Find the first boundary where meaning disappears
Trace a failed example backwards from the record your trainer consumed. Can the reviewer locate its source passage, original table or native object? Does the extracted value match that source? Does the instruction actually ask for the output being taught? Stop at the earliest mismatch rather than adjusting model settings around corrupted input.
Keep separate counts for delivered files, readable files, extracted objects and accepted task records. Those are different preparation stages. An archive can be complete while most candidate examples remain unsuitable for your chosen task. Agree on which stage the supplier delivers and which stages your own team must complete.
A parser can succeed while the content becomes wrong
Inspect extraction according to the source structure. A specification may carry an exception in a footnote, a table may depend on merged headings, and a drawing may express a relationship through placement. Plain text can preserve the words while losing their meaning. Native object extraction can retain properties while omitting linked references or application-specific behavior.
Build a review set that exercises those structures. Compare converted records with the file opened in an appropriate viewer. Record missing rows, detached headings, unresolved references and omitted annotations separately. A successful process exit or a nonempty export establishes that conversion ran; it does not establish that the task evidence survived.
Treat units and coordinate context as data
Illustrative example: a source object stores a dimension in millimetres, while a derived record labels the unchanged value as metres. The model can learn a perfectly consistent answer from a semantically incorrect target. A similar failure occurs when local coordinates are treated as project coordinates or an area is confused with a length.
Preserve the original value, source unit, conversion rule and resulting unit. For geometry tasks, identify the coordinate frame and any transformation applied. Include checks for missing or conflicting metadata. A reviewer should be able to reproduce the conversion without guessing what the preparation script assumed. Unknown units should remain unknown until resolved.
Match the requirement to the correct artifact version
A requirement from a revised brief may be paired with an earlier drawing because their filenames look related. That pairing teaches the wrong relationship even if each file was extracted accurately. Use project identifiers, explicit revision information and supporting source references to establish links. Mark inferred relationships for review rather than promoting filename similarity to evidence.
Document revisions also need careful interpretation. A changed paragraph shows that the document changed. It does not necessarily identify the responsible conversation, the corresponding artifact revision or approval of the result. Preserve the relationship you can establish and record the stages you cannot reconstruct. Missing history is preferable to an invented trajectory.
Define the record before preparing the corpus
Write an example contract around one target behavior. Specify what the model receives, what it must produce, which source evidence supports the target and how the record will be accepted. Requirement extraction, object matching and instruction following need different targets. A generic document dump cannot substitute for that decision.
Separate observed outputs from reviewed reference answers. Include evidence identifiers, preparation version and unresolved issues in the record metadata. Datasheets for Datasets recommends documenting motivation, composition, collection and recommended uses. For this preparation decision, add a transformation log so a buyer can trace a derived field back to its source and distinguish source limitations from processing errors.
Replay the same task through each preparation stage
Choose held-out tasks with inspectable source evidence. First give a reviewer the native source, then the extracted representation, then the assembled task record. Ask whether each representation still supports the same answer. This comparison can reveal a conversion problem before a model experiment obscures it.
Next run the same model with context prepared through competing conversion approaches. Keep the task, prompt and scoring rules stable. Compare source fidelity, accepted record yield, review effort and downstream task errors. Exclude related project revisions from training. Track rejected records as well as accepted ones, so a preparation method cannot look efficient by silently discarding the hard material.
Buy the next preparation step with a clear acceptance test
The pilot should end with a concrete diagnosis: unreadable source, faulty extraction, lost measurement context, unresolved linkage or an undefined target. Each points to different work. Additional files help only when they address a coverage need that the team can convert into usable records.
Agree on native-file delivery and prepared-record acceptance separately. Specify required metadata, review responsibilities, known exclusions and what happens to rejected examples. Scope permissions for processing and the intended model use alongside those deliverables. A source sample can establish preparation feasibility; it cannot establish model improvement by itself. Expand the corpus when the preparation path and your evaluation support that decision.