| # Data quality research plan |
|
|
| Quality is multidimensional and purpose-dependent. A single score must not replace the underlying |
| measurements or review decisions. |
|
|
| ## Required dimensions |
|
|
| | Dimension | Example evidence | Common failure | |
| |---|---|---| |
| | Integrity | content digest, byte count, decode errors | silent truncation or corruption | |
| | Duplication | exact and near-duplicate rates by source and split | evaluation leakage | |
| | Structure | parse success, field completeness, document boundaries | merged or fragmented records | |
| | Language | versioned detector plus manual sample | dialect and code-switch misclassification | |
| | Content quality | transparent heuristics and rejection counts | opaque quality score | |
| | Privacy | detector results plus human review | identifiers missed by automation | |
| | Safety | category definitions, sample review, subgroup counts | context-blind filtering | |
| | Coverage | domain, time, language, geography, and format distribution | hidden sampling bias | |
| | Contamination | train/eval overlap checks and hash strategy | benchmark memorisation | |
| | Reproducibility | fixed inputs, code revision, configuration, hashes | irreproducible `latest` snapshot | |
|
|
| ## Reporting rules |
|
|
| - Report absolute counts and rates. |
| - Break results down by source and split; aggregates can hide failures. |
| - Record every exclusion reason and overlapping filter count. |
| - Version detectors and thresholds, and retain calibration evidence. |
| - Use deterministic sampling for manual review and record the sampling method. |
| - Publish known blind spots and uncertainty. |
| - Never call a source clean, safe, representative, or unbiased without scoped evidence. |
|
|
