q734738781/mace-dataset-curation

mace-dataset-curation

Overview

Use this skill to convert collected VASP runs into a reusable extxyz dataset directory while staying close to the validated reference_scripts/mace_training_example export contract.

Quick Start

1. Point build_dataset_from_runs at one result root.
2. Choose frame_mode deliberately: final or all_ionic_steps.
3. If the downstream training uses a multi-head foundation model, set head_label explicitly, typically omat_pbe.
4. Keep the split fractions explicit and reproducible, and leave split_unit="source_run" unless you intentionally want frame-level leakage.
5. Leave require_converged=false unless you are intentionally building a guessed-converged subset.
6. Carry forward the emitted summary JSON and split file paths.

Allowed tools

• build_dataset_from_runs

Workflow

1. Choose the frame policy first

• The validated reference workflow starts from ionic-step data rather than only final frames, so all_ionic_steps is the default starting point when you want to match that path.
• final is only for deliberately reduced datasets where relaxed endpoints alone are the training target.

2. Keep split artifacts stable

• Treat dataset.extxyz, train.extxyz, valid.extxyz, and test.extxyz as the canonical handoff set.
• Use the summary JSON as the ledger for skipped runs, frame counts, stress encoding assumptions, and any head_label / config_type tags.

3. Keep the reference labeling contract explicit

• The validated reference export path writes REF_energy, REF_forces, REF_stress, config_type, and optionally head.
• If the downstream finetune uses mace-mh-1 with omat_pbe, set head_label="omat_pbe" instead of assuming the head will be inferred later.

4. Stop before training

• This skill ends at a curated dataset directory.
• Hand the resulting dataset to mace-finetuning-and-benchmark or active-learning-relabel-loop instead of mixing curation and training into one opaque step.
• Use this skill once the working artifact is a dataset directory rather than a structure-screening batch.

Method-critical defaults

• The reference-aligned default is to keep all ionic steps and store step_electronic_converged_guess for later filtering; use require_converged=true only when that hard subset is the actual dataset target.
• Keep alignment_check=true unless you are deliberately debugging malformed XML; XML/ASE step-order mismatches should cause the run to be skipped, not silently truncated into the dataset.
• Report whether the dataset contains final frames only or full ionic trajectories.
• Surface the head_label when the dataset is intended for multi-head foundation-model finetuning.
• Keep split_unit="source_run" for trajectory-style data unless you intentionally want frame-level mixing across train/valid/test.
• Do not silently reshuffle split fractions between model comparisons.
• Keep the handoff artifact explicit: this skill should start from collected result trees and end with a reproducible dataset directory plus split files.

Output Contract

Return:

• dataset directory
• split file paths
• dataset summary JSON
• any skipped-run ledger

References

• Reference flow: vasp_to_mace_finetune.md
• Export conventions: export_ase_db_to_mace_xyz.py