Zovda
A question bank looks fine in a spreadsheet. Stems read cleanly, options run A to D, the answer column is filled. Then the model gets 92% on the held-out split and still cannot solve the same question when the wording changes.
The usual reason is answer leakage: traces of the answer left inside the question, or reasoning that only works because the answer was visible during training.
Leakage rarely looks like a bug
It shows up as small habits. The correct option is the longest one. The stem repeats a number that only appears in the right answer. Leftover numbering from the original source points at option B. A solution paragraph starts with "as computed above" when nothing above survives the export.
None of these break a validator, and all of them teach the model a shortcut. The model is not learning the subject; it is learning the artifact.
What a clean record carries
A usable question record keeps the parts separate and named:
- The stem itself, with no answer markers
- Structured options, one field per choice
- The answer, referencing an option rather than restating it
- The worked solution, step by step
- The knowledge points the item tests
- For vocational items, the category, trade, and question type
Separate fields matter because they let you build different training shapes from one source: answer-only for evaluation, solution-plus-answer for SFT, and stem-only for generation tasks.
Coverage beats cleverness
Nine school subjects across three stages, plus sixteen vocational categories, is not a marketing number — it is a filter. A model fine-tuned on one subject learns that subject's phrasing. A model that sees the same reasoning pattern across subjects learns the pattern.
That is also why we keep subject and stage as fields. Buyers slice by grade band or by trade instead of taking a package they have to re-label.
The checklist we run before delivery
- Does any stem contain a phrase that only makes sense if the answer is known?
- Do the options stand alone, without cross-references?
- Is the solution derivable from the stem, not from the answer key?
- Are held-out items truly held out, or copied from the training split?
- Can a buyer filter by subject and stage without opening every row?
Where to start
If you are training reasoning rather than recall, the question bank and reasoning collection is the closest fit. The corpus catalogue explains how the same items feed evaluation versus SFT.
Zovda assembles these question banks through Datuva, where every item is checked for leakage before it enters a collection and the check itself is logged.
Describe the subjects, stages, and volume you need. We reply within one business day, and data cards are shared after the scope is confirmed.