What we are actually finding out.

We are a small lab, so we publish the work rather than a roadmap. Everything below is something we built and measured, written up whether or not the result was flattering.

The work so far

Tino Rangatiratanga Evals

Most model evaluations are written somewhere else, for somewhere else. We built our own to ask a narrower question: how well does a model actually hold tikanga Māori, rather than recognise the vocabulary. The write-up covers how Tino Rangatiratanga Evals is built and where models fall over.

Post-training against Tino Rangatiratanga Evals

A benchmark is only worth having if you act on it. This is what happened when we post-trained a model against Tino Rangatiratanga Evals, including the parts that did not move.

Māori data and AI

The reason Tino Rangatiratanga Evals matters sits upstream of it, in who holds the data and who decides how it is used. Our reading of Māori data sovereignty, and the frameworks we work against, live in their own section.

Where it runs

We host our own infrastructure in New Zealand under New Zealand governance. That is a research constraint as much as a commercial one: it decides what data we are allowed to train on and what we can promise about where it stays.

Working with us on it

If you are researching in the same space, or you have a dataset where the sovereignty question is the hard part, we would rather talk early than read about it later.

Get in touch