Data Changelog
cleaned-20261107
The main branch provides the current cleaned data. All 779 questions, 292 dataset folders, and 214 test_current.csv oracles are retained. Question IDs and the inference/evaluation JSON structure are unchanged.
- Corrected reviewed qualitative/categorical descriptions, codebook interpretations, feature aliases, outcome proxies, and ranking direction metadata.
- Applied strict visible-input group holdout to seven hotel batches: removed 387 overlapping history rows, without changing current candidates or outcome labels.
- Restored 541 hidden health target cells and 2,877,541 CSV target cells from exact original 2020 records using the documented
GenHealth/GeneralHealthcrosswalk. CSV counts include repeated views, not unique people. Removed the redundant input alias. - Replaced three genuinely unlabelled health candidates with observed 2022 records, preserving IDs and filter eligibility; removed their six matching history records. Rebuilt complete classification oracles from observed labels.
- Made the placed-only salary cohort explicit in seven campus recruitment questions and filters. Non-placed salaries remain missing and outside the scoring cohort, not imputed as zero.
- Removed one observed health-grade statement from a query that asks for that grade; clarified one wine target class and one mall spending objective. These edits do not change outcome labels.
- Corrected all eight forest-fire questions to monthly rather than daily counts and explained the original January-1 year marker. Restored corrupted month/state characters against the original Latin-1 Kaggle archive, with full-record matching and numeric values preserved.
- Restored the corrupted chocolate maker name against the original rating publisher's matching records; rating values are unchanged.
- Clarified seven gun-incident comparisons so they do not claim that distinct source records have identical report contents or identify a causal effect. Their provided numerical conditions, observed outcomes, and gold trends are unchanged. Auxiliary source-record differences remain a documented limitation.
Compatibility And Limits
The original used-car archive contains 104 released history descriptions with one unrecoverable replacement character each. Exact descriptions were verified against the original archive and retained verbatim, not guessed or silently dropped. These source-text artifacts are separate from the repaired structured-category encoding errors.
Historical model outputs and summary scores are not results for this cleaned version. Re-run inference and evaluation, preserving previous outputs separately. test_current.csv and JSON gold must remain evaluator-only.
Deterministic validation and full-dataset automatic semantic screening, followed by a second pass over flagged cases, were supplemented by source-based adjudication. This is not exhaustive human annotation or a guarantee that every qualitative phrase is error-free. Qualitative language has no published one-to-one quantile mapping. Treatment Effect Analysis is predictive scenario comparison rather than causal identification.
Character-Restoration Sources
- Forest Fires in Brazil: original Kaggle data (Latin-1 archive; not the already-lossy local UTF-8 copy).
- Chocolate Bar Ratings: Kaggle data and the original Flavors of Cacao database. The old Kaggle CSV itself contains the corrupted maker spelling; the publisher's matching reference/year/origin/cocoa/rating records support the spelling repair.