active-learning-relabel-loop

active-learning-relabel-loop

Overview

Use this skill to rank candidate structures for the next relabel round and maintain a clean handoff between current dataset, current model, and selected next-step structures.

Quick Start

1. Start from either a candidate structure pool or a curated dataset.
2. Use calculate_al_candidates to rank and select the next structures.
3. After new reference calculations are collected, rebuild the dataset with build_dataset_from_runs.
4. Retrain and benchmark by preparing mace_train_dir / mace_eval_dir stage layouts and submitting them with remote_submission.

Allowed tools

• calculate_al_candidates
• build_dataset_from_runs
• remote_submission
• get_avail_remote_task

Workflow

1. Select under one explicit ranking rule

• If a committee is available, say whether disagreement is part of the ranking.
• If not, state that the round is diversity-first.
• Treat the current selector as a baseline heuristic: diversity in a simple structure-feature space plus optional committee disagreement from per-atom energy variance.

2. Keep relabel bookkeeping clean

• Preserve the emitted ranking JSON/CSV and any selected-structure export as the loop ledger.
• Do not overwrite the previous round’s selection artifacts.

3. Rebuild, then retrain

• After the new reference calculations finish, rebuild the dataset rather than manually appending hidden frames.
• Retrain and re-benchmark from the new dataset state so round-to-round changes are auditable.
• Keep the MACE retrain leg aligned with the same validated finetune path used in mace-finetuning-and-benchmark, especially the foundation-model head and E0 strategy.

4. Use this skill for the model-update portion of the loop

• Candidate structures may originate from materials-side screening, but this skill starts when the loop object is selection, dataset refresh, retraining, and benchmark comparison.
• Keep the selection ledger, rebuilt dataset, and updated model artifacts explicit so the next relabel round is auditable.

Method-critical defaults

• Surface the candidate source, committee size, and selection size in every loop round.
• Do not describe the current scorer as a richer uncertainty engine than it is; it is not force-uncertainty AL, Bayesian optimization, or a calibrated acquisition function.
• Do not compare active-learning rounds if the held-out benchmark split changed without being reported.
• Keep the round boundary explicit: each cycle should expose the incoming candidate pool, the selected subset, the rebuilt dataset, and the updated benchmark outputs.

Output Contract

Return:

• AL ranking JSON/CSV path
• selected candidate artifact path
• rebuilt dataset summary path after relabel
• updated training/evaluation artifact paths

References

• This skill coordinates the loop contract; the expensive reference calculations still happen outside the ML lane.
• Use the same validated training recipe documented in vasp_to_mace_finetune.md when comparing AL rounds.