A Full QC Pipeline for RL Data: Intake Review, Active Testing, Packaging

joecole · x · 2026-10-14

Sean Cai publishes "QC for RL Data," detailing the full quality-control process for evaluating RL environment datasets produced by third-party vendors before they are presented to frontier labs as off-the-shelf (OTS) training data.

The process has three stages:

Intake review — documentation check; datasets that fail do not advance.
Active testing — post-train a small model and probe it.
Packaging — format for lab delivery.

The article stresses that QA plans must be scoped by training method:

SFT: the unit of data is a prompt-response pair; the minimum pipeline is intake review, log-probability shift analysis, and a catastrophic forgetting check. Active testing with rollouts and graders doesn't apply because the data itself is the target behavior rather than a reward signal to be hillclimbed, so required artifacts shrink (no grader, no rubric, no vendor-run post-training rollouts).
Different training architectures impose different minimum QA requirements; applying the full pipeline to data it was never designed for generates spurious findings and strains the vendor relationship.

The author also ships an impromptu skill that gives agents context on these design principles, updated alongside the article.

Original post →

A Full QC Pipeline for RL Data: Intake Review, Active Testing, Packaging

More from coding & agent