The Mix Report
Post-training is where the money is
Pretraining data is a commodity with an effectively unbounded supply. Post-training data is a manufactured good with a labour cost, and labs pay accordingly. The publicly documented post-training suites split into three distinct mixes — supervised demonstrations, preference comparisons, and reinforcement learning with verifiable rewards — and they are three different products with three different suppliers. Most data owners can supply exactly one of them and are pitching the wrong one.
The three mixes
| SFT | Preference (DPO and relatives) | RLVR | |
|---|---|---|---|
| Teaches | How to do the task at all — format, register, procedure | Which of two acceptable answers is better | To search for solutions that a checker will accept |
| Unit | Prompt → response, often with reasoning shown | Prompt → (chosen, rejected) pair with a reason | Task + environment + verifier, no gold answer required |
| Who can make it | Domain practitioners doing the work | Domain experts judging outputs | Engineers who can encode success as a check |
| Scale bought | 104–106 items | 104–105 pairs | 103–105 tasks |
| Priced by | Expert hours embodied | Judgement hours plus annotator calibration | Engineering hours plus task difficulty |
| Fails when | Demonstrations are inconsistent or idealised | Annotators disagree and disagreement is unmeasured | The checker is gameable |
What each one is really buying
SFT — behaviour under a distribution nobody else has
Supervised fine-tuning data is the closest thing to "recorded professional work" in this market. Its value is not the answers; it is that the prompts come from a real distribution of things people actually needed done, in the proportions they actually needed them. Synthetic prompt sets are cheap and plentiful; a real distribution of requests, with real edge cases in real proportions, is not manufacturable.
The premium sits on shown reasoning. A decision alone teaches the model to guess; a decision with the practitioner's path to it teaches the model a procedure. If your records contain the reasoning implicitly — a consultation transcript, a case file, an incident write-up — that is a higher grade of asset than a clean answer key, and it should be priced as one.
Preference data — calibrated judgement, and the disagreement
Preference mixes need comparisons where the label is defensible. The buyer's real question is not "which did your annotator prefer" but "how much do your annotators agree, and on what do they systematically disagree?" A preference set with measured inter-annotator agreement, kept disagreements, and documented rater qualifications is worth several times one delivered as a clean binary column.
This is the one mix where an organisation with expertise but no archive can still sell: judgement can be produced to order, against the buyer's own outputs, under contract.
RLVR — tasks with a checker, which is a different business
Reinforcement learning with verifiable rewards does not consume answers. It consumes problems plus a programmatic way to tell whether an attempt succeeded: a maths problem with a checkable result, a repository with a failing test, a workflow in a real application with a terminal state that can be inspected. The model generates, the verifier judges, the signal comes from the verifier rather than from a human label.
This is where the market is growing fastest and where the supply is thinnest, because it requires something most data owners have never built: an executable definition of success. It is also where the assets are worth the most per unit, and where the buyer's diligence is hardest — a gameable checker produces a model that games it. If your organisation runs a real application with real workflows, you are sitting on RLVR material and probably describing it as "logs". The RL gym series is the build guide.
Why the money concentrates here
Three structural reasons, all of which favour the seller:
- Supply is manufactured, not found. Nobody can scrape a preference set. The marginal cost of the next unit is an expert hour, so the price floor is a labour rate rather than a storage cost.
- Effect is measurable in weeks. A post-training set can be trialled on a small model and evaluated the same week. Measurable effect is purchasable effect; this is the stage where a contribution test actually settles arguments.
- The capability frontier moved here. Token counts in pretraining have flattened across successive open-weight releases while benchmark scores kept climbing. Whatever is producing those gains is downstream of pretraining, and downstream of pretraining is where budgets follow.
Which one are you?
| If you hold… | You are selling into… | Prepare |
|---|---|---|
| Archives of completed professional work with reasoning visible | SFT | Segment by task type; keep the reasoning; document who produced it |
| Practitioners who could judge outputs, but no archive | Preference | An annotation protocol, rater credentials, an agreement study |
| A real application with real workflows and observable end states | RLVR | Deterministic reset, a checker per task, and rights over the application itself |
| Bulk text in a specialist domain | Midtraining, not post-training | See midtraining |
Read next: The saturation board — which capability areas stopped improving, and are therefore the ones being bought into.