Example Unit

An example unit defines what one row, item, case, or training example represents. It is the first data question because labels, splits, metrics, and leakage all depend on the unit.

A supervised dataset can be written as:

$$\mathcal{D}=\{(u_i,x_i,y_i,m_i){\}}_{i=1}^n$$

$u_i$ is the example unit identifier, $x_i$ is input, $y_i$ is the target, and $m_i$ is metadata.

Common Units

Domain	Possible Example Unit	Typical Context
Vision	image, video clip, frame, object instance	camera, scene, timestamp, annotation source
Language	document, passage, prompt, message, conversation	source, user request, retrieval context
Molecules	standardized molecule, conformer, ligand, assay row	target, assay, endpoint, chemical state
Proteins	sequence, residue, domain, structure, protein family	species, construct, structure source, MSA context
Structure-based modeling	receptor-ligand pair, pocket-ligand pair, pose, complex	receptor state, pocket definition, ligand pose source
Genome sequence	sequence window, genomic region, variant	reference genome, strand, coordinate system, cell or tissue context when public
Systems	request, job, run, event, user action	environment, version, time block, policy

Why It Matters

Changing the example unit changes the task. A molecule-level prediction, molecule-target prediction, and assay-measurement prediction can use similar features but answer different questions.

For pair or context-dependent tasks, the unit may be a tuple:

$$u_i=(e_i,c_i)$$

where $e_i$ is the main entity and $c_i$ is context such as target, assay, prompt, pocket, source, or time. The same entity can appear in multiple examples if the context changes.

Ambiguous Cases

• Molecule property: unit may be molecule, assay row, molecule-target pair, or measurement.
• Protein task: unit may be sequence, domain, family, structure, residue, or complex.
• Docking task: unit may be receptor-ligand pair, pose, pocket-ligand pair, or screening candidate.
• LLM task: unit may be prompt, message, conversation, document chunk, or user request.

Unit Changes That Change the Claim

If the unit is	The model claims	Main leakage risk
molecule	a property attached to a chemical identity	duplicate salts, tautomers, stereo variants, scaffold overlap
molecule-target pair	target-conditioned activity or affinity	same molecule or same target appearing across splits
assay measurement	a protocol-dependent observed value	assay, source, threshold, or campaign leakage
protein sequence	sequence-level behavior	close homologs across train and test
residue	local annotation or residue-level output	same protein split into train and test residues
protein-ligand pose	pose scoring or pose classification	generated poses from the same complex crossing splits
genomic region	region-level prediction	adjacent windows or same locus crossing splits

The unit should be chosen before model design. Otherwise the same score can be read as a stronger generalization claim than the dataset supports.

Minimal Contract

For a reusable note, state:

• Unit identifier: what makes two examples the same or different.
• Input fields: raw object, context object, metadata, and required preprocessing.
• Label owner: whether the label belongs to the entity, pair, assay, pose, region, or run.
• Split implication: which broader group must stay together.
• Metric implication: whether metrics aggregate by row, entity, target, scaffold, family, or source.

Checks

• What is exactly one example?
• Can multiple rows describe the same underlying entity?
• Does the target label belong to the example, a pair, a group, or a context?
• What metadata is needed to interpret the label?
• Does the split unit need to be broader than the example unit?
• Would changing the unit change the loss, metric, or leakage risk?

Related

• Dataset construction checklist
• Data schema
• Split unit
• Dataset
• Entity relation map
• Leakage