FluidInference/verdict-coreml

1{2  "source_repo": "heman10x/rlcd-modernbert-151m",3  "source_revision": "8af2496eb63c7fa66d7d234e1f62629380030eb4",4  "package": "verdict_fp16_L128_candidates25.mlpackage",5  "package_files_sha256": {6    "Data/com.apple.CoreML/model.mlmodel": "6cc079930a63c1069a7a2c259f4c086879981021047767f2bb4d68b6d1f02b00",7    "Data/com.apple.CoreML/weights/weight.bin": "a982e14147982df9abb99a3637075a0193835c6b99edc396a52c1a43c2a4854a",8    "Manifest.json": "fc5d31df174a9f450bfb9ba406307c379c30fabba5cad38fcaf8717b806c69a8"9  },10  "chip": "arm",11  "macos": "27.0",12  "max_calibrated_probability_error": 0.003407791923494219,13  "selected_id_agreements": 4,14  "questions": 4,15  "rows": [16    {17      "name": "lost card",18      "type": "choice",19      "tokens": 53,20      "candidates_including_abstention": 3,21      "native": {22        "type": "choice",23        "selected_id": "card_lost",24        "selected_probability": 0.627995456676331,25        "probabilities": {26          "card_lost": 0.627995456676331,27          "pin_reset": 0.12699368118294335,28          "__insufficient_evidence__": 0.245010862140725629        },30        "concentration": 0.18185863532617996,31        "is_abstention": false,32        "p_abstain": 0.2450108621407256,33        "calibration_status": "calibrated_for_scope",34        "choice": "card_lost"35      },36      "coreml": {37        "type": "choice",38        "selected_id": "card_lost",39        "selected_probability": 0.6283802556009165,40        "probabilities": {41          "card_lost": 0.6283802556009165,42          "pin_reset": 0.1261597182435837,43          "__insufficient_evidence__": 0.2454600261554998344        },45        "concentration": 0.18269014752444745,46        "is_abstention": false,47        "p_abstain": 0.24546002615549983,48        "calibration_status": "calibrated_for_scope",49        "choice": "card_lost"50      },51      "max_calibrated_probability_error": 0.0008339629393596393,52      "selected_id_agrees": true,53      "median_model_call_ms": 3.60970804467797354    },55    {56      "name": "binary inquiry",57      "type": "noul",58      "tokens": 50,59      "candidates_including_abstention": 3,60      "native": {61        "type": "noul",62        "selected_id": "true",63        "selected_probability": 0.46928461522757403,64        "probabilities": {65          "true": 0.46928461522757403,66          "false": 0.24040802067067693,67          "__insufficient_evidence__": 0.2903073641017490768        },69        "concentration": 0.03808313508410699,70        "is_abstention": false,71        "p_abstain": 0.29030736410174907,72        "calibration_status": "calibrated_for_scope",73        "noul": 0.6612505068952915,74        "selected_outcome": "true"75      },76      "coreml": {77        "type": "noul",78        "selected_id": "true",79        "selected_probability": 0.4658768233040798,80        "probabilities": {81          "true": 0.4658768233040798,82          "false": 0.2409705632084786,83          "__insufficient_evidence__": 0.2931526134874416584        },85        "concentration": 0.036521350138502884,86        "is_abstention": false,87        "p_abstain": 0.29315261348744165,88        "calibration_status": "calibrated_for_scope",89        "noul": 0.6590911025400003,90        "selected_outcome": "true"91      },92      "max_calibrated_probability_error": 0.003407791923494219,93      "selected_id_agrees": true,94      "median_model_call_ms": 3.403667011298239295    },96    {97      "name": "urgency score",98      "type": "score",99      "tokens": 55,100      "candidates_including_abstention": 4,101      "native": {102        "type": "score",103        "selected_id": "__insufficient_evidence__",104        "selected_probability": 0.3259147064757007,105        "probabilities": {106          "low": 0.1979320574725608,107          "medium": 0.289823573347529,108          "high": 0.18632966270420953,109          "__insufficient_evidence__": 0.3259147064757007110        },111        "concentration": 0.02039159311536931,112        "is_abstention": true,113        "p_abstain": 0.3259147064757007,114        "calibration_status": "calibrated_for_scope",115        "score": null,116        "selected_level_id": "__insufficient_evidence__",117        "selected_value": null118      },119      "coreml": {120        "type": "score",121        "selected_id": "__insufficient_evidence__",122        "selected_probability": 0.32674006867380345,123        "probabilities": {124          "low": 0.19782407464261623,125          "medium": 0.2888377991986832,126          "high": 0.18659805748489713,127          "__insufficient_evidence__": 0.32674006867380345128        },129        "concentration": 0.020407770698545158,130        "is_abstention": true,131        "p_abstain": 0.32674006867380345,132        "calibration_status": "calibrated_for_scope",133        "score": null,134        "selected_level_id": "__insufficient_evidence__",135        "selected_value": null136      },137      "max_calibrated_probability_error": 0.0009857741488458194,138      "selected_id_agrees": true,139      "median_model_call_ms": 3.3971249940805137140    },141    {142      "name": "insufficient evidence",143      "type": "choice",144      "tokens": 41,145      "candidates_including_abstention": 3,146      "native": {147        "type": "choice",148        "selected_id": "__insufficient_evidence__",149        "selected_probability": 0.6003394036234859,150        "probabilities": {151          "visa": 0.19581438090218925,152          "mastercard": 0.20384621547432477,153          "__insufficient_evidence__": 0.6003394036234859154        },155        "concentration": 0.1354395684855455,156        "is_abstention": true,157        "p_abstain": 0.6003394036234859,158        "calibration_status": "calibrated_for_scope",159        "choice": "__insufficient_evidence__"160      },161      "coreml": {162        "type": "choice",163        "selected_id": "__insufficient_evidence__",164        "selected_probability": 0.5997858326963917,165        "probabilities": {166          "visa": 0.19632166854967723,167          "mastercard": 0.20389249875393106,168          "__insufficient_evidence__": 0.5997858326963917169        },170        "concentration": 0.1348775837045153,171        "is_abstention": true,172        "p_abstain": 0.5997858326963917,173        "calibration_status": "calibrated_for_scope",174        "choice": "__insufficient_evidence__"175      },176      "max_calibrated_probability_error": 0.0005535709270941824,177      "selected_id_agrees": true,178      "median_model_call_ms": 3.4325000015087426179    }180  ],181  "passed": true182}183