DoccyHealth/Solomon

1{2 "application": {3  "by_head_key": {4   "boolean/state4": [5    "boolean",6    "multilabel"7   ],8   "ordered/choiceS": "ordered",9   "single/choiceR": "single"10  },11  "choice": {12   "branches": "|R single choice (single/choiceR), |S ordered choice (ordered/choiceS)",13   "confidence": "the listed top-1 probability",14   "expression": "probabilities = softmax(x[:n] / T)",15   "note": "slice to the n listed options first, then divide by T. Reserved slots are never scored.",16   "rule": "SLICE THEN TEMPER"17  },18  "granularity": "per answer type. The merged yes/no head serves two types, so by_head_key maps it to both and the served type selects the scalar.",19  "idempotence": "apply exactly once; the returned probability and the listed top-1 derive from the same tempered read.",20  "note": "Nouls and choices are tempered DIFFERENTLY. Implement exactly as written.",21  "noul": {22   "branches": "yes/no and every multi-label candidate (head_key boolean/state4, the merged head)",23   "confidence": "max(P(yes), 1 - P(yes))",24   "expression": "z = x[0] - logsumexp(x[1:]);  P(yes) = sigmoid(z / T)",25   "note": "x is the full four-letter logit vector. Do NOT compute softmax(x / T)[0].",26   "rule": "COLLAPSE THEN TEMPER"27  }28 },29 "fit": {30  "calibration_file_sha256": "ea069d224501af950caaae6914d6fdf14ce6d4d2bda566c41549b1e8c30a6222",31  "decision": "served at T = 1.0 for every type: the fitted scalars did not improve held-out calibration (test ECE worse in 8 of 10 type x modality cells; n-weighted 0.0212 fitted vs 0.0199 unscaled)",32  "modality": "image rows use the same per-type scalar (no modality key in serving)",33  "scored_heads_note": "scores were produced with a heads file whose two extra (entity/multilabel) slots were never read; its other 16 arrays are byte-identical to the shipped heads file, so the served logits are the fitted logits",34  "scored_heads_sha256": "96ea51416bbeb32d991b7b38d7f0c22ff3e82539c8910284c4fb1961f2869ace",35  "source": "real development documents (held out from test), this model's BF16 scores, one NLL-minimising scalar per answer type"36 },37 "fitted_on_model": {38  "adapter_sha256": "d122466d430a058bb6457d919f811160e97fbd20149f4f24ca455c5d83e360a0",39  "trained_heads_sha256": "f766d752d7768a419a9657155cf27f042834d9de29392cf7470d8725130e67ab"40 },41 "frozen": true,42 "models": {43  "boolean": {44   "applied": false,45   "fit_units": 192,46   "fitted_temperature": 0.8175095705097734,47   "head_key": "boolean/state4",48   "kind": "noul",49   "task": "boolean",50   "temperature": 1.0,51   "unit": "question"52  },53  "multilabel": {54   "applied": false,55   "fit_units": 1862,56   "fitted_temperature": 0.842297230286191,57   "head_key": "boolean/state4",58   "kind": "noul",59   "task": "multilabel",60   "temperature": 1.0,61   "unit": "candidate noul"62  },63  "ordered": {64   "applied": false,65   "fit_units": 134,66   "fitted_temperature": 1.2561869742268443,67   "head_key": "ordered/choiceS",68   "kind": "choice",69   "task": "ordered",70   "temperature": 1.0,71   "unit": "question"72  },73  "single": {74   "applied": false,75   "fit_units": 134,76   "fitted_temperature": 1.107722547236206,77   "head_key": "single/choiceR",78   "kind": "choice",79   "task": "single",80   "temperature": 1.0,81   "unit": "question"82  }83 },84 "schema": "solomon-readout-temperature-v3",85 "sha256": "945bad449b7f5ffc88e597277d632fbab81c3c8729e22c8babd3f4a45fe1378b"86}87