Hurricane Refinery v2
This is the stored snapshot for the approved document version. The diff below shows what changed from the previous version.
Document snapshot
Hurricane Refinery
Fine-tune models on your own GPU hardware, from the Sentinel console.
Point Refinery at a machine you control, give it a corpus, and it trains a LoRA/QLoRA adapter there inside a pinned container. Your data, your compute, your weights.
- HOWITWORKS.md — the architecture and the decisions behind it
- SECURITY.md — the trust boundary; read this before adding a target
- HELPDESK.md — setup and what to do when something breaks
The short version
A training job outlives the process that starts it, so jobs are durable files that reconcile against the target rather than entries in memory. A corpus is validated and scanned for credentials before it can leave this machine, and the receiver hashes what actually arrived. Weights stay on the target in v1.
Drop a .jsonl onto the Datasets tab to bring a corpus in. The training image is
pinned by digest — so what produced a set of weights is always recorded — and
Refinery tells you when that pin is out of date, fetches the new one, and offers
to clear the old ones out.
Training runs on hardware you supply, and Sentinel's sandbox does not extend to it. That is stated deliberately rather than glossed: see SECURITY.md.
Diff from previous
--- README@1 +++ README@2 @@ -17,5 +17,10 @@ validated and scanned for credentials before it can leave this machine, and the receiver hashes what actually arrived. Weights stay on the target in v1. +Drop a `.jsonl` onto the Datasets tab to bring a corpus in. The training image is +pinned by digest — so what produced a set of weights is always recorded — and +Refinery tells you when that pin is out of date, fetches the new one, and offers +to clear the old ones out. + **Training runs on hardware you supply, and Sentinel's sandbox does not extend to it.** That is stated deliberately rather than glossed: see SECURITY.md.