Data Lineage

Data lineage records how a dataset was derived from source records through curation, filtering, preprocessing, labeling, splitting, and evaluation. It connects Metadata and provenance, Data versioning, Preprocessing contract, and Evaluation protocol.

A compact lineage graph is:

$$\mathcal{S}\stackrel{\text{curation}}{\to }{\mathcal{D}}_{\mathrm{raw}}\stackrel{\text{filter}}{\to }{\mathcal{D}}_{\mathrm{clean}}\stackrel{P_{\varphi }}{\to }{\mathcal{D}}_{\mathrm{model}}\stackrel{s}{\to }({\mathcal{D}}_{\mathrm{train}},{\mathcal{D}}_{\mathrm{val}},{\mathcal{D}}_{\mathrm{test}})$$

where $\mathcal{S}$ is the source record set, $P_{\varphi }$ is the preprocessing contract, and $s$ is the split rule.

For a run, lineage should identify:

$$L_{\mathcal{D}}=(\text{source},\text{curation},\text{filters},\text{labels},\text{preprocessing},\text{splits},\text{version})$$

Why It Matters

• A small filtering change can change benchmark difficulty.
• Split leakage often comes from lineage mistakes, not model code.
• Reproducibility requires tracing metrics back to a dataset version and split.
• Public claims should only depend on public or safely summarized sources.
• Dataset lineage separates data changes from architecture or optimizer changes.

Lineage Record

A useful lineage record should include:

• Public source name, citation, license, and retrieval date or version.
• Raw record schema and stable identifiers.
• Inclusion and exclusion filters.
• Deduplication and near-duplicate policy.
• Label generation, harmonization, censoring, and missing-label policy.
• Preprocessing, tokenization, standardization, featurization, and normalization contracts.
• Split rule, split unit, random seed, group IDs, and leakage checks.
• Dataset version identifier and any content hashes that can be shared publicly.
• Link to the training run, evaluation protocol, and dataset card.

Molecular Modeling Lineage

For structure-based and molecular datasets, lineage should preserve entity relationships:

$$(\text{molecule},\text{target},\text{assay},\text{label},\text{unit},\text{source})$$

or, for structure data:

$$(\text{protein},\text{ligand},\text{complex},\text{pose},\text{pocket},\text{source})$$

Flattening these records too early can hide assay changes, target ambiguity, template leakage, or pocket-definition leakage.

Checks

• Can every processed example be traced to a public or safely summarized source?
• Are filters and deduplication rules recorded before metric interpretation?
• Is the split generated from the intended split unit rather than row order?
• Are labels tied to source, unit, threshold, and censoring context?
• Can a run be associated with one dataset lineage identifier?
• Are private paths, private dataset names, collaborator details, and unpublished results excluded from public notes?

Related

• Data
• Metadata and provenance
• Data versioning
• Dataset card
• Dataset construction checklist
• Dataset split contract
• Preprocessing contract
• Leakage
• Evaluation protocol
• Run artifact
• Dataset