September 12, 2026 · 3 min read

How Our Data Approach Has Evolved

A short update on how we build training data — the approach has changed, and it's worth explaining plainly rather than letting it show up quietly in model cards.

What changed

Most models we've shipped so far — across Prism, Amethyst, and Copal — leaned on a teacher model generating synthetic examples as the entire training signal. That was never an absolute rule, though: Prism Roleplay 1.5 already blended real, human-written roleplay transcripts into the same training set alongside synthetic examples, so it's not accurate to say generated data was the only option we ever used. What's changing now is making that flexibility the explicit, default way we think about data, rather than an exception.

Going forward, we treat real (human-authored, human-collected) data and generated (synthetic, teacher-distilled) data as two ingredients, not one fixed recipe. Depending on the model and what it's meant to do, we'll:

• combine real and generated data in the same training run
• use generated data alone, when no suitable real dataset exists for the task
• use real data alone, when a clean human-authored dataset already covers exactly what's needed

The decision is made per model, based on what actually moves accuracy — not a standing policy that one source is always better.

Why

Synthetic data is cheap to produce and easy to steer toward a narrow task, which is exactly why it was the right call for single-purpose models like Prism's titling and safety classifiers. But it inherits the teacher's blind spots, and at scale it can quietly drift toward whatever patterns the teacher over-produces. Real data doesn't have that failure mode, but it's rarely shaped exactly like the task at hand, and for a lot of what we build, a matching real dataset simply doesn't exist. Treating them as interchangeable defaults — always synthetic, or always real — was leaving accuracy on the table in both directions.

What this looks like already

Prism Roleplay 1.5 is the “combine real and generated” case: real, human-written roleplay transcripts blended into the same training set as synthetic examples. The instruction-tuning stage of Quartz Micro Preview is the “real data alone” case: full-parameter SFT on databricks-dolly-15k, a human-written instruction dataset, with no synthetic generation involved at all. Most of the rest of the Prism, Amethyst, and Copal lineup is still synthetic-only for now, since that continues to fit narrow, single-purpose distillation well — but as Roleplay 1.5 already showed, that's a per-model choice, not a rule. Expect more releases to blend sources, or lean on just one, wherever that tests better.

Same standing rule as always: a held-out base-vs-tuned evaluation before we call anything done, regardless of where the training data came from.

— The Vertex AGI team