vm-ranker can now compute the value model itself (off by default)
vm-ranker/gets its ownparams.rswith 44param!defaults. These include 27 ranking weights that mirror home-mixer's names, keys and current defaults (all of home-mixer's weights exceptContActiveSecs5mResidualNormWeight; e.g.FavoriteWeight= 0.5,ReplyWeight= 5.0), plus DPP settings (DppEnabledon,DppTheta= 0.65,DppMaxSelectedRank= 150) and author-diversity settings (EnableAuthorDiversityon,AuthorDiversityDecay= 0.5,AuthorDiversityFloor= 0.25).- New
vm-ranker/scoring/value_model.rsbuilds value-model weights from those params (weights_from_params). It is gated byComputeValueModel, which defaults to off. Newranking_config.rsandconfig_sync.rsfiles are added alongside. - Home-mixer side: new
vm_ranker_debias_payload.rsand aVMRankerSendDebiasInputsflag (off).PhoenixMOEMaxResults200→0andPhoenixMoeColdStartMaxResults0→200.PhoenixRetrievalMOEInferenceClusterId"Experiment3Memy04"→"Experiment2Memy04". - Grox:
GROK_GEMMA_FOLLOWER_SPLIT200_000→225_000.
Full analysis → Summary: hand-written notes, checked against the diff