shiqiao123/Muon-MACE-data

Muon-MACE: selection inputs, benchmark results and figure data

Data companion to Muon-MACE, a MACE research implementation with hybrid Muon–Adam optimization.

This repository contains 52 individually accessible data files (220,102,043 bytes). Browse a table, download an array or clone the directory tree; no ZIP extraction is needed. Code, configurations and plotting programs live on GitHub. 中文指南 · File catalog · Checksums and download mapping · Data dictionary

Find the data you need

Goal	Start here	Contents
Compare the five matched selections	Results table	Energy RMSE, force RMSE and GPU-hours, one row per arm
Rebuild the selections	Selection cache	Six arrays containing metadata, descriptors, fixed candidate pool and PCA
Use an accepted selection	Selected indices	Five unique 49,000-index arrays and their original manifest
Inspect optimizer results	Results	Per-seed, per-family and summary tables
Reproduce a paper figure	Figure data	Five figure-specific input directories and manuscript mapping
Verify the validation partition	Splits	Seed-123 positions within the 49k selected subset
Audit provenance	Provenance	Source-data checks, arithmetic acceptance and historical file hashes

selection/
  cache/                   metadata, descriptors, pool and PCA: 6 arrays
  indices/                 5 accepted index arrays + original manifest
results/                   benchmark rows, summaries and 3BPA predictions
figures/                   inputs grouped by the 5 plotting programs
splits/                    frozen validation-position list
provenance/                source checks and numerical acceptance
LICENSES/                  third-party license
manifest.json              file sizes, SHA-256, groups and local destinations
catalog.json               exact CSV columns/row counts and array shapes/dtypes
DATA_DICTIONARY.md          scientific meaning and units

Download directly

Git clone

With Git LFS installed:

git lfs install
git clone https://huggingface.co/datasets/shiqiao123/Muon-MACE-data
cd Muon-MACE-data
git lfs pull

CSV/JSON/text files are directly versioned by Git; NumPy arrays use Git LFS. A clone with large-file downloading disabled contains pointer files until git lfs pull completes. For an immutable copy, check out the dataset commit recorded in your Muon-MACE code release.

One table or one array

Every file has its own download link. For example:

import pandas as pd

url = (
    "https://huggingface.co/datasets/shiqiao123/Muon-MACE-data"
    "/resolve/main/results/off24_49k_same_pool_seed123.csv"
)
results = pd.read_csv(url)
print(results[["arm", "energy_rmse_mev_per_atom", "force_rmse_mev_per_a"]])

After a full clone:

from pathlib import Path
import numpy as np

root = Path("Muon-MACE-data")
selected = np.load(root / "selection/indices/combined_80_20.npy", allow_pickle=False)
assert selected.shape == (49000,)
assert np.unique(selected).size == 49000
pool = np.load(root / "selection/cache/baseline_candidate_indices.npy", allow_pickle=False)
assert np.isin(selected, pool).all()

Verified downloads for the reproduction scripts

Install Muon-MACE 0.2.1 or later from the code repository. These commands download individual files, verify each SHA-256 and arrange them in the layout expected by the existing reproduction scripts:

mace_muon_download --artifact paper --output-dir outputs/paper-artifacts
mace_muon_download --artifact selector --output-dir outputs/selector-cache
python tools/verify_paper_results.py --data-root outputs/paper-artifacts
python tools/verify_selector_realization.py --input outputs/selector-cache \
  --reference-dir outputs/paper-artifacts/data/selected_indices
python figures/fig4_off24_selection_49k/plot.py --data-root outputs/paper-artifacts

The code client pins a dataset Git commit and the manifest hash; the browser example above uses main for convenience. The manifest's local_path field maps this browsing layout to the historical plotting/verification layout.

Scientific scope

The selector uses an available 951,005-record molecular training corpus. The baseline pool contains 300,000 records; every matched arm contains 49,000 unique records at seed 123. Random-within-pool, representative-only, residual-only, 80:20 and 50:50 share the same label-aware pool and upstream selection plan. The two-column errors and runtime table reports these single-seed point estimates.

The 200k optimizer tables contain three training seeds and complete executed configurations. The historical 49k point in the size sweep belongs to a pre-PCA selection realization; figure notes identify it. The selector cache retains the same teacher-derived residual information, but the teacher checkpoint identity is not recoverable from surviving provenance.

3BPA files cover 7,047 dihedral conformers. Their relative energies are anchored at the DFT-minimum conformer. Historical fields named barrier_height_error refer to full-landscape energy-span error: absolute difference between the predicted and reference max–min energy spans. See the data dictionary for units and legacy names.

Source data and rights

The available train/test XYZ archives were verified against Cambridge's Research data supporting MACE-OFF23, whose repository rights metadata specifies MIT. Those raw archives remain at the official source; this repository contains the derived inputs and results listed in the manifest. The archive title is OFF23, although the study evaluates an OFF24 checkpoint.

3BPA references come from BOTNet-datasets, MIT licensed, Copyright (c) 2022 davkovacs. Its original license is preserved in LICENSES. The associated paper is 10.1021/acs.jctc.1c00647. Provenance and source-member hashes are in provenance.

No paper model weights are published here. Model releases belong in separate Hugging Face model repositories with checkpoint-specific cards and licenses.

Versioning

The directory release preserves every original data-file byte. The two former ZIP files have been removed from the current tree; old commit c10dd8e36c6cc6040d90ef0636c2c459a4ef1ff8 still resolves for older code clients. No history was rewritten. The legacy inventories are retained as provenance, while manifest.json and catalog.json describe the current directory release.