sentry-ai/nz-sovereign-mistral-lora

nz-sovereign-mistral-lora

Status: placeholder. There are no weights here yet. This repo is reserved for Step 5 of the Sovereign Agentic Pipeline, scheduled for November 2026. Nothing has been trained. Do not cite this repo as a result.

What this will be

A LoRA adapter over Mistral open weights, fine-tuned on New Zealand material, forming the model layer of a fully open-source sovereign AI pipeline: an AI system a New Zealand business can run end to end on New Zealand renewable compute, where no data leaves the country and every layer is inspectable and swappable.

The exact base checkpoint is not fixed yet and will be recorded here when it is. The licence above is provisional and will be reconciled with whatever base is chosen.

Why the adapter and not a full model

The binding constraint on the whole project: the moment one layer is a foreign API, the sovereignty claim dies, and so does the energy claim, because you no longer know what grid the compute ran on. Open weights plus a local adapter keeps every layer inspectable and runnable on hardware we can point at.

Publishing weights here is distribution, not hosted inference, so it does not breach that constraint.

Related

• —Training data: `sentry-ai/nz-sovereign-sft`
• —Held-out eval: sentry-ai/nz-sovereign-eval (private and protected, see below)
• —Code, methodology and the full research trail: https://github.com/JamesWeb3/sovereign-agentic-pipeline

The evaluation rule

The held-out eval set is never used for training and never tuned against. If Step 5 runs short on time, the dataset gets cut, never the eval set. A small dataset with a rigorous eval is a result; a large one with a rushed eval measures nothing.

Team

James Oldham (Sentry AI), engineering lead. Karl Belleza (@kbcoding13), researcher.