Six categories, ranked by what's actually moving right now — not by what's easiest to produce.
| Rank | Category | What it is | Demand |
|---|---|---|---|
| 1 | Agentic / tool-use data | Full multi-step trajectories — a system pulling data, checking rules, taking an action — with a labeled success/failure outcome. | 100 |
| 2 | Preference / RLHF data | Two or more responses to the same situation, ranked by quality. This is the fuel for continued post-training. | 88 |
| 3 | Evaluation / benchmark data | Hard cases with a verified correct answer — especially ones public benchmarks don't cover. | 78 |
| 4 | Domain corpora unavailable on the open web | Text that simply doesn't exist anywhere else — internal jargon, proprietary process documentation, industry-specific patterns. | 64 |
| 5 | Supervised fine-tuning (SFT) data | Clean instruction → correct-response pairs. Still useful, increasingly commoditized for generic tasks. | 40 |
| 6 | Retrieval / RAG grounding data | Documents plus verified Q&A pairs used to build or evaluate retrieval systems. | 28 |
Every lab has scraped the open web already — what's left has to come from somewhere real. Agentic and preference data win because they encode judgment, not just facts: a model can read a textbook on its own, but it can't observe how a real team decided a borderline edge case, or which of two draft responses an expert preferred and why.
RAG data ranks lowest not because it's not useful, but because most buyers can assemble it themselves cheaply from their own or licensed corpora — it's the one category on this list where a data seller isn't the only possible source.