Glossary

Definitions of the terms used throughout the ALF documentation.

Active learning

A sequential strategy that, given a fixed labelling budget, repeatedly chooses the most informative or highest-value candidates to label next, rather than labelling everything or a random sample. ALF implements active learning as an ask/tell loop.

Acquisition function

The rule that scores candidates by their expected value to the search and decides which to pick next, trading off exploitation (high predicted value) against exploration (high uncertainty). Built-in examples: Greedy, UCB, Expected Improvement, Thompson Sampling, CoreSet. Distinct from a Search function, which produces the pool the acquisition function scores.

Best-found

The value of the best candidate discovered so far, tracked as a function of round. The headline “are we winning?” curve for a design experiment (see Task).

Calibration

How well a model’s stated uncertainty matches its actual error. A well-calibrated Surrogate “knows what it doesn’t know”, which is what lets an Acquisition function explore intelligently. Often summarised by ECE.

Candidate

A single point in the search space (e.g. a protein sequence or a SMILES string), optionally carrying features and a label. Represented by Candidate / LabelledCandidates in alf_core.

Coverage

The fraction of the truly high-value region that an experiment has discovered. Distinguishes finding the peak from finding all the good designs; related to Recall.

Dataset

The candidate pool together with its labels and train/validation splits, plus a query interface. Base class BaseDataset; examples include GFP, FLIP, ProteinGym, and GuacaMol.

ECE

Expected Calibration Error: a scalar summary of Calibration, the average gap between predicted confidence and observed accuracy across confidence bins.

Modality

The kind of a candidate — used to match datasets with compatible Models and to pick a metric, not how the data is stored. It is the data’s domain where one exists (SEQUENCE, MOLECULE), with TABULAR the domain-agnostic case for raw numeric feature vectors. Storage type (used for serialisation) is inferred separately from type(data).

Model

A learnable predictor with a common interface: featurise, train, predict, sample. Base class BaseModel; examples include CNN, Gaussian Process, ESM-2, and Chemprop. A model becomes a Surrogate or an Oracle depending on the role it plays in the loop.

Offline

A loop in which the Oracle reads labels from a held-out, pre-scored pool. Fully reproducible with no external calls; the mode used for benchmarking and method development.

Online

A loop in which the Oracle obtains labels from a live scorer (a trained model or simulator) on demand; the mode used to drive a real experiment. The loop is otherwise identical to Offline.

Optimizer

The object that bundles an Acquisition function and a Search function and exposes the ask (propose a batch) and tell (retrain on new labels) interface. Class Optimizer.

Oracle

The source of ground-truth labels for a proposed batch. Wraps a scorer that is either a trained Model (online) or a Dataset (offline). Class Oracle.

Precision

Of the candidates a method selected as high-value, the fraction that truly are. Reported alongside Recall for classification-style objectives.

Recall

Of all the truly high-value candidates, the fraction the experiment has selected. Measures whether the method is finding the good region, not just one point; see also Coverage.

Regret

The gap between the best achievable value and the best value found so far. Lower is better; regret-versus-round is the primary way ALF compares the speed of methods.

Search function

The component that generates the pool of Candidate objects an Acquisition function then scores, e.g. enumerating a dataset (DatasetSearch) or sampling from a generative Model (GeneratorSearch). Base class BaseSearch.

Spearman correlation

A rank-correlation metric used to judge how well a Surrogate’s predicted ordering of candidates matches the true ordering: what matters for selection, regardless of absolute scale.

State

The mutable carrier threaded through the loop: the current Dataset, Surrogate, round index, acquisition history, and per-round metrics. Class State.

Surrogate

A cheap-to-evaluate probabilistic approximation of the true objective, wrapping a Model and retrained each round on the data acquired so far. Its predictions (and uncertainty) drive the Acquisition function. Class Surrogate.

Task

The orchestrator that runs the loop end to end via setup then run. The family of task (DesignTask, SupervisedTask, or ZeroShotTask) determines what is being optimised. Base class BaseTask.