WEBVTT

00:00:00.000 --> 00:00:08.799
Here's a sneaky training data bug. These two traces both use call seventeen. That's fine. They're separate conversations.

00:00:08.799 --> 00:00:18.312
But trace C calls eighteen, and its result points to seventeen. Searching for that ID anywhere in the dataset would miss the problem.

00:00:18.312 --> 00:00:27.831
Check the pair: the trace, and the tool call. Keep the source file and row, so you can get back to the exact record.

00:00:27.831 --> 00:00:39.979
Our tiny mock example checks JSON Lines, Parquet, and Arrow. Inspect the records locally in Row Cove, and use your trainer's own validation too.