File size: 4,190 Bytes
9bb1626 51ae9e4 9bb1626 51ae9e4 9bb1626 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 | {
"version": "local_joint_precision_v1",
"seed": 1729,
"steps": 4000,
"lr": 0.0001,
"weight_decay": 0.01,
"validation_interval": 500,
"initialization": {
"sha256": "0834dba67b83c7ccc1b26f3c8a964d1573fda410ebe0ea0a7cb545e756a13b8a",
"width": 64
},
"architecture": "Unchanged single local bat/rodent CNN; all parameters trainable, no scene veto or external model.",
"objective": "Per-example/group/polarity balanced local loss; BAT negatives weighted2;75% worst2% time-cell negative loss; +.25 same-location identity margin; weak TRAIN field recording MIL every second update weight.25. Identity margin disabled for synthetic co-occurrence.",
"sampling": "Three annotated bat contexts;two bounded noise contexts;three captive rodent contexts;three whole field noise contexts;one additional captive hard negative;one digital-silence control. Negative pools half parent-balanced hard mining and half parent-balanced uniform. No rank from held audio.",
"augmentation": "Same TRAIN gain -24..+12dB and continuous-tone augmentation as prior run;25% bat/rodent mixtures with conservative negative-mask intersection. One weak positive complete5sec bag every other step; no annotation/pseudo-model selects its witness.",
"optimizer": "AdamW100step warmup/cosine to10%;gradient norm5;EMA .99;fresh optimizer",
"selection": "Mean VAL focal-bat recall and whole-recording bat recall at lowest cut satisfying <=5% observed FPR separately on VAL reviewed-noise regions,captive-rodent contexts,and whole reviewed background recordings. VAL temporary cut selects weights only, never exported. Earliest tie, original base included as eligible fallback.",
"calibration": "After weight selection, lowest float32 BAT threshold with <=5% observed FPR in EACH corresponding CAL negative bucket. No recall guarantee; report achieved positive retention. Rodent threshold highest retaining90% CAL focal rodent regions, diagnostic only; one captive CAL parent.",
"controls": "Apply identical new CAL rule to unchanged base weights, alongside original base rule. Separates threshold-only changes from training. Never choose rule or checkpoint from transfer/private outcomes.",
"evaluation": "Known transfer cohorts were scored after freezing weights and calibration; not an untouched final test.",
"forbidden_training_calibration": [
"Held-out evaluation recordings",
"Calibration/validation used only for their assigned roles",
"Private user recordings"
],
"data_scope": "Reuse corrected base roles. Extra captive peak crops from existing hard-negative TRAIN rows only: BAT-negative context, no invented rodent-positive label. All weak human-verified TRAIN positive recordings supply MIL, not dense targets. No source held role is promoted.",
"limitations": [
"No new pup training collection is added. Existing pup VAL/test cannot be moved into training.",
"The5% budget is an observed CAL fraction, not a guaranteed deployment FPR. Background calibration exposure is short.",
"No fresh device-held final cohort; report history and sample counts."
],
"code_sha256": {
"local_joint_precision.py": "b9cf262a87af507b98323083344c95b24ca1fe43b90fb70ffd6f366fbe5e83fc",
"prepare_local_joint_precision.py": "56ddcb289333947567deca2bd3f9afad9ac28c52a196482b9fe5f26cc8d90976",
"train_local_joint_precision.py": "22357a7b6eba08ff93faa89ec6fdf31afee0e3cb211dea53ff2b2c9dfd7ea313",
"local_joint_field_model.py": "2523b722b38f83fab5e99cd8a23787508e76c407ac53b3c6a8cfda88ec9915db",
"train_local_joint_field_refined.py": "f5516ce28ea7e2586249ede0436dd7036dce5c08b0d138530ada8bfe6f439841"
},
"pre_gradient_amendment": {
"previous_protocol_sha256": "55bd8a4bcd2533dbc221d11c2c4f0431fe4158219e58a5c3472cb04e0c63260b",
"reason": "Auxiliary identity contrast requires explicit opposite-class absence; unknown labels must not acquire negative supervision.",
"weights_fitted_so_far": 0,
"mining_reused_sha256": "ea60e1f3ca5da41589487d3e525c2d94886206107532056408fd65ed1ff493ab"
},
"public_redaction": "Absolute paths and source identifiers omitted. Original frozen protocol SHA256 is recorded in calibration.json."
}
|