SegueWhat the model must do, the one metric that decides between candidates, and the target.
Review every 30 daysCurated public templates
ML project — current state
What a model-development project currently believes: what's current, what's settled, and what has been tried. Point to systems of record; don't copy them.
Your brief starts empty. Fill it with your own project’s answers.
Take a look inside
Sections in this template
Expand a section to see what goes inside.
Current state
7 sectionsYour latest answers, in one place.
Objective and deciding metric
Section instructions
Current best model
Section instructions
The checkpoint currently considered best on the canonical evaluation: its run or model ID, the deciding metric, and where the weights live. Link to the tracker; don't paste tables.…
Review every 14 daysComplete section instructions
The checkpoint currently considered best on the canonical evaluation: its run or model ID, the deciding metric, and where the weights live. Link to the tracker; don't paste tables. Never call a model best just because it is the most recent.
Canonical data version
Section instructions
The dataset version everyone should train and evaluate on, what changed from the previous one, and which versions are deprecated and why.
Review every 14 daysData sources and access
Section instructions
Where each dataset lives, who controls access, and known quality issues.
Review every 30 daysEvaluation protocol
Section instructions
The canonical evaluation set, how to run it, and known pitfalls such as leakage.
Review every 30 daysOpen questions and owners
Section instructions
What's unresolved, and who is expected to answer it.
Where to look
Section instructions
Repos, tracker projects, dashboards and documents, one line each.