🜂 kompress.vaked.dev
learnable context compression · open science
kompress ∞
Asymmetric Loss Modulation Resolves the Voting Ensemble Paradox in Learned Context-Pruning Ensembles. An open-science manuscript: formalised, proven, and fixed — with every model, dataset and line of code in the open.
the manuscript
The Voting Ensemble Paradox
Under k-of-N drop voting the ensemble collapses to a single order statistic — its weakest member. Three mechanisms resolve it.
KOMPRESS v2 — the PDF ⤓
Geometric-Mean Consensus Distillation & Spectral Rigidity for Edge Reasoning. The earlier manuscript, the full derivation.
Read the write-up
You build an ensemble. You expect it to be better. It's worse. Formalised, proven, and fixed.
The loop shipped
17 models, 8 teachers, 4 architectures, $38.95 total — the whole experiment behind the paper.
the number
0.955heretic-exact @ 15% compression
0.993exact-keep rate
~78%token savings
149MModernBERT base
everything else
ModelsKompress v2 → v8 in the open, on Hugging Face.
Research repositoryHarnesses, baselines, run scripts, raw results.
LoopKitA loop-experiment-researcher template, so anyone can scaffold from here.
Eval spaceInteractive headroom evaluation.
Blogpocoo.vaked.dev — the loop, written down.
♥ supportKeep the loop running.