Data
Dataset validation
What is dataset validation?
Dataset validation checks a training file before any compute is leased: that it parses, that every record has the required fields, that sequences fit the context window, and that duplicates and empty targets are flagged. It is the cheapest possible place to catch an error.
The economics are stark. A malformed dataset discovered by validation costs nothing; the same dataset discovered three hours into a training run costs three hours of GPU time and the time it takes a person to work out what went wrong.
The checks worth running are unglamorous: encoding, field presence, token length distribution, duplicate rate, class balance, and the proportion of examples that would be truncated. Truncation in particular is silent and common.
Related terms
Start with the free tier
A magic link creates your account, your tenant and your first API key. No card until you ask for compute.