Skip to content
Optimal Lab

AI Engineering · interactive

Synthetic data

Let a model write your training examples and you get thousands in an hour. Train on them, generate again from the new model, and two things creep in: the rare cases disappear, and errors feed on themselves. Run eight generations, then mix real data back in and verify what you keep.

Real data

0% mixed in

Model

gen 0 · 40 topics

Generate

2,000 examples

Verify

off

Train the next

gen 1

topics in generation 0 · 40 of 40

2% wrong

common topics ← → rare topicsgrey: the real data

topics left
40/40
top 5 topics
53%
wrong examples
2%

Generation 0 is the real data: all 40 topics, with a long tail of rare ones, and 2% of examples wrong.

generation 0 of 8

after 8 generations: 3/40 topics, 60% wrong

  1. 1

    Real support tickets cover forty topics, a few of them often and most of them rarely. A few are mislabelled: real data isn't clean either.

    40 topics · top 5 = 53% of tickets · rarest topic 0.6%

  2. 2

  3. 3

  4. 4

  5. 5

In production

Tools and technologies

Synthetic data pipelines generate, filter and verify; the research names the trap.

Not tied to one language — use these names when you search docs or design a system.