Dataset Card

A dataset card is a structured summary of what a dataset contains, how it was built, what it can support, and where it can mislead.

A compact view is:

$$C_{\mathcal{D}}=(\text{source},\text{schema},\text{examples},\text{labels},\text{splits},\text{limits})$$

Suggested Sections

• Purpose: what the dataset is for.
• Source: public origin, license, and collection process.
• Lineage: how raw records were filtered, transformed, labeled, and split.
• Example unit: what one example represents.
• Schema: fields, identifiers, units, and relationships.
• Labels: target meaning, annotation process, uncertainty, missing values, censoring, and weak labels.
• Entity contract: molecule, target, assay, label, unit, threshold, censoring, and source when relevant.
• Preprocessing: filtering, normalization, tokenization, featurization, and version.
• Splits: split unit, policy, sizes, and leakage checks.
• Evaluation: supported tasks and metrics.
• Limitations: sampling bias, coverage gaps, label noise, and domain shift.

Claim Boundary

A dataset card should state what claims the dataset can and cannot support:

Claim type	Card should specify
interpolation	duplicate, near-neighbor, and source coverage
new chemistry	scaffold or cluster split policy
new protein/family	sequence identity or family split policy
future deployment	time split and source drift
probabilistic decision	calibration set and label uncertainty
generation quality	validity, novelty, diversity, and filter policy

The card is not just documentation; it constrains how benchmark results may be interpreted.

Versioned Identity

Dataset identity should include data and processing versions:

$$\mathrm{id}(\mathcal{D})=H(\text{raw source},\text{filters},\text{labels},\text{splits},\text{preprocessing})$$

If any of these change, results should not be treated as directly comparable without an update note.

Checks

• Can the dataset be reconstructed from public information?
• Are label semantics and units clear?
• Does the split support the intended benchmark claim?
• Are limitations explicit enough to narrow claims?
• Does the card avoid private sources, paths, or unpublished results?
• Are preprocessing and split decisions versioned?
• Is the intended evaluation protocol stated, not only the metric?
• Does the card distinguish raw data, curated data, features, and benchmark splits?
• Does the card state unsupported or out-of-scope claims?

Minimum Public Card

For a public wiki note, a minimal dataset card should include:

• Public source and citation status.
• Example unit and split unit.
• Label semantics and missing-label policy.
• Censoring and weak-label policy when applicable.
• Target-assay-label contract for chem-bio datasets when relevant.
• Preprocessing contract summary.
• Split policy and leakage checks.
• Supported tasks and metrics.
• Known limitations and out-of-scope claims.

Related

• Dataset construction checklist
• Data versioning
• Metadata and provenance
• Data lineage
• Example unit
• Split unit
• Dataset split contract
• Preprocessing contract
• Target-assay-label contract
• Evaluation protocol
• Benchmark
• Benchmark card