SmallDataRegime.ai

Research

Notes from the small-data regime.

Post-training gets more sample-efficient when the target is narrow, the negatives are real, and a held-out eval can say no.

01

The last mile is a small-data problem

Base models already cover the broad distribution. What remains is a narrow behavior: a format, a tool policy, a citation habit, a safety boundary. Those behaviors move on dozens or hundreds of carefully chosen examples, and they stall when the next thousand look like the first.

02

Yield, not volume

A preference is useful when the policy’s held-out win rate moves after it is included. SmallDataRegime.ai ranks new pairs by that marginal lift so a labeler spends the next hour on the slice that is still wrong, not on another easy win.

03

The spec is the governor

A reward model will exploit whatever you fail to write down. The spec — desired behavior, negative examples, and the evals that can veto a release — is the object teams review. The weights are a consequence.