What makes a dataset worth fine-tuning on

What most enterprise archives are missing

A support ticket, a strategy deck, a wiki page: each records what someone wrote, once, usually after the fact. None of them say whether the choice behind it worked. Train a model on that and it learns to sound like the archive, fluent and confident, without being any better informed about which of the choices inside it actually paid off. The gap is not something text cleaning fixes. Most archives capture intention and description. A training-grade example needs the other half: not just what was decided, but what happened as a result, measured against what was expected.

The five things a training-grade example needs

Strip a genuinely useful example down and five things are present, every time.

• A real input: the situation as it actually arrived, not a summary written afterward for training.
• A real decision: the call that was actually made, not the one that reads best in hindsight.
• The expected result: written down before the outcome existed, so it cannot be invented later.
• The measured outcome: what actually happened, checked against that expectation rather than assumed.
• A human judgment: a second, different person confirming what the example teaches, so a mistake is never graded as a lesson by accident.

Remove any one of the five and the example teaches something narrower: a style, a vocabulary, a plausible-sounding answer. Keep all five and it teaches a decision.

Fine-tuning is no longer a research-lab project

The reason the question matters now, and did not as much five years ago, is cost. Fine-tuning an open-weight model on domain data is estimated to cost 1,000 to 10,000 times less than training a model from scratch, and reaches production in two to six months rather than the multi-year cycle a foundation model takes. The enterprise market for these deployments was estimated at around 7.57 billion dollars in 2026. The economics changed. The question of what to fine-tune on did not go away. It got more important, because a cheap process run on the wrong data is still a cheap way to reach a worse model.

Volume versus verification

The 2026 consensus among the people who build these models is not what it was two years earlier: curated, expert-verified data beats volume, and some models now reach top results with a hundred times less data than earlier generations needed. That tracks with what a graded record actually contains. A few thousand verified decisions in one domain, each with its expectation, its outcome and a named approval, teach a model more about that domain than a million pages scraped from wherever they could be found. Verification is the expensive part, and it was never for sale.

Why the well is running dry

The alternative source is narrowing too. Researchers at Epoch AI project that the stock of high-quality public text will be exhausted somewhere between 2026 and 2032. What increasingly replaces it on the open web is machine-written, and that creates a second problem: in a study published in Nature in 2024, Shumailov and colleagues showed that models trained repeatedly on their own output degrade across generations instead of improving. Scraping more of the web is not a growing supply anymore. It is a shrinking, increasingly self-referential one.

Whose decision this is

None of this makes the choice for you. A record built from graded decisions has the shape fine-tuning needs: real inputs, real outcomes, a human judgment on each one. But whether it is ever used that way is the owner's decision, not a default the system makes on your behalf, and today that default is no. If you later publish an approved lesson to the Almanexa Hub, a standard pack grants organizational use only. Training rights exist solely where a publisher explicitly grants them on a separate, priced offer, recorded in the signed license, and nothing changes that quietly.

Read more about the record as training data, or see what decision memory is.