SmallDataRegime.ai

Docs

From a reward spec to a gated update.

Three commands cover the first run. Everything after that is review, yield, and whether the policy is allowed to leave the lab.

  1. 01

    Install

    > npm install -g smalldataregime

    The CLI talks to a cluster, a lab machine, or a local checkpoint directory.

  2. 02

    Point at a spec

    > smalldataregime spec check reward.md

    A spec needs a desired behavior, at least one negative, and a held-out eval command.

  3. 03

    Train inside a budget

    > smalldataregime train --budget 200 --reward prefs.jsonl

    The run stops at the preference budget or when the KL cap trips, then waits on the eval gate.