KET Language Modeling Details

This detail layer keeps the experiment visible but subordinate to the story: KET is first demonstrated as a sequence-modeling architecture, then reused as the categorical transport principle behind foundry and UDL examples.

Experiment Provenance

These perplexity curves are experiment artifacts, not schematic drawings. They come from KET language-modeling comparison runs over PTB and WikiText-style benchmarks, with validation perplexity logged by the runner and rendered into the plot files shown below.

Paper anchor

Kan Extension Transformers arXiv paper

Supplementary source bundle

README for runnable KET language-modeling code

Plot-generating runner

LM causal comparison script

Demo plot generator

Compact attention/KET/TopoCoend comparison script

Reproducibility index

Experiment settings, artifacts, and rerun commands

Language Modeling Evidence

KET versus attention validation perplexity curve. — KET versus attention: a compact entry point into the language-modeling behavior.

Attention, KET, and TopoCoend validation perplexity curves. — Attention, KET, and TopoCoend variants: the detail view for comparing transport-style architectures.

Why This Matters For The Rest Of The Tutorial

In language modeling, KET transports predictive structure across token contexts. In Odyssey, the transported object becomes a foundry artifact: a claim, skill, dashboard, or causal cell. In UDL, the transported object becomes decision behavior: rollout extends a local policy, while pullback diagnostics check whether the extension remains compatible with gates and evidence.