Data Is A Pipeline, Not A Number

When someone asks how much training data we have, the honest answer is another question: for what, and how hard does it need to be? A single number is the wrong unit. The data that’s worth anything gets made on demand, shaped to a specific domain, and graded before it ever ships.

The big number is a tell

“We have three million tasks” tells you almost nothing on its own, because it says nothing about where they came from. A hundred tasks that are hand-checked, correct for the domain, and hard in the places that matter will beat a million scraped-and-templated near-duplicates, and it won’t be close.

What compounds is the machine that makes them

The machine has to do three jobs well. The first is generation: producing fresh tasks in a domain and letting you dial the difficulty on purpose. Grading is where most setups quietly fall down, because rubber-stamping whatever the model did is so much easier than actually separating the good runs from the bad ones. Then there’s provenance, which nobody wants to think about until it bites them: knowing where every task came from, and whether the model has already seen it in training. Miss any one of the three and the other two start lying to you.

Get that machine right and the “how much data” question answers itself: as much as the buyer needs, freshly made, with a paper trail attached. The count stops being the pitch and turns into a side effect, which is roughly the moment you stop sounding like everyone else selling data by the pound.