Dataset

A dataset is a curated collection of examples, labels, metadata, and splits used to train or evaluate models.

For chem-bio datasets, one row should usually preserve the Target-assay-label contract instead of collapsing the row into only molecule -> label.

A useful dataset note treats the dataset as an object with lineage:

$$\mathcal{D}=(\mathcal{S},\mathcal{U},\mathcal{X},\mathcal{Y},\mathcal{M},\Pi )$$

where $\mathcal{S}$ is source records, $\mathcal{U}$ is the example-unit set, $\mathcal{X}$ is model input, $\mathcal{Y}$ is labels or targets, $\mathcal{M}$ is metadata, and $\Pi$ is the split policy.

Dataset Object

Part	Question
Source	Where did the raw records come from, and can the source be cited publicly?
Schema	What fields, identifiers, units, and relationships exist?
Example unit	What is exactly one example?
Label semantics	What does the target mean, and under what measurement process?
Metadata	What context is required to interpret each example?
Split policy	What grouping prevents leakage for the intended claim?
Preprocessing	What transformations produce model-ready inputs?
Artifact boundary	Which files, hashes, versions, or scripts define this dataset version?

Example Unit vs Split Unit

The row identifier is rarely enough. A dataset should state both:

$$u_i=\text{example unit},g(u_i)=\text{split unit}$$

If two examples share the same split unit, they should stay on the same side of train, validation, and test for that claim.

Task	Example Unit	Split Unit
molecule property	standardized molecule or measurement row	scaffold, molecular cluster, source
target-conditioned activity	molecule-target-assay row	scaffold, protein family, assay/source
protein sequence task	sequence, domain, residue, or family record	sequence cluster or protein family
docking or pose task	receptor-ligand pair, pose, or complex	receptor, ligand scaffold, complex, pocket family
genome sequence task	window, region, variant, or locus	chromosome, locus, species, source batch

The split unit names the generalization claim. A random row split and a scaffold split are different scientific statements even if the model and metric are identical.

Label and Metadata Boundary

For supervised datasets, label meaning should not be inferred from a column name alone.

Label Issue	Required Context
binary activity	endpoint, threshold, unit, source, missing-label policy
affinity or potency	unit, transformation, censoring, assay protocol
pose correctness	reference structure, tolerance, failed-pose handling
generated-object validity	validity rule, invalid denominator, filtering policy
sequence annotation	reference database, version, evidence level, region definition

For chem-bio rows, the label often belongs to a tuple:

$$y_i=h(e_i,c_i)$$

where $e_i$ is the main entity and $c_i$ is context such as target, assay, pocket, source, time, construct, or measurement protocol.

Dataset Version

A dataset version is not just a file name. It should include enough information to reconstruct or audit the claim:

Version Field	Example
source snapshot	public dataset release, paper supplement, database date
filtering rule	allowed records, excluded records, failure handling
standardization	molecule normalization, sequence filtering, structure preparation
split rule	grouping key, validation policy, held-out claim
preprocessing artifacts	vocabulary, scaler, featurizer, conformer generator, graph builder
checksums or manifests	public-safe hashes or manifest shape
limitations	coverage gaps, bias, label noise, domain shift

Why It Matters

• Splits, duplicate handling, and label provenance shape what a benchmark measures.
• Molecular datasets often mix molecules, proteins, assays, and structures.
• Clear dataset notes make leakage risks and task definitions easier to audit.
• Dataset lineage explains how the dataset version was produced from source records.
• Dataset cards keep model claims from drifting beyond the data’s supported scope.

Public Dataset Note Shape

Use this shape when writing about a public dataset:

Section	Include
Purpose	task family and intended use
Public source	citation, license, release, and to verify for missing metadata
Example unit	row/entity/pair/pose/region definition
Schema	identifiers, fields, units, optional fields
Labels	meaning, unit, transform, censoring, missingness, weak labels
Splits	split unit, validation rule, leakage checks
Preprocessing	transformations fit on train only, versioned featurization
Evaluation	supported metrics and unsupported claims
Limitations	sampling bias, label noise, assay/source drift, applicability domain

Do not publish private raw paths, collaborator-only datasets, internal labels, unpublished metrics, or unreleased split files. If a dataset is private, write the general dataset pattern rather than the private artifact.

Checks

• What is one training example, and what is the target label?
• Is the example unit different from the row identifier?
• What split unit matches the intended generalization claim?
• Are duplicate molecules, protein families, or near-identical structures separated?
• Does the split match the intended generalization setting?
• Are missing labels, censored measurements, and assay metadata handled explicitly?
• Is the split key the same entity that defines the intended generalization claim?
• Is label provenance kept with each row?
• Does each row keep target, assay, unit, threshold, censoring, and source context when relevant?
• Are preprocessing artifacts fit only on training data?
• Can the public note support the claim without exposing private sources or unpublished results?

Related

• Entity relation map
• Target-assay-label contract
• Data
• Dataset card
• Example unit
• Split unit
• Dataset split contract
• Dataset construction checklist
• Preprocessing contract
• Label semantics
• Data curation
• Data lineage
• Annotation and labeling
• Metadata and provenance
• Bioactivity label
• Benchmark
• Data preprocessing
• test split
• Assay harmonization
• Activity cliff
• Applicability domain
• Assay
• Molecule
• Sequence
• Structure