Back to projects

Experimental project

TensionLM

The earlier model-science line testing sigmoid pairwise tension attention, telemetry, curriculum behaviour, and bounded formal-domain signals.

Claim discipline: Claims on this page should be treated as hypotheses unless linked to a receipt, benchmark, or reproducible artifact.
Token pairs
Sigmoid tension field
Value aggregation
Inspectable telemetry

What problem it addresses

Standard attention offers strong capability but does not directly expose an independent pairwise pressure field with the intended TS interpretation.

How it works

Replace or compare softmax attention with sigmoid pairwise tension, export telemetry, train small models, and evaluate matched tasks.

What exists now

A valuable previous-generation evidence base; Ten-SON-LM is the clearer current architecture question.

Evidence / receipts

The repository contains runnable models, public artifacts, small-scale matched comparisons, curriculum experiments, and formal-domain runs with both positive and negative outcomes.

Limits

It has not established a general capability advantage over standard attention. Older strong claims should be read under the newer, narrower public evidence boundary.

Next milestone

Keep the repository stable as a baseline and evidence archive rather than adding another loosely scoped model branch.

How to run / inspect

Install / open

git clone https://github.com/BoggersTheFish/TensionLM && cd TensionLM

Inspect

Read the public evidence boundary first, then inspect matched softmax comparisons, formal-domain checkpoints, and failed or regressed final runs.

Links