefined;
|
|
863
865
|
version: string;
|
|
864
866
|
weight: number;
|
|
865
867
|
hardGate: boolean;
|
|
866
868
|
rewardEligible: boolean;
|
|
867
869
|
privileged: boolean;
|
|
868
870
|
kind: "custom_verifier";
|
|
871
|
+
runtime?: "isolated_javascript" | "sandbox_process" | undefined;
|
|
869
872
|
verifierRef: {
|
|
870
873
|
id: string;
|
|
871
874
|
path: string;
|
|
872
875
|
contentHash: string;
|
|
873
876
|
sizeBytes: number;
|
|
874
877
|
mediaType: string;
|
|
875
|
-
visibility: "
|
|
878
|
+
visibility: "host_private" | "policy" | "verifier";
|
|
876
879
|
};
|
|
880
|
+
exportName?: string | undefined;
|
|
877
881
|
timeoutMs: number;
|
|
878
882
|
networkPolicy: "none";
|
|
879
|
-
taskIds?: string[] | undefined;
|
|
880
|
-
runtime?: "isolated_javascript" | "sandbox_process" | undefined;
|
|
881
|
-
exportName?: string | undefined;
|
|
882
883
|
} | {
|
|
883
884
|
id: string;
|
|
885
|
+
taskIds?: string[] | undefined;
|
|
884
886
|
version: string;
|
|
885
887
|
weight: number;
|
|
886
888
|
hardGate: boolean;
|
|
@@ -893,10 +895,9 @@ export declare function prepareImportedTasksetPackage(input: {
|
|
|
893
895
|
contentHash: string;
|
|
894
896
|
sizeBytes: number;
|
|
895
897
|
mediaType: string;
|
|
896
|
-
visibility: "
|
|
898
|
+
visibility: "host_private" | "policy" | "verifier";
|
|
897
899
|
};
|
|
898
900
|
reviewerRole: string;
|
|
899
|
-
taskIds?: string[] | undefined;
|
|
900
901
|
form?: {
|
|
901
902
|
schemaVersion: "openpond.humanForm.v1";
|
|
902
903
|
mode: "individual" | "pairwise";
|
|
@@ -907,73 +908,73 @@ export declare function prepareImportedTasksetPackage(input: {
|
|
|
907
908
|
instructions: string;
|
|
908
909
|
required: boolean;
|
|
909
910
|
allowAbstain: boolean;
|
|
910
|
-
kind: "score";
|
|
911
|
-
minimum: number;
|
|
912
|
-
maximum: number;
|
|
913
|
-
step: number;
|
|
914
911
|
requireRationale?: boolean | undefined;
|
|
915
912
|
anchors?: {
|
|
916
913
|
value: string | number | boolean;
|
|
917
914
|
description: string;
|
|
918
915
|
}[] | undefined;
|
|
916
|
+
kind: "score";
|
|
917
|
+
minimum: number;
|
|
918
|
+
maximum: number;
|
|
919
|
+
step: number;
|
|
919
920
|
} | {
|
|
920
921
|
id: string;
|
|
921
922
|
label: string;
|
|
922
923
|
instructions: string;
|
|
923
924
|
required: boolean;
|
|
924
925
|
allowAbstain: boolean;
|
|
925
|
-
kind: "category";
|
|
926
|
-
options: string[];
|
|
927
926
|
requireRationale?: boolean | undefined;
|
|
928
927
|
anchors?: {
|
|
929
928
|
value: string | number | boolean;
|
|
930
929
|
description: string;
|
|
931
930
|
}[] | undefined;
|
|
931
|
+
kind: "category";
|
|
932
|
+
options: string[];
|
|
932
933
|
} | {
|
|
933
934
|
id: string;
|
|
934
935
|
label: string;
|
|
935
936
|
instructions: string;
|
|
936
937
|
required: boolean;
|
|
937
938
|
allowAbstain: boolean;
|
|
938
|
-
kind: "boolean";
|
|
939
939
|
requireRationale?: boolean | undefined;
|
|
940
940
|
anchors?: {
|
|
941
941
|
value: string | number | boolean;
|
|
942
942
|
description: string;
|
|
943
943
|
}[] | undefined;
|
|
944
|
+
kind: "boolean";
|
|
944
945
|
} | {
|
|
945
946
|
id: string;
|
|
946
947
|
label: string;
|
|
947
948
|
instructions: string;
|
|
948
949
|
required: boolean;
|
|
949
950
|
allowAbstain: boolean;
|
|
950
|
-
kind: "text";
|
|
951
|
-
maximumLength: number;
|
|
952
951
|
requireRationale?: boolean | undefined;
|
|
953
952
|
anchors?: {
|
|
954
953
|
value: string | number | boolean;
|
|
955
954
|
description: string;
|
|
956
955
|
}[] | undefined;
|
|
956
|
+
kind: "text";
|
|
957
957
|
minimumLength?: number | undefined;
|
|
958
|
+
maximumLength: number;
|
|
958
959
|
} | {
|
|
959
960
|
id: string;
|
|
960
961
|
label: string;
|
|
961
962
|
instructions: string;
|
|
962
963
|
required: boolean;
|
|
963
964
|
allowAbstain: boolean;
|
|
964
|
-
kind: "preference";
|
|
965
|
-
allowTie: boolean;
|
|
966
965
|
requireRationale?: boolean | undefined;
|
|
967
966
|
anchors?: {
|
|
968
967
|
value: string | number | boolean;
|
|
969
968
|
description: string;
|
|
970
969
|
}[] | undefined;
|
|
970
|
+
kind: "preference";
|
|
971
971
|
requireStrength?: boolean | undefined;
|
|
972
|
+
allowTie: boolean;
|
|
972
973
|
})[];
|
|
973
974
|
} | undefined;
|
|
974
975
|
})[];
|
|
975
976
|
isolation: {
|
|
976
|
-
processBoundary: "
|
|
977
|
+
processBoundary: "container" | "isolated_process" | "same_process";
|
|
977
978
|
networkPolicy: "none";
|
|
978
979
|
defaultTimeoutMs: number;
|
|
979
980
|
};
|
|
@@ -994,7 +995,7 @@ export declare function prepareImportedTasksetPackage(input: {
|
|
|
994
995
|
contentHash: string;
|
|
995
996
|
sizeBytes: number;
|
|
996
997
|
mediaType: string;
|
|
997
|
-
visibility: "
|
|
998
|
+
visibility: "host_private" | "policy" | "verifier";
|
|
998
999
|
};
|
|
999
1000
|
tools: {
|
|
1000
1001
|
name: string;
|
|
@@ -1008,41 +1009,41 @@ export declare function prepareImportedTasksetPackage(input: {
|
|
|
1008
1009
|
maxStateBytes: number;
|
|
1009
1010
|
maxObservationBytes: number;
|
|
1010
1011
|
operationTimeoutMs: number;
|
|
1011
|
-
contentHash: string;
|
|
1012
1012
|
executionServices?: ({
|
|
1013
|
+
id: string;
|
|
1014
|
+
operation: "collect" | "step";
|
|
1015
|
+
toolName?: string | undefined;
|
|
1016
|
+
timeoutMs: number;
|
|
1013
1017
|
kind: "sqlite.v1";
|
|
1014
1018
|
sql: {
|
|
1015
|
-
scope: "
|
|
1019
|
+
scope: "arguments" | "initialState" | "input" | "state";
|
|
1016
1020
|
path: string[];
|
|
1017
1021
|
};
|
|
1018
1022
|
snapshot: {
|
|
1019
|
-
scope: "
|
|
1023
|
+
scope: "arguments" | "initialState" | "input" | "state";
|
|
1020
1024
|
path: string[];
|
|
1021
1025
|
};
|
|
1022
1026
|
maxRows: number;
|
|
1023
1027
|
maxResultBytes: number;
|
|
1028
|
+
} | {
|
|
1024
1029
|
id: string;
|
|
1025
|
-
operation: "
|
|
1026
|
-
timeoutMs: number;
|
|
1030
|
+
operation: "collect" | "step";
|
|
1027
1031
|
toolName?: string | undefined;
|
|
1028
|
-
|
|
1032
|
+
timeoutMs: number;
|
|
1029
1033
|
kind: "javascript.v1";
|
|
1030
1034
|
source: {
|
|
1031
|
-
scope: "
|
|
1035
|
+
scope: "arguments" | "initialState" | "input" | "state";
|
|
1032
1036
|
path: string[];
|
|
1033
1037
|
};
|
|
1034
1038
|
cases: {
|
|
1035
|
-
scope: "
|
|
1039
|
+
scope: "arguments" | "initialState" | "input" | "state";
|
|
1036
1040
|
path: string[];
|
|
1037
1041
|
};
|
|
1038
1042
|
exportName: string;
|
|
1039
1043
|
maxCases: number;
|
|
1040
1044
|
maxResultBytes: number;
|
|
1041
|
-
id: string;
|
|
1042
|
-
operation: "step" | "collect";
|
|
1043
|
-
timeoutMs: number;
|
|
1044
|
-
toolName?: string | undefined;
|
|
1045
1045
|
})[] | undefined;
|
|
1046
|
+
contentHash: string;
|
|
1046
1047
|
};
|
|
1047
1048
|
} | undefined;
|
|
1048
1049
|
instructionMode?: "per_task" | undefined;
|
|
@@ -1062,7 +1063,7 @@ export declare function prepareImportedTasksetPackage(input: {
|
|
|
1062
1063
|
contentHash: string;
|
|
1063
1064
|
revision: number;
|
|
1064
1065
|
};
|
|
1065
|
-
purpose: "evaluation" | "
|
|
1066
|
+
purpose: "evaluation" | "reward_training" | "supervised_training";
|
|
1066
1067
|
examples: {
|
|
1067
1068
|
evidence: {
|
|
1068
1069
|
id: string;
|
|
@@ -1076,7 +1077,7 @@ export declare function prepareImportedTasksetPackage(input: {
|
|
|
1076
1077
|
};
|
|
1077
1078
|
familyKey: string;
|
|
1078
1079
|
inputHash: string;
|
|
1079
|
-
split: "
|
|
1080
|
+
split: "frozen_eval" | "test" | "train" | "validation";
|
|
1080
1081
|
}[];
|
|
1081
1082
|
sealedAt: string;
|
|
1082
1083
|
sealedBy: string;
|
|
@@ -1104,23 +1105,19 @@ export declare function prepareImportedTasksetPackage(input: {
|
|
|
1104
1105
|
attemptId: string;
|
|
1105
1106
|
occurredAt: string;
|
|
1106
1107
|
familyKey: string | null;
|
|
1107
|
-
split: "
|
|
1108
|
-
input: Record<string, z.
|
|
1109
|
-
observedOutput: Record<string, z.
|
|
1110
|
-
expected: Record<string, z.
|
|
1111
|
-
evaluatorContext: Record<string, z.
|
|
1108
|
+
split: "frozen_eval" | "test" | "train" | "validation";
|
|
1109
|
+
input: Record<string, z.JSONType>;
|
|
1110
|
+
observedOutput: Record<string, z.JSONType> | null;
|
|
1111
|
+
expected: Record<string, z.JSONType> | null;
|
|
1112
|
+
evaluatorContext: Record<string, z.JSONType> | null;
|
|
1112
1113
|
assets: {
|
|
1113
1114
|
id: string;
|
|
1114
1115
|
path: string;
|
|
1115
1116
|
contentHash: string;
|
|
1116
1117
|
sizeBytes: number;
|
|
1117
1118
|
mediaType: string;
|
|
1118
|
-
visibility: "
|
|
1119
|
+
visibility: "host_private" | "policy" | "verifier";
|
|
1119
1120
|
}[];
|
|
1120
|
-
provenance: {
|
|
1121
|
-
sourceRecordRef: string | null;
|
|
1122
|
-
mappingHash: string | null;
|
|
1123
|
-
};
|
|
1124
1121
|
requiredOutputs?: {
|
|
1125
1122
|
path: string;
|
|
1126
1123
|
mediaType: string;
|
|
@@ -1130,11 +1127,16 @@ export declare function prepareImportedTasksetPackage(input: {
|
|
|
1130
1127
|
contentHash: string;
|
|
1131
1128
|
sizeBytes: number;
|
|
1132
1129
|
mediaType: string;
|
|
1133
|
-
visibility: "
|
|
1130
|
+
visibility: "host_private" | "policy" | "verifier";
|
|
1134
1131
|
} | null;
|
|
1132
|
+
submissionMode?: "structured_json" | undefined;
|
|
1135
1133
|
maxBytes: number | null;
|
|
1136
1134
|
metadata: Record<string, unknown>;
|
|
1137
1135
|
}[] | undefined;
|
|
1136
|
+
provenance: {
|
|
1137
|
+
sourceRecordRef: string | null;
|
|
1138
|
+
mappingHash: string | null;
|
|
1139
|
+
};
|
|
1138
1140
|
};
|
|
1139
1141
|
supersedes: {
|
|
1140
1142
|
id: string;
|
|
@@ -1168,11 +1170,11 @@ export declare function prepareImportedTasksetPackage(input: {
|
|
|
1168
1170
|
revision: number;
|
|
1169
1171
|
} | null;
|
|
1170
1172
|
};
|
|
1171
|
-
evidenceValidity: "
|
|
1172
|
-
taskAdmissibility: "
|
|
1173
|
-
observedQuality: "
|
|
1174
|
-
targetApproval: "
|
|
1175
|
-
approvedTarget: Record<string, z.
|
|
1173
|
+
evidenceValidity: "invalid" | "pending" | "valid";
|
|
1174
|
+
taskAdmissibility: "approved" | "pending" | "rejected";
|
|
1175
|
+
observedQuality: "failed" | "passed" | "unavailable" | "unscored";
|
|
1176
|
+
targetApproval: "approved" | "not_required" | "pending" | "rejected";
|
|
1177
|
+
approvedTarget: Record<string, z.JSONType> | null;
|
|
1176
1178
|
grade: {
|
|
1177
1179
|
schemaVersion: "openpond.rewardComposition.v1";
|
|
1178
1180
|
binding: {
|
|
@@ -1181,6 +1183,7 @@ export declare function prepareImportedTasksetPackage(input: {
|
|
|
1181
1183
|
revision: number;
|
|
1182
1184
|
};
|
|
1183
1185
|
taskHash: string;
|
|
1186
|
+
taskId?: string | undefined;
|
|
1184
1187
|
outputHash: string;
|
|
1185
1188
|
results: {
|
|
1186
1189
|
graderId: string;
|
|
@@ -1189,26 +1192,24 @@ export declare function prepareImportedTasksetPackage(input: {
|
|
|
1189
1192
|
contentHash: string;
|
|
1190
1193
|
revision: number;
|
|
1191
1194
|
};
|
|
1192
|
-
role: "
|
|
1193
|
-
status: "
|
|
1195
|
+
role: "evaluation" | "training";
|
|
1196
|
+
status: "failed" | "not_applicable" | "pending" | "scored" | "unavailable";
|
|
1194
1197
|
rawScore: number | null;
|
|
1195
1198
|
normalizedScore: number | null;
|
|
1196
1199
|
passed: boolean | null;
|
|
1197
1200
|
evidenceHashes: string[];
|
|
1198
|
-
message: string | null;
|
|
1199
1201
|
graderEvidence?: {
|
|
1200
1202
|
schemaVersion: "openpond.graderEvidence.v1";
|
|
1201
1203
|
graderId: string;
|
|
1202
1204
|
graderVersion: string;
|
|
1205
|
+
status?: "failed" | "pending" | "scored" | "unavailable" | undefined;
|
|
1203
1206
|
score: number | null;
|
|
1204
1207
|
passed: boolean;
|
|
1205
1208
|
rewardEligible: boolean;
|
|
1206
|
-
failureClass: "
|
|
1209
|
+
failureClass: "cancelled" | "environment_failure" | "grader_failure" | "infrastructure_failure" | "policy_failure" | "timeout" | null;
|
|
1207
1210
|
feedback: string[];
|
|
1208
1211
|
visibleEvidenceRefs: string[];
|
|
1209
1212
|
privilegedEvidenceRefs: string[];
|
|
1210
|
-
contentHash: string;
|
|
1211
|
-
status?: "scored" | "pending" | "failed" | "unavailable" | undefined;
|
|
1212
1213
|
modelJudgeReceipt?: {
|
|
1213
1214
|
schemaVersion: "openpond.modelJudgeReceipt.v1";
|
|
1214
1215
|
providerId: string;
|
|
@@ -1223,27 +1224,28 @@ export declare function prepareImportedTasksetPackage(input: {
|
|
|
1223
1224
|
transportPolicy?: {
|
|
1224
1225
|
schemaVersion: "openpond.modelJudgeTransportPolicy.v1";
|
|
1225
1226
|
responseFormat: "judgment_json_schema";
|
|
1226
|
-
reasoningEffort: "high" | "low" | "
|
|
1227
|
+
reasoningEffort: "high" | "low" | "max" | "medium" | "off" | "xhigh";
|
|
1227
1228
|
maximumOutputTokens: number;
|
|
1228
1229
|
} | undefined;
|
|
1229
1230
|
effectiveRequestHash?: string | undefined;
|
|
1230
1231
|
finishReason?: string | null | undefined;
|
|
1231
1232
|
reasoningTokens?: number | null | undefined;
|
|
1232
1233
|
} | undefined;
|
|
1234
|
+
contentHash: string;
|
|
1233
1235
|
} | undefined;
|
|
1236
|
+
message: string | null;
|
|
1234
1237
|
}[];
|
|
1235
1238
|
training: {
|
|
1236
|
-
status: "
|
|
1239
|
+
status: "not_configured" | "pending" | "scored" | "unscorable";
|
|
1237
1240
|
score: number | null;
|
|
1238
1241
|
passed: boolean | null;
|
|
1239
1242
|
};
|
|
1240
1243
|
evaluation: {
|
|
1241
|
-
status: "
|
|
1244
|
+
status: "not_configured" | "pending" | "scored" | "unscorable";
|
|
1242
1245
|
score: number | null;
|
|
1243
1246
|
passed: boolean | null;
|
|
1244
1247
|
};
|
|
1245
1248
|
contentHash: string;
|
|
1246
|
-
taskId?: string | undefined;
|
|
1247
1249
|
} | null;
|
|
1248
1250
|
targetGrade: {
|
|
1249
1251
|
schemaVersion: "openpond.rewardComposition.v1";
|
|
@@ -1253,6 +1255,7 @@ export declare function prepareImportedTasksetPackage(input: {
|
|
|
1253
1255
|
revision: number;
|
|
1254
1256
|
};
|
|
1255
1257
|
taskHash: string;
|
|
1258
|
+
taskId?: string | undefined;
|
|
1256
1259
|
outputHash: string;
|
|
1257
1260
|
results: {
|
|
1258
1261
|
graderId: string;
|
|
@@ -1261,26 +1264,24 @@ export declare function prepareImportedTasksetPackage(input: {
|
|
|
1261
1264
|
contentHash: string;
|
|
1262
1265
|
revision: number;
|
|
1263
1266
|
};
|
|
1264
|
-
role: "
|
|
1265
|
-
status: "
|
|
1267
|
+
role: "evaluation" | "training";
|
|
1268
|
+
status: "failed" | "not_applicable" | "pending" | "scored" | "unavailable";
|
|
1266
1269
|
rawScore: number | null;
|
|
1267
1270
|
normalizedScore: number | null;
|
|
1268
1271
|
passed: boolean | null;
|
|
1269
1272
|
evidenceHashes: string[];
|
|
1270
|
-
message: string | null;
|
|
1271
1273
|
graderEvidence?: {
|
|
1272
1274
|
schemaVersion: "openpond.graderEvidence.v1";
|
|
1273
1275
|
graderId: string;
|
|
1274
1276
|
graderVersion: string;
|
|
1277
|
+
status?: "failed" | "pending" | "scored" | "unavailable" | undefined;
|
|
1275
1278
|
score: number | null;
|
|
1276
1279
|
passed: boolean;
|
|
1277
1280
|
rewardEligible: boolean;
|
|
1278
|
-
failureClass: "
|
|
1281
|
+
failureClass: "cancelled" | "environment_failure" | "grader_failure" | "infrastructure_failure" | "policy_failure" | "timeout" | null;
|
|
1279
1282
|
feedback: string[];
|
|
1280
1283
|
visibleEvidenceRefs: string[];
|
|
1281
1284
|
privilegedEvidenceRefs: string[];
|
|
1282
|
-
contentHash: string;
|
|
1283
|
-
status?: "scored" | "pending" | "failed" | "unavailable" | undefined;
|
|
1284
1285
|
modelJudgeReceipt?: {
|
|
1285
1286
|
schemaVersion: "openpond.modelJudgeReceipt.v1";
|
|
1286
1287
|
providerId: string;
|
|
@@ -1295,27 +1296,28 @@ export declare function prepareImportedTasksetPackage(input: {
|
|
|
1295
1296
|
transportPolicy?: {
|
|
1296
1297
|
schemaVersion: "openpond.modelJudgeTransportPolicy.v1";
|
|
1297
1298
|
responseFormat: "judgment_json_schema";
|
|
1298
|
-
reasoningEffort: "high" | "low" | "
|
|
1299
|
+
reasoningEffort: "high" | "low" | "max" | "medium" | "off" | "xhigh";
|
|
1299
1300
|
maximumOutputTokens: number;
|
|
1300
1301
|
} | undefined;
|
|
1301
1302
|
effectiveRequestHash?: string | undefined;
|
|
1302
1303
|
finishReason?: string | null | undefined;
|
|
1303
1304
|
reasoningTokens?: number | null | undefined;
|
|
1304
1305
|
} | undefined;
|
|
1306
|
+
contentHash: string;
|
|
1305
1307
|
} | undefined;
|
|
1308
|
+
message: string | null;
|
|
1306
1309
|
}[];
|
|
1307
1310
|
training: {
|
|
1308
|
-
status: "
|
|
1311
|
+
status: "not_configured" | "pending" | "scored" | "unscorable";
|
|
1309
1312
|
score: number | null;
|
|
1310
1313
|
passed: boolean | null;
|
|
1311
1314
|
};
|
|
1312
1315
|
evaluation: {
|
|
1313
|
-
status: "
|
|
1316
|
+
status: "not_configured" | "pending" | "scored" | "unscorable";
|
|
1314
1317
|
score: number | null;
|
|
1315
1318
|
passed: boolean | null;
|
|
1316
1319
|
};
|
|
1317
1320
|
contentHash: string;
|
|
1318
|
-
taskId?: string | undefined;
|
|
1319
1321
|
} | null;
|
|
1320
1322
|
note: string;
|
|
1321
1323
|
decidedAt: string;
|
|
@@ -1326,14 +1328,14 @@ export declare function prepareImportedTasksetPackage(input: {
|
|
|
1326
1328
|
id: string;
|
|
1327
1329
|
revision: number;
|
|
1328
1330
|
name: string;
|
|
1329
|
-
kind: "
|
|
1331
|
+
kind: "benchmark" | "direct" | "opentelemetry" | "provider" | "work";
|
|
1330
1332
|
taskDefinition: {
|
|
1331
1333
|
id: string;
|
|
1332
1334
|
contentHash: string;
|
|
1333
1335
|
revision: number;
|
|
1334
1336
|
};
|
|
1335
1337
|
enabled: boolean;
|
|
1336
|
-
allowedSplits: ("
|
|
1338
|
+
allowedSplits: ("frozen_eval" | "test" | "train" | "validation")[];
|
|
1337
1339
|
mapping: {
|
|
1338
1340
|
schemaVersion: "openpond.taskSourceMapping.v1";
|
|
1339
1341
|
exampleId: string;
|
|
@@ -1344,10 +1346,9 @@ export declare function prepareImportedTasksetPackage(input: {
|
|
|
1344
1346
|
expected: string | null;
|
|
1345
1347
|
evaluatorContext: string | null;
|
|
1346
1348
|
familyKey: string;
|
|
1347
|
-
split: "
|
|
1349
|
+
split: "frozen_eval" | "test" | "train" | "validation";
|
|
1348
1350
|
} | null;
|
|
1349
1351
|
adapterVersion: string | null;
|
|
1350
|
-
contentHash: string;
|
|
1351
1352
|
reviewOrigin?: {
|
|
1352
1353
|
modelId: string;
|
|
1353
1354
|
packageHash: string;
|
|
@@ -1378,6 +1379,7 @@ export declare function prepareImportedTasksetPackage(input: {
|
|
|
1378
1379
|
revision: number;
|
|
1379
1380
|
};
|
|
1380
1381
|
} | undefined;
|
|
1382
|
+
contentHash: string;
|
|
1381
1383
|
}[];
|
|
1382
1384
|
assets: {
|
|
1383
1385
|
schemaVersion: "openpond.learningTextAsset.v1";
|
|
@@ -1389,32 +1391,38 @@ export declare function prepareImportedTasksetPackage(input: {
|
|
|
1389
1391
|
contentHash: string;
|
|
1390
1392
|
sizeBytes: number;
|
|
1391
1393
|
mediaType: string;
|
|
1392
|
-
visibility: "
|
|
1394
|
+
visibility: "host_private" | "policy" | "verifier";
|
|
1393
1395
|
};
|
|
1394
1396
|
text: string;
|
|
1395
1397
|
contentHash: string;
|
|
1396
1398
|
}[];
|
|
1397
1399
|
} | undefined;
|
|
1400
|
+
contentHash: string;
|
|
1398
1401
|
};
|
|
1399
1402
|
taskset: {
|
|
1400
1403
|
schemaVersion: "openpond.taskset.v1";
|
|
1401
1404
|
id: string;
|
|
1402
1405
|
revision: number;
|
|
1403
1406
|
profileId: string;
|
|
1407
|
+
profileRelease?: {
|
|
1408
|
+
id: string;
|
|
1409
|
+
contentHash: string;
|
|
1410
|
+
revision: number;
|
|
1411
|
+
} | null | undefined;
|
|
1404
1412
|
createImproveRunId: string | null;
|
|
1405
1413
|
name: string;
|
|
1406
1414
|
objective: string;
|
|
1407
|
-
purpose: "
|
|
1415
|
+
purpose: "benchmark" | "general";
|
|
1408
1416
|
benchmark: {
|
|
1409
1417
|
schemaVersion: "openpond.tasksetBenchmark.v1";
|
|
1410
1418
|
definitionId: string;
|
|
1411
1419
|
releaseId: string;
|
|
1412
1420
|
releaseHash: string;
|
|
1413
1421
|
managedReleasePath: string;
|
|
1414
|
-
adaptationSplit: "
|
|
1415
|
-
evaluationSplit: "
|
|
1416
|
-
primaryMetric: "
|
|
1417
|
-
qualityGate: "
|
|
1422
|
+
adaptationSplit: "frozen_eval" | "test" | "train" | "validation";
|
|
1423
|
+
evaluationSplit: "frozen_eval" | "test" | "train" | "validation";
|
|
1424
|
+
primaryMetric: "cost_usd" | "foreground_tokens" | "latency_ms" | "success_rate";
|
|
1425
|
+
qualityGate: "all_pass" | "non_regression" | "none";
|
|
1418
1426
|
source: "builtin" | "imported";
|
|
1419
1427
|
metadata: Record<string, unknown>;
|
|
1420
1428
|
} | null;
|
|
@@ -1425,7 +1433,7 @@ export declare function prepareImportedTasksetPackage(input: {
|
|
|
1425
1433
|
publishedAt: string;
|
|
1426
1434
|
metadata: Record<string, unknown>;
|
|
1427
1435
|
} | null;
|
|
1428
|
-
status: "
|
|
1436
|
+
status: "archived" | "awaiting_disclosure_approval" | "awaiting_materialization_approval" | "baselining" | "blocked" | "draft" | "failed" | "materializing" | "needs_review" | "ready" | "validating";
|
|
1429
1437
|
sourceRefs: ({
|
|
1430
1438
|
schemaVersion: "openpond.connectedDatasetSource.v1";
|
|
1431
1439
|
kind: "connected_activity";
|
|
@@ -1434,27 +1442,27 @@ export declare function prepareImportedTasksetPackage(input: {
|
|
|
1434
1442
|
title: string;
|
|
1435
1443
|
sourceHash: string;
|
|
1436
1444
|
occurredAt: string;
|
|
1437
|
-
licensingStatus: "
|
|
1438
|
-
secretScanStatus: "
|
|
1439
|
-
piiScanStatus: "
|
|
1440
|
-
origin: "
|
|
1445
|
+
licensingStatus: "approved" | "blocked" | "pending" | "review";
|
|
1446
|
+
secretScanStatus: "blocked" | "passed" | "pending";
|
|
1447
|
+
piiScanStatus: "blocked" | "passed" | "pending" | "review";
|
|
1448
|
+
origin: "claude_code" | "codex" | "grok_build" | "hermes" | "native_chat" | "native_work" | "oh_my_pi" | "openclaw" | "opencode" | "pi";
|
|
1441
1449
|
sourceId: string;
|
|
1442
1450
|
sessionId: string;
|
|
1443
1451
|
normalizerVersion: string;
|
|
1444
|
-
projection: "
|
|
1452
|
+
projection: "conversation" | "turn";
|
|
1445
1453
|
ownerUserId: string;
|
|
1446
1454
|
purpose: "recorded_evaluation";
|
|
1447
|
-
metadata: Record<string, unknown>;
|
|
1448
1455
|
sourceFileHashes?: string[] | undefined;
|
|
1456
|
+
metadata: Record<string, unknown>;
|
|
1449
1457
|
} | {
|
|
1450
1458
|
id: string;
|
|
1451
1459
|
profileId: string;
|
|
1452
1460
|
title: string;
|
|
1453
1461
|
sourceHash: string;
|
|
1454
1462
|
occurredAt: string;
|
|
1455
|
-
licensingStatus: "
|
|
1456
|
-
secretScanStatus: "
|
|
1457
|
-
piiScanStatus: "
|
|
1463
|
+
licensingStatus: "approved" | "blocked" | "pending" | "review";
|
|
1464
|
+
secretScanStatus: "blocked" | "passed" | "pending";
|
|
1465
|
+
piiScanStatus: "blocked" | "passed" | "pending" | "review";
|
|
1458
1466
|
metadata: Record<string, unknown>;
|
|
1459
1467
|
schemaVersion: "openpond.huggingFaceDatasetSource.v1";
|
|
1460
1468
|
kind: "huggingface";
|
|
@@ -1473,9 +1481,9 @@ export declare function prepareImportedTasksetPackage(input: {
|
|
|
1473
1481
|
title: string;
|
|
1474
1482
|
sourceHash: string;
|
|
1475
1483
|
occurredAt: string;
|
|
1476
|
-
licensingStatus: "
|
|
1477
|
-
secretScanStatus: "
|
|
1478
|
-
piiScanStatus: "
|
|
1484
|
+
licensingStatus: "approved" | "blocked" | "pending" | "review";
|
|
1485
|
+
secretScanStatus: "blocked" | "passed" | "pending";
|
|
1486
|
+
piiScanStatus: "blocked" | "passed" | "pending" | "review";
|
|
1479
1487
|
metadata: Record<string, unknown>;
|
|
1480
1488
|
schemaVersion: "openpond.uploadedFileDatasetSource.v1";
|
|
1481
1489
|
kind: "uploaded_file";
|
|
@@ -1490,9 +1498,9 @@ export declare function prepareImportedTasksetPackage(input: {
|
|
|
1490
1498
|
title: string;
|
|
1491
1499
|
sourceHash: string;
|
|
1492
1500
|
occurredAt: string;
|
|
1493
|
-
licensingStatus: "
|
|
1494
|
-
secretScanStatus: "
|
|
1495
|
-
piiScanStatus: "
|
|
1501
|
+
licensingStatus: "approved" | "blocked" | "pending" | "review";
|
|
1502
|
+
secretScanStatus: "blocked" | "passed" | "pending";
|
|
1503
|
+
piiScanStatus: "blocked" | "passed" | "pending" | "review";
|
|
1496
1504
|
metadata: Record<string, unknown>;
|
|
1497
1505
|
schemaVersion: "openpond.generatedDatasetSource.v1";
|
|
1498
1506
|
kind: "generated";
|
|
@@ -1506,9 +1514,9 @@ export declare function prepareImportedTasksetPackage(input: {
|
|
|
1506
1514
|
title: string;
|
|
1507
1515
|
sourceHash: string;
|
|
1508
1516
|
occurredAt: string;
|
|
1509
|
-
licensingStatus: "
|
|
1510
|
-
secretScanStatus: "
|
|
1511
|
-
piiScanStatus: "
|
|
1517
|
+
licensingStatus: "approved" | "blocked" | "pending" | "review";
|
|
1518
|
+
secretScanStatus: "blocked" | "passed" | "pending";
|
|
1519
|
+
piiScanStatus: "blocked" | "passed" | "pending" | "review";
|
|
1512
1520
|
metadata: Record<string, unknown>;
|
|
1513
1521
|
schemaVersion: "openpond.learningBatchDatasetSource.v1";
|
|
1514
1522
|
kind: "learning_batch";
|
|
@@ -1535,18 +1543,53 @@ export declare function prepareImportedTasksetPackage(input: {
|
|
|
1535
1543
|
title: string;
|
|
1536
1544
|
occurredAt: string;
|
|
1537
1545
|
consent: {
|
|
1538
|
-
status: "
|
|
1539
|
-
scope: "
|
|
1546
|
+
status: "denied" | "granted" | "pending" | "revoked";
|
|
1547
|
+
scope: "full_session" | "metadata_only" | "selected_turns";
|
|
1540
1548
|
grantedBy: string | null;
|
|
1541
1549
|
grantedAt: string | null;
|
|
1542
1550
|
purpose: "task_authoring_and_evaluation";
|
|
1543
1551
|
};
|
|
1544
1552
|
connectedAppIds: string[];
|
|
1545
|
-
secretScanStatus: "
|
|
1546
|
-
piiScanStatus: "
|
|
1547
|
-
licensingStatus: "
|
|
1553
|
+
secretScanStatus: "blocked" | "passed" | "pending";
|
|
1554
|
+
piiScanStatus: "blocked" | "passed" | "pending" | "review";
|
|
1555
|
+
licensingStatus: "approved" | "blocked" | "pending" | "review";
|
|
1548
1556
|
metadata: Record<string, unknown>;
|
|
1549
1557
|
})[];
|
|
1558
|
+
datasetArtifact?: {
|
|
1559
|
+
schemaVersion: "openpond.datasetArtifact.v1";
|
|
1560
|
+
id: string;
|
|
1561
|
+
tasksetId: string;
|
|
1562
|
+
tasksetRevision: number;
|
|
1563
|
+
contentHash: string;
|
|
1564
|
+
format: "parquet";
|
|
1565
|
+
schema: {
|
|
1566
|
+
schemaVersion: "openpond.datasetSemanticSchema.v1";
|
|
1567
|
+
fields: {
|
|
1568
|
+
name: string;
|
|
1569
|
+
semanticRole: "chosen" | "cluster_id" | "demonstration" | "expected_output" | "feedback" | "messages" | "metadata" | "privileged_context" | "prompt" | "rejected" | "reward" | "row_id" | "tag";
|
|
1570
|
+
logicalType: "boolean" | "float" | "integer" | "json" | "messages" | "string";
|
|
1571
|
+
nullable: boolean;
|
|
1572
|
+
policy: "metadata" | "privileged" | "visible";
|
|
1573
|
+
}[];
|
|
1574
|
+
schemaHash: string;
|
|
1575
|
+
};
|
|
1576
|
+
shards: {
|
|
1577
|
+
id: string;
|
|
1578
|
+
split: "frozen_eval" | "test" | "train" | "validation";
|
|
1579
|
+
path: string;
|
|
1580
|
+
contentHash: string;
|
|
1581
|
+
schemaHash: string;
|
|
1582
|
+
sizeBytes: number;
|
|
1583
|
+
rowCount: number;
|
|
1584
|
+
rowGroupCount: number;
|
|
1585
|
+
}[];
|
|
1586
|
+
rowCount: number;
|
|
1587
|
+
splitCounts: Record<"frozen_eval" | "test" | "train" | "validation", number>;
|
|
1588
|
+
sourceReceiptRefs: string[];
|
|
1589
|
+
mappingHash: string;
|
|
1590
|
+
qualityReportHash: string;
|
|
1591
|
+
createdAt: string;
|
|
1592
|
+
} | null | undefined;
|
|
1550
1593
|
policy: {
|
|
1551
1594
|
policyVisibleFields: string[];
|
|
1552
1595
|
privilegedFields: string[];
|
|
@@ -1555,15 +1598,11 @@ export declare function prepareImportedTasksetPackage(input: {
|
|
|
1555
1598
|
};
|
|
1556
1599
|
environment: {
|
|
1557
1600
|
protocolVersion: "openpond.taskEnvironment.v1";
|
|
1558
|
-
kind: "agent" | "
|
|
1601
|
+
kind: "agent" | "chat" | "program" | "stateful_harness" | "work";
|
|
1559
1602
|
entrypoint: string;
|
|
1560
1603
|
stateful: boolean;
|
|
1561
1604
|
deterministicSeeds: boolean;
|
|
1562
1605
|
toolNames: string[];
|
|
1563
|
-
lifecycle: ("grade" | "create" | "reset" | "step" | "cleanup")[];
|
|
1564
|
-
defaultTimeoutMs: number;
|
|
1565
|
-
networkPolicy: "none" | "declared_read_only" | "declared_scoped";
|
|
1566
|
-
metadata: Record<string, unknown>;
|
|
1567
1606
|
actionBindings?: {
|
|
1568
1607
|
schemaVersion: "openpond.harnessActionBinding.v1";
|
|
1569
1608
|
actionId: string;
|
|
@@ -1586,64 +1625,84 @@ export declare function prepareImportedTasksetPackage(input: {
|
|
|
1586
1625
|
source: "case_id";
|
|
1587
1626
|
}[];
|
|
1588
1627
|
}[] | undefined;
|
|
1628
|
+
lifecycle: ("cleanup" | "create" | "grade" | "reset" | "step")[];
|
|
1629
|
+
defaultTimeoutMs: number;
|
|
1630
|
+
networkPolicy: "declared_read_only" | "declared_scoped" | "none";
|
|
1589
1631
|
resources?: {
|
|
1590
1632
|
id: string;
|
|
1591
|
-
kind: "
|
|
1633
|
+
kind: "catalog" | "code_module" | "configuration" | "file";
|
|
1592
1634
|
path: string;
|
|
1593
|
-
|
|
1635
|
+
mediaType?: string | null | undefined;
|
|
1636
|
+
visibility: "policy_hidden" | "policy_visible" | "privileged";
|
|
1594
1637
|
required: boolean;
|
|
1595
1638
|
metadata: Record<string, unknown>;
|
|
1596
|
-
mediaType?: string | null | undefined;
|
|
1597
1639
|
}[] | undefined;
|
|
1640
|
+
metadata: Record<string, unknown>;
|
|
1598
1641
|
};
|
|
1599
1642
|
capabilities: {
|
|
1600
1643
|
schemaVersion: "openpond.tasksetCapabilities.v1";
|
|
1601
|
-
taskKind: "chat" | "
|
|
1602
|
-
supportedSignals: ("
|
|
1603
|
-
compatibleMethods: ("
|
|
1604
|
-
rewardKinds: ("
|
|
1644
|
+
taskKind: "chat" | "custom_program" | "multi_agent" | "single_agent";
|
|
1645
|
+
supportedSignals: ("correction" | "demonstration" | "feedback" | "label" | "preference" | "reward")[];
|
|
1646
|
+
compatibleMethods: ("dpo" | "grpo" | "none" | "opd" | "opsd" | "ppo" | "retrieval" | "sdft" | "sdpo" | "sft")[];
|
|
1647
|
+
rewardKinds: ("deterministic" | "exact" | "human" | "model_judge" | "none")[];
|
|
1605
1648
|
requiresTools: boolean;
|
|
1606
1649
|
requiresState: boolean;
|
|
1607
1650
|
requiresPrivilegedGrading: boolean;
|
|
1608
|
-
environmentPlacements: ("
|
|
1651
|
+
environmentPlacements: ("colocated" | "local" | "provider_native" | "remote")[];
|
|
1609
1652
|
exportable: boolean;
|
|
1610
1653
|
portabilityBlockers: string[];
|
|
1611
1654
|
};
|
|
1655
|
+
metrics?: {
|
|
1656
|
+
schemaVersion: "openpond.tasksetMetricPolicy.v1";
|
|
1657
|
+
primaryMetric: string;
|
|
1658
|
+
aggregation: "custom" | "mean_score" | "pass_rate" | "weighted_mean";
|
|
1659
|
+
missingReward: "exclude" | "zero";
|
|
1660
|
+
taskWeights?: Record<string, number> | undefined;
|
|
1661
|
+
customAggregator: {
|
|
1662
|
+
module: string;
|
|
1663
|
+
exportName: string;
|
|
1664
|
+
contentHash: string;
|
|
1665
|
+
timeoutMs: number;
|
|
1666
|
+
networkPolicy: "none";
|
|
1667
|
+
} | null;
|
|
1668
|
+
} | undefined;
|
|
1612
1669
|
tasks: {
|
|
1613
1670
|
schemaVersion: "openpond.taskData.v1";
|
|
1614
1671
|
id: string;
|
|
1615
1672
|
clusterKey: string;
|
|
1616
|
-
split: "
|
|
1673
|
+
split: "frozen_eval" | "test" | "train" | "validation";
|
|
1617
1674
|
input: Record<string, unknown>;
|
|
1618
1675
|
expectedOutput: Record<string, unknown> | null;
|
|
1619
1676
|
policyVisibleContext: Record<string, unknown>;
|
|
1620
1677
|
privilegedContextRef: string | null;
|
|
1621
1678
|
sourceRefs: string[];
|
|
1622
|
-
tags: string[];
|
|
1623
|
-
metadata: Record<string, unknown>;
|
|
1624
1679
|
assets?: {
|
|
1625
1680
|
id: string;
|
|
1626
1681
|
sourceRefId: string;
|
|
1627
1682
|
artifactRef: string;
|
|
1628
1683
|
fileName: string;
|
|
1684
|
+
stagingPath?: string | undefined;
|
|
1629
1685
|
mediaType: string;
|
|
1630
1686
|
sha256: string;
|
|
1631
1687
|
sizeBytes: number;
|
|
1632
|
-
split: "
|
|
1688
|
+
split: "frozen_eval" | "test" | "train" | "validation";
|
|
1633
1689
|
metadata: Record<string, unknown>;
|
|
1634
|
-
stagingPath?: string | undefined;
|
|
1635
1690
|
}[] | undefined;
|
|
1636
1691
|
resourceRefs?: string[] | undefined;
|
|
1637
1692
|
requiredOutputs?: {
|
|
1638
1693
|
path: string;
|
|
1639
1694
|
mediaType: string;
|
|
1640
|
-
metadata: Record<string, unknown>;
|
|
1641
1695
|
schemaRef?: string | null | undefined;
|
|
1696
|
+
submissionMode?: "structured_json" | undefined;
|
|
1642
1697
|
maxBytes?: number | undefined;
|
|
1698
|
+
metadata: Record<string, unknown>;
|
|
1643
1699
|
}[] | undefined;
|
|
1700
|
+
tags: string[];
|
|
1701
|
+
metadata: Record<string, unknown>;
|
|
1644
1702
|
}[];
|
|
1645
1703
|
graders: ({
|
|
1646
1704
|
id: string;
|
|
1705
|
+
taskIds?: string[] | undefined;
|
|
1647
1706
|
version: string;
|
|
1648
1707
|
label: string;
|
|
1649
1708
|
weight: number;
|
|
@@ -1651,11 +1710,11 @@ export declare function prepareImportedTasksetPackage(input: {
|
|
|
1651
1710
|
rewardEligible: boolean;
|
|
1652
1711
|
privileged: boolean;
|
|
1653
1712
|
metadata: Record<string, unknown>;
|
|
1654
|
-
kind: "
|
|
1713
|
+
kind: "content" | "diff" | "file" | "runtime_event" | "schema" | "state" | "test";
|
|
1655
1714
|
config: Record<string, unknown>;
|
|
1656
|
-
taskIds?: string[] | undefined;
|
|
1657
1715
|
} | {
|
|
1658
1716
|
id: string;
|
|
1717
|
+
taskIds?: string[] | undefined;
|
|
1659
1718
|
version: string;
|
|
1660
1719
|
label: string;
|
|
1661
1720
|
weight: number;
|
|
@@ -1666,15 +1725,15 @@ export declare function prepareImportedTasksetPackage(input: {
|
|
|
1666
1725
|
kind: "model_judge";
|
|
1667
1726
|
rubric: string;
|
|
1668
1727
|
judge: {
|
|
1669
|
-
providerId: "claude-code" | "codex" | "custom-openai-compatible" | "
|
|
1728
|
+
providerId: "anthropic" | "claude-code" | "codex" | "custom-openai-compatible" | "deepseek" | "google" | "grok-build" | "groq" | "moonshot" | "openai" | "opencode" | "openpond" | "openrouter" | "together" | "xai" | "zai" | `acp:${string}`;
|
|
1670
1729
|
modelId: string;
|
|
1671
1730
|
};
|
|
1672
1731
|
calibrationFixtureRefs: string[];
|
|
1673
|
-
calibrationStatus: "
|
|
1732
|
+
calibrationStatus: "failed" | "passed" | "pending";
|
|
1674
1733
|
temperature: number;
|
|
1675
|
-
taskIds?: string[] | undefined;
|
|
1676
1734
|
} | {
|
|
1677
1735
|
id: string;
|
|
1736
|
+
taskIds?: string[] | undefined;
|
|
1678
1737
|
version: string;
|
|
1679
1738
|
label: string;
|
|
1680
1739
|
weight: number;
|
|
@@ -1683,9 +1742,6 @@ export declare function prepareImportedTasksetPackage(input: {
|
|
|
1683
1742
|
privileged: boolean;
|
|
1684
1743
|
metadata: Record<string, unknown>;
|
|
1685
1744
|
kind: "human";
|
|
1686
|
-
rubric: string;
|
|
1687
|
-
reviewerRole: string;
|
|
1688
|
-
taskIds?: string[] | undefined;
|
|
1689
1745
|
form?: {
|
|
1690
1746
|
schemaVersion: "openpond.humanForm.v1";
|
|
1691
1747
|
mode: "individual" | "pairwise";
|
|
@@ -1696,72 +1752,75 @@ export declare function prepareImportedTasksetPackage(input: {
|
|
|
1696
1752
|
instructions: string;
|
|
1697
1753
|
required: boolean;
|
|
1698
1754
|
allowAbstain: boolean;
|
|
1699
|
-
kind: "score";
|
|
1700
|
-
minimum: number;
|
|
1701
|
-
maximum: number;
|
|
1702
|
-
step: number;
|
|
1703
1755
|
requireRationale?: boolean | undefined;
|
|
1704
1756
|
anchors?: {
|
|
1705
1757
|
value: string | number | boolean;
|
|
1706
1758
|
description: string;
|
|
1707
1759
|
}[] | undefined;
|
|
1760
|
+
kind: "score";
|
|
1761
|
+
minimum: number;
|
|
1762
|
+
maximum: number;
|
|
1763
|
+
step: number;
|
|
1708
1764
|
} | {
|
|
1709
1765
|
id: string;
|
|
1710
1766
|
label: string;
|
|
1711
1767
|
instructions: string;
|
|
1712
1768
|
required: boolean;
|
|
1713
1769
|
allowAbstain: boolean;
|
|
1714
|
-
kind: "category";
|
|
1715
|
-
options: string[];
|
|
1716
1770
|
requireRationale?: boolean | undefined;
|
|
1717
1771
|
anchors?: {
|
|
1718
1772
|
value: string | number | boolean;
|
|
1719
1773
|
description: string;
|
|
1720
1774
|
}[] | undefined;
|
|
1775
|
+
kind: "category";
|
|
1776
|
+
options: string[];
|
|
1721
1777
|
} | {
|
|
1722
1778
|
id: string;
|
|
1723
1779
|
label: string;
|
|
1724
1780
|
instructions: string;
|
|
1725
1781
|
required: boolean;
|
|
1726
1782
|
allowAbstain: boolean;
|
|
1727
|
-
kind: "boolean";
|
|
1728
1783
|
requireRationale?: boolean | undefined;
|
|
1729
1784
|
anchors?: {
|
|
1730
1785
|
value: string | number | boolean;
|
|
1731
1786
|
description: string;
|
|
1732
1787
|
}[] | undefined;
|
|
1788
|
+
kind: "boolean";
|
|
1733
1789
|
} | {
|
|
1734
1790
|
id: string;
|
|
1735
1791
|
label: string;
|
|
1736
1792
|
instructions: string;
|
|
1737
1793
|
required: boolean;
|
|
1738
1794
|
allowAbstain: boolean;
|
|
1739
|
-
kind: "text";
|
|
1740
|
-
maximumLength: number;
|
|
1741
1795
|
requireRationale?: boolean | undefined;
|
|
1742
1796
|
anchors?: {
|
|
1743
1797
|
value: string | number | boolean;
|
|
1744
1798
|
description: string;
|
|
1745
1799
|
}[] | undefined;
|
|
1800
|
+
kind: "text";
|
|
1746
1801
|
minimumLength?: number | undefined;
|
|
1802
|
+
maximumLength: number;
|
|
1747
1803
|
} | {
|
|
1748
1804
|
id: string;
|
|
1749
1805
|
label: string;
|
|
1750
1806
|
instructions: string;
|
|
1751
1807
|
required: boolean;
|
|
1752
1808
|
allowAbstain: boolean;
|
|
1753
|
-
kind: "preference";
|
|
1754
|
-
allowTie: boolean;
|
|
1755
1809
|
requireRationale?: boolean | undefined;
|
|
1756
1810
|
anchors?: {
|
|
1757
1811
|
value: string | number | boolean;
|
|
1758
1812
|
description: string;
|
|
1759
1813
|
}[] | undefined;
|
|
1814
|
+
kind: "preference";
|
|
1760
1815
|
requireStrength?: boolean | undefined;
|
|
1816
|
+
allowTie: boolean;
|
|
1761
1817
|
})[];
|
|
1762
1818
|
} | undefined;
|
|
1819
|
+
rubric: string;
|
|
1820
|
+
reviewerRole: string;
|
|
1763
1821
|
} | {
|
|
1764
1822
|
id: string;
|
|
1823
|
+
taskIds?: string[] | undefined;
|
|
1765
1824
|
version: string;
|
|
1766
1825
|
label: string;
|
|
1767
1826
|
weight: number;
|
|
@@ -1770,17 +1829,16 @@ export declare function prepareImportedTasksetPackage(input: {
|
|
|
1770
1829
|
privileged: boolean;
|
|
1771
1830
|
metadata: Record<string, unknown>;
|
|
1772
1831
|
kind: "custom_verifier";
|
|
1832
|
+
runtime?: "isolated_javascript" | "sandbox_process" | undefined;
|
|
1773
1833
|
module: string;
|
|
1774
1834
|
exportName: string;
|
|
1775
1835
|
timeoutMs: number;
|
|
1776
1836
|
networkPolicy: "none";
|
|
1777
|
-
taskIds?: string[] | undefined;
|
|
1778
|
-
runtime?: "isolated_javascript" | "sandbox_process" | undefined;
|
|
1779
1837
|
})[];
|
|
1780
1838
|
graderFixtures: {
|
|
1781
1839
|
id: string;
|
|
1782
1840
|
taskId: string;
|
|
1783
|
-
label: "
|
|
1841
|
+
label: "adversarial" | "boundary" | "infrastructure_failure" | "negative" | "positive" | "prompt_injection";
|
|
1784
1842
|
output: Record<string, unknown>;
|
|
1785
1843
|
infrastructureError: string | null;
|
|
1786
1844
|
expectedPassed: boolean;
|
|
@@ -1837,7 +1895,7 @@ export declare function prepareImportedTasksetPackage(input: {
|
|
|
1837
1895
|
metadata: Record<string, unknown>;
|
|
1838
1896
|
kind: "feedback";
|
|
1839
1897
|
feedback: string;
|
|
1840
|
-
polarity: "
|
|
1898
|
+
polarity: "mixed" | "negative" | "neutral" | "positive";
|
|
1841
1899
|
}[];
|
|
1842
1900
|
rewards: {
|
|
1843
1901
|
id: string;
|
|
@@ -1883,13 +1941,13 @@ export declare function prepareImportedTasksetPackage(input: {
|
|
|
1883
1941
|
authoringProvenance: {
|
|
1884
1942
|
schemaVersion: "openpond.taskAuthoringProvenance.v1";
|
|
1885
1943
|
model: {
|
|
1886
|
-
providerId: "claude-code" | "codex" | "custom-openai-compatible" | "
|
|
1944
|
+
providerId: "anthropic" | "claude-code" | "codex" | "custom-openai-compatible" | "deepseek" | "google" | "grok-build" | "groq" | "moonshot" | "openai" | "opencode" | "openpond" | "openrouter" | "together" | "xai" | "zai" | `acp:${string}`;
|
|
1887
1945
|
modelId: string;
|
|
1888
1946
|
} | null;
|
|
1889
1947
|
modelConfig: Record<string, unknown>;
|
|
1890
1948
|
skillHash: string;
|
|
1891
1949
|
promptTemplateVersion: string;
|
|
1892
|
-
buildIntent: "
|
|
1950
|
+
buildIntent: "demonstrations" | "discovery" | "preferences" | "rubric" | "verifiable_reward";
|
|
1893
1951
|
buildSpecification: {
|
|
1894
1952
|
kind: "demonstrations";
|
|
1895
1953
|
behavior: string;
|
|
@@ -1944,9 +2002,9 @@ export declare function prepareImportedTasksetPackage(input: {
|
|
|
1944
2002
|
tasksetId: string;
|
|
1945
2003
|
tasksetHash: string;
|
|
1946
2004
|
ready: boolean;
|
|
1947
|
-
recommendedMethod: "
|
|
2005
|
+
recommendedMethod: "dpo" | "grpo" | "none" | "opd" | "opsd" | "ppo" | "retrieval" | "sdft" | "sdpo" | "sft";
|
|
1948
2006
|
trainingPath: {
|
|
1949
|
-
primaryMethod: "
|
|
2007
|
+
primaryMethod: "dpo" | "grpo" | "opsd" | "ppo" | "sdft" | "sdpo" | "sft";
|
|
1950
2008
|
bootstrap: {
|
|
1951
2009
|
method: "sft";
|
|
1952
2010
|
purpose: "trajectory_bootstrap";
|
|
@@ -1955,9 +2013,9 @@ export declare function prepareImportedTasksetPackage(input: {
|
|
|
1955
2013
|
} | null;
|
|
1956
2014
|
} | null;
|
|
1957
2015
|
methodReadiness: {
|
|
1958
|
-
method: "
|
|
1959
|
-
status: "
|
|
1960
|
-
reasonCodes: ("
|
|
2016
|
+
method: "dpo" | "grpo" | "ppo" | "sft";
|
|
2017
|
+
status: "compatible" | "needs_dataset_work" | "recommended";
|
|
2018
|
+
reasonCodes: ("demonstrations_missing" | "executable_reward_missing" | "frozen_eval_missing" | "preference_pairs_invalid" | "preference_pairs_missing" | "reward_model_missing" | "reward_not_calibrated" | "taskset_not_ready" | "value_model_required")[];
|
|
1961
2019
|
reasons: string[];
|
|
1962
2020
|
}[];
|
|
1963
2021
|
compatibleDestinationClasses: ("custom" | "export" | "hosted_managed")[];
|
|
@@ -1978,60 +2036,6 @@ export declare function prepareImportedTasksetPackage(input: {
|
|
|
1978
2036
|
createdAt: string;
|
|
1979
2037
|
updatedAt: string;
|
|
1980
2038
|
metadata: Record<string, unknown>;
|
|
1981
|
-
profileRelease?: {
|
|
1982
|
-
id: string;
|
|
1983
|
-
contentHash: string;
|
|
1984
|
-
revision: number;
|
|
1985
|
-
} | null | undefined;
|
|
1986
|
-
datasetArtifact?: {
|
|
1987
|
-
schemaVersion: "openpond.datasetArtifact.v1";
|
|
1988
|
-
id: string;
|
|
1989
|
-
tasksetId: string;
|
|
1990
|
-
tasksetRevision: number;
|
|
1991
|
-
contentHash: string;
|
|
1992
|
-
format: "parquet";
|
|
1993
|
-
schema: {
|
|
1994
|
-
schemaVersion: "openpond.datasetSemanticSchema.v1";
|
|
1995
|
-
fields: {
|
|
1996
|
-
name: string;
|
|
1997
|
-
semanticRole: "metadata" | "feedback" | "messages" | "prompt" | "row_id" | "cluster_id" | "demonstration" | "chosen" | "rejected" | "expected_output" | "privileged_context" | "reward" | "tag";
|
|
1998
|
-
logicalType: "string" | "boolean" | "messages" | "integer" | "float" | "json";
|
|
1999
|
-
nullable: boolean;
|
|
2000
|
-
policy: "metadata" | "visible" | "privileged";
|
|
2001
|
-
}[];
|
|
2002
|
-
schemaHash: string;
|
|
2003
|
-
};
|
|
2004
|
-
shards: {
|
|
2005
|
-
id: string;
|
|
2006
|
-
split: "train" | "validation" | "test" | "frozen_eval";
|
|
2007
|
-
path: string;
|
|
2008
|
-
contentHash: string;
|
|
2009
|
-
schemaHash: string;
|
|
2010
|
-
sizeBytes: number;
|
|
2011
|
-
rowCount: number;
|
|
2012
|
-
rowGroupCount: number;
|
|
2013
|
-
}[];
|
|
2014
|
-
rowCount: number;
|
|
2015
|
-
splitCounts: Record<"train" | "validation" | "test" | "frozen_eval", number>;
|
|
2016
|
-
sourceReceiptRefs: string[];
|
|
2017
|
-
mappingHash: string;
|
|
2018
|
-
qualityReportHash: string;
|
|
2019
|
-
createdAt: string;
|
|
2020
|
-
} | null | undefined;
|
|
2021
|
-
metrics?: {
|
|
2022
|
-
schemaVersion: "openpond.tasksetMetricPolicy.v1";
|
|
2023
|
-
primaryMetric: string;
|
|
2024
|
-
aggregation: "custom" | "mean_score" | "pass_rate" | "weighted_mean";
|
|
2025
|
-
missingReward: "zero" | "exclude";
|
|
2026
|
-
customAggregator: {
|
|
2027
|
-
module: string;
|
|
2028
|
-
exportName: string;
|
|
2029
|
-
contentHash: string;
|
|
2030
|
-
timeoutMs: number;
|
|
2031
|
-
networkPolicy: "none";
|
|
2032
|
-
} | null;
|
|
2033
|
-
taskWeights?: Record<string, number> | undefined;
|
|
2034
|
-
} | undefined;
|
|
2035
2039
|
};
|
|
2036
2040
|
generatedFiles: {
|
|
2037
2041
|
path: string;
|