File size: 7,649 Bytes
0c1817e | 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 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 | {
"schema": "AutonomaFableRouterCurriculum.v2",
"status": "data_contract_frozen_training_not_authorized_nonrouting",
"nonRouting": true,
"trainingAuthorized": false,
"goalOrder": [
"capability",
"succinctness_and_termination",
"expert_efficiency",
"throughput"
],
"immutableSources": {
"host": {
"repo": "ermiaazarkhalili/LFM2.5-2.6B-SFT-Fable5",
"revision": "72d68fc"
},
"reasoning": {
"repo": "HelioAI/Claude-Fable-5-5500x",
"revision": "440267fbdb1b00a40216e7233dbce25530a0ed09"
},
"agentCoding": {
"repo": "greghavens/fable-5-coding-and-debugging-traces",
"revision": "c63e82adec30798edcbd6e1dcb0014d2b15de236"
},
"selectedBanks": {
"path": "config/benching/fable-donor-selected-banks.v1.json",
"sha256": "158689cca0679423925cfe3c2eefbd6a248045ed01d2f7dad4d74e05e3ad0d81"
}
},
"dataPolicy": {
"splitIdentity": {
"reasoning": "sha256(trimmed prompt)",
"agentCoding": "source_trajectory_sha256"
},
"splitBuckets": {
"train": "0..899",
"validation": "900..949",
"test": "950..999"
},
"reasoningDeduplication": "retain the shortest admitted non-empty completion per normalized prompt",
"historicalTrajectorySampling": "at most the first tool-producing prefix plus the final prefix",
"attestedTrajectorySampling": "retain all verifier-attested assistant decisions",
"maximumTokensBeforeTemplate": 1900,
"finalTemplateMaximumTokens": 2048,
"finalTrainerMustRetokenizeAndRejectOverflow": true,
"frozenEvaluationIsolation": "never read sealed/frozen evaluation prompts or answers while building or training"
},
"lanes": {
"host_preservation": {
"routerEligibility": "preserve",
"sftWeight": 1.0,
"baseFableKlWeight": 1.0,
"expertActivityTarget": "zero unless routed loss beats frozen host by the benefit margin"
},
"verified_expert": {
"routerEligibility": "expert",
"sftWeight": 1.0,
"baseFableKlWeight": 0.25,
"requires": "model_attested=true and non-empty verifier"
},
"interaction_pattern": {
"routerEligibility": "conditional",
"sftWeight": 0.1,
"baseFableKlWeight": 0.5,
"purpose": "tool topology and recovery only; not correctness supervision"
},
"loop_negative": {
"routerEligibility": "inherit_positive",
"preferenceWeight": 0.5,
"purpose": "reject repeated completions and duplicate identical tool calls"
}
},
"routingObjective": {
"counterfactual": "compare frozen-host token NLL with independently explored routed token NLL; never rely only on the router's current choice",
"offClass": "host-only is an explicit route target when no sampled expert helps",
"exploration": {
"expertScaleDuringDiscovery": "initialize at 0.005 and cap at 0.025 so incorrect trials cannot materially override the host",
"oracleProbeScales": [
0.025,
0.05,
0.1
],
"probeIsolation": "counterfactual probe branches are detached from the served/main output; only the low-scale chosen branch can affect host logits during discovery",
"candidateExpertsPerEligibleToken": 4,
"sampling": "stratified without replacement across steps and layers, independent of router logits; half uniform coverage and half frozen-profile-prior sampling",
"coverageBeforePruning": "every selected expert must receive counterfactual trials on validation-eligible token classes",
"forcedExplorationProbability": {
"initial": 0.25,
"final": 0.05,
"annealSteps": 300
},
"eligibleTokenEntropyBonus": {
"initial": 0.01,
"final": 0.0,
"annealSteps": 300
},
"minimumEligibleRouteMass": {
"initial": 0.15,
"final": 0.0,
"annealSteps": 300,
"scope": "expert-eligible tokens only; never host-preservation tokens"
},
"positiveDiscoveryReplay": "retain positive-benefit token/expert pairs in a bounded balanced replay buffer"
},
"routerSupervision": {
"target": "best sampled positive-benefit expert, with host-only off as a competing class",
"loss": "benefit-weighted ranking/classification plus end-to-end token NLL",
"negativeExperts": "sampled wrong experts teach relative ranking without requiring their high-impact activation",
"allOffProtection": "ranking targets and the temporary eligible-token route floor provide gradients even when current router logits prefer off"
},
"benefitMarginNats": 0.02,
"benefitMarginSchedule": "0.0 during discovery, anneal to 0.02 nats over steps 100..300",
"routeActivation": "after discovery, reward only when routed NLL improves beyond the annealed margin",
"routeCost": {
"expertScaleL1": 0.015,
"activeExpertPenalty": 0.005,
"targetActiveExpertsPerEligibleToken": "1..2",
"schedule": "zero for the first 100 discovery steps, then linearly anneal to full weight by step 300"
},
"conditionalLoadBalance": {
"weight": 0.01,
"applyOnlyTo": "expert-eligible tokens that pass the benefit gate",
"neverApplyTo": [
"host_preservation tokens",
"tokens where the host-only path is equal or better"
]
},
"antiDeadExpert": {
"measurement": "usage and marginal NLL improvement per layer, expert, lane, and donor bank",
"minimumUse": "no global quota; require coverage only among positive-benefit eligible tokens",
"pruneRule": "remove experts only after completed stratified exploration finds no validation-set marginal benefit; router non-use alone is not evidence",
"collapseGate": "reject checkpoints that choose off for all eligible tokens while the oracle replay buffer contains positive-benefit routes"
}
},
"behaviorObjective": {
"stopBoundaryLoss": 0.25,
"loopUnlikelihood": 0.5,
"duplicateToolCallPenalty": 0.75,
"maximumRepeatedNgramRateRegression": 0.0,
"maximumMedianAnswerLengthRegression": 0.05,
"maximumToolCallsAcrossFrozenSuite": 24,
"reject": [
"verbatim answer repetition",
"duplicate identical tool calls without an intervening result",
"continued reasoning after a valid terminal answer",
"expert activation without measured token-level benefit"
]
},
"curriculum": [
{
"stage": "A-host-anchor",
"steps": 200,
"mixture": {
"host_preservation": 0.75,
"verified_expert": 0.2,
"interaction_pattern": 0.05
},
"routerScaleMaximum": 0.025
},
{
"stage": "B-benefit-gated-routing",
"steps": 600,
"mixture": {
"host_preservation": 0.3,
"verified_expert": 0.6,
"interaction_pattern": 0.1
},
"routerScaleMaximum": 0.1
},
{
"stage": "C-succinct-calibration",
"steps": 200,
"mixture": {
"host_preservation": 0.55,
"verified_expert": 0.35,
"interaction_pattern": 0.1
},
"loopNegativeRatio": 0.25,
"routerScaleMaximum": 0.1
}
],
"checkpointSelection": {
"order": [
"verified coding and tool correctness",
"host capability non-regression",
"succinctness and clean termination",
"expert-use efficiency",
"decode throughput"
],
"rejectOn": [
"any sealed correctness regression beyond the predeclared tolerance",
"loop or duplicate-tool regression",
"median verbosity regression above 5%",
"expert traffic without positive marginal validation benefit"
]
}
}
|