| { |
| "application": { |
| "by_head_key": { |
| "boolean/state4": [ |
| "boolean", |
| "multilabel" |
| ], |
| "ordered/choiceS": "ordered", |
| "single/choiceR": "single" |
| }, |
| "choice": { |
| "branches": "|R single choice (single/choiceR), |S ordered choice (ordered/choiceS)", |
| "confidence": "the listed top-1 probability", |
| "expression": "probabilities = softmax(x[:n] / T)", |
| "note": "slice to the n listed options first, then divide by T. Reserved slots are never scored.", |
| "rule": "SLICE THEN TEMPER" |
| }, |
| "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.", |
| "idempotence": "apply exactly once; the returned probability and the listed top-1 derive from the same tempered read.", |
| "note": "Nouls and choices are tempered DIFFERENTLY. Implement exactly as written.", |
| "noul": { |
| "branches": "yes/no and every multi-label candidate (head_key boolean/state4, the merged head)", |
| "confidence": "max(P(yes), 1 - P(yes))", |
| "expression": "z = x[0] - logsumexp(x[1:]); P(yes) = sigmoid(z / T)", |
| "note": "x is the full four-letter logit vector. Do NOT compute softmax(x / T)[0].", |
| "rule": "COLLAPSE THEN TEMPER" |
| } |
| }, |
| "fit": { |
| "calibration_file_sha256": "ea069d224501af950caaae6914d6fdf14ce6d4d2bda566c41549b1e8c30a6222", |
| "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)", |
| "modality": "image rows use the same per-type scalar (no modality key in serving)", |
| "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", |
| "scored_heads_sha256": "96ea51416bbeb32d991b7b38d7f0c22ff3e82539c8910284c4fb1961f2869ace", |
| "source": "real development documents (held out from test), this model's BF16 scores, one NLL-minimising scalar per answer type" |
| }, |
| "fitted_on_model": { |
| "adapter_sha256": "d122466d430a058bb6457d919f811160e97fbd20149f4f24ca455c5d83e360a0", |
| "trained_heads_sha256": "f766d752d7768a419a9657155cf27f042834d9de29392cf7470d8725130e67ab" |
| }, |
| "frozen": true, |
| "models": { |
| "boolean": { |
| "applied": false, |
| "fit_units": 192, |
| "fitted_temperature": 0.8175095705097734, |
| "head_key": "boolean/state4", |
| "kind": "noul", |
| "task": "boolean", |
| "temperature": 1.0, |
| "unit": "question" |
| }, |
| "multilabel": { |
| "applied": false, |
| "fit_units": 1862, |
| "fitted_temperature": 0.842297230286191, |
| "head_key": "boolean/state4", |
| "kind": "noul", |
| "task": "multilabel", |
| "temperature": 1.0, |
| "unit": "candidate noul" |
| }, |
| "ordered": { |
| "applied": false, |
| "fit_units": 134, |
| "fitted_temperature": 1.2561869742268443, |
| "head_key": "ordered/choiceS", |
| "kind": "choice", |
| "task": "ordered", |
| "temperature": 1.0, |
| "unit": "question" |
| }, |
| "single": { |
| "applied": false, |
| "fit_units": 134, |
| "fitted_temperature": 1.107722547236206, |
| "head_key": "single/choiceR", |
| "kind": "choice", |
| "task": "single", |
| "temperature": 1.0, |
| "unit": "question" |
| } |
| }, |
| "schema": "solomon-readout-temperature-v3", |
| "sha256": "945bad449b7f5ffc88e597277d632fbab81c3c8729e22c8babd3f4a45fe1378b" |
| } |
|
|
