Nine Million Questions: The Quiet Market for Exam-Question Datasets

The hottest training material in EdTech AI isn’t essays — it’s exam questions: millions of them, structured, answer-keyed, and licensable. Study apps, tutoring copilots, and eval-hungry model teams all draw from the same scarce resource, and almost nobody outside the niche knows a market exists. Three use cases below explain who buys question banks, why building one yourself is a copyright lawsuit with extra steps, and what “9 million questions” inventory actually means.

Who Buys Questions by the Million

Use case 1: The eval team. Model teams building education copilots need benchmarks that aren’t contaminated — and every public benchmark is, by definition, already in everyone’s training data. A question answer dataset for LLM training, freshly sourced and structured with answers and explanations, becomes a private eval: the model is graded on questions it provably hasn’t seen. The eval use case is why the same dataset gets bought by teams with no tutoring product at all.

Use case 2: The fine-tuning team. Tutoring copilots need QA pairs in the style of real curricula — not synthetic paraphrases but actual exam phrasing, distractor structure, and difficulty gradients. Fine-tuning on licensed question banks produces models that speak the language of the exam room; fine-tuning on synthetic questions produces models that speak the language of a language model.

Use case 3: The product team. Practice-test features, spaced-repetition drills, diagnostic quizzes — content pipelines that used to run on licensed textbook arrangements now run on structured question data. A medical exam questions dataset becomes a USMLE prep feature; a vocational exam question dataset becomes a trade-certification app; the same schema powers both.

Why Nobody Builds Their Own

The naive path — scrape past papers from the open web — runs into the three problems that define this niche:

The licensed-dataset path resolves all three at once: provenance per record, structured fields (question, options, answer, explanation, subject, difficulty where available), and a commercial-use license in writing. Thordata’s datamall lists this at a scale few buyers expect — a 9-million-question bank spanning general exams, with a 400,000-question medical exam vertical inside it — delivered in the catalog’s standard formats (JSON/CSV/Parquet) with sample fields viewable before purchase, standard datasets starting around $0.25 per 1,000 records.

The Structure That Makes a Bank Usable

The difference between a question dump and a usable exam question bank dataset is the schema:

question_id        stable key across releases
subject / topic    curriculum alignment (the field tutors actually ask about)
question_text      with options where the format is multiple choice
answer             verified, not majority-voted
explanation        present for a large fraction — the tutoring payload
difficulty         where available — the eval-team payload
source / license   provenance that survives a procurement review

The last line is the quiet differentiator. Education companies get acquired; acquisition diligence asks where the training content came from, and “the open web” is the answer that kills deals. A per-record license scope turns that diligence question into a five-minute answer.

FAQ

Can we license the medical question bank for a USMLE product?
That’s the flagship use case the 400K medical vertical exists for — but check jurisdiction scope with the vendor when scoping: medical licensing exams differ by country, and the license should name the markets you ship in.

How do question banks feed LLM training versus eval?
Split the purchase: a fine-tuning shard (question + answer + explanation as instruction-style pairs) and a held-out eval shard never touched in training, refreshed as the vendor updates the bank. The same license covers both; the discipline of splitting them is yours.

What about answer accuracy — who guarantees it?
Ask for the verification methodology and the error rate, and audit a random sample before fine-tuning. Any serious vendor expects the audit; the sample-first flow exists so you can run it before the invoice.

Is a 9-million-question bank overkill for a tutoring startup?
It’s priced like a component, not a commitment — per-record pricing means a focused 100K-question slice is a pilot, and difficulty/subject fields let you license exactly your vertical. Start narrow (one subject, two grade bands), measure learning outcomes, and expand the license with the product.

Does search data belong in an EdTech data strategy?
Very much — twice. Once for product: education queries are intensely seasonal and exam-specific, and SERP monitoring at about $0.70 per 1,000 structured responses shows exactly which exams your market is anxious about, week by week. And once for visibility: study content ranks or it doesn’t exist, and continuous SERP data crawling tracks where your practice pages surface against the exam boards themselves. The exam calendar drives your users’ year; the search feed is how you read it — and query monitoring on your top exam keywords is the cheapest product-research tool an EdTech team can run.

The niche in one sentence: exam questions became an industrial input the day tutoring AI got serious, the serious buyers all license rather than scrape, and the inventory to license from — nine million questions deep — is sitting on a shelf most of the industry hasn’t looked at yet.