v0.1.0

The first stable release of ALF, under the Apache 2.0 license.

Published to PyPI as alf-core and alf-tools 0.1.0.

Added

• RandomAcquisition and UncertaintySampling acquisition functions, giving a random baseline and a pure-exploration strategy.
• Matbench dataset and Modality.MATERIALS, behind a new optional matbench/materials extra.
• SmilesMutationSearch and ElementSubstitutionSearch, search protocols that propose novel molecules and crystals each round.
• GuacaMolOracle, for scoring arbitrary SMILES in online loops.
• top_k on SingleMutantSearch, to seed mutants from more than just the single best sequence.
• BaseModel.embed() and Surrogate.embed(), giving embeddings a contract separate from featurise().

Changed

• CoreSet now calls embed() instead of featurise(). BaseModel.embed() raises NotImplementedError by default, so a custom model that relied on featurise() returning a 2-D embedding must now override embed() — return self.featurise(inputs) restores the previous behaviour.
• Tied labels in LabelledCandidates.get_top_k and sort now break by position, rather than in an unspecified order that varied with array size. Ties are routine, so benchmark numbers may shift and reference results may need regenerating.
• SingleMutantSearch now deduplicates its proposals, so the pool it returns can be smaller than before. It also rejects empty training sets and multi-output labels instead of silently mis-selecting.
• chemprop is capped below 2.3.0, and uv overrides raise the matminer, scipy, monty and scikit-learn floors workspace-wide.

Fixed

• one_hot_encode() pads variable-length batches instead of raising IndexError.
• BoTorchAcquisition featurises candidates, fixing GP models whose Candidate.data is not already numeric.

Full changelog: v0.1.0b0...v0.1.0

v0.1.0b0

ALF was released as a beta version under the Apache 2.0 license. First stable release will follow shortly. The following features are included:

• Introduced alf_core, a lightweight framework package providing base classes, core data structures (Candidate, LabelledCandidates, Predictions, State), and an active learning loop abstraction with no ML-framework dependencies.
• Introduced alf_tools, a ready-to-use package of surrogate models, datasets, and acquisition functions built on top of alf_core with PyTorch.
• Added surrogate models: CNNModel, MLPModel, Gaussian Process (GPModel), EnsembleWrapper for deep ensembles, and ChempropModel (message-passing neural network via Chemprop v2).
• Added ESM-2 protein language model surrogate with BERT-style masked language model fine-tuning, and ESMFold oracle for protein structure confidence scoring.
• Integrated BoTorch for continuous search spaces, including BoTorch-based surrogate models and acquisition functions (Expected Improvement, Upper Confidence Bound, Thompson Sampling) alongside discrete acquisition functions (Greedy, CoreSet with greedy k-centres).
• Added benchmark datasets: GFP (green fluorescent protein), FLIP, ProteinGym (with cross-validation support), and GuacaMol (small-molecule optimisation).
• Added classification support across the core framework, surrogate models, and metrics.
• Added input normalisation utilities including InputStandardiser and a configurable normalisation strategy.
• Added benchmarking metrics for active learning campaigns (recall, regret, and acquisition batch metrics) and benchmark example scripts comparing surrogates and acquisition functions.
• Added training metric logging via TerminalStateLogger and configurable per-round state recording.
• Published comprehensive documentation including installation guide, API reference, conceptual explanations, how-to recipes, and tutorials following the Diátaxis structure.