hermanhugging/qmof_project

AIXELO MOF Research Data Artifacts

Gated research repository. Access is granted manually for authorized AIXELO work and approved non-commercial study or research. The complete and canonical collaborative codebase is maintained in AIXELO’s private GitLab; this Hugging Face repository is a data-distribution and reproducibility endpoint for selected project artifacts.

At a Glance

Item	Status
Repository purpose	Distribution of MOF datasets, preprocessed LMDB archives, source archives, and preparation utilities
Primary users	AIXELO personnel, authorized collaborators, and approved non-commercial researchers
Access	Gated; manual approval
Commercial use	Not permitted for AIXELO-controlled materials without prior written permission
Canonical company code	Private AIXELO GitLab
Dataset viewer	Disabled because most artifacts are archives or LMDB databases
Contact	contact@aixelo.com

Important Licensing Boundary

This repository contains materials of mixed origin.

• AIXELO-controlled materials include project-specific preprocessing, conversion scripts, metadata, analysis, split definitions, documentation, and derived artifacts. They are available under the AIXELO Research-Only Data Artifact License.
• Third-party datasets remain subject to their original licenses, terms, attribution requirements, database rights, and access conditions.
• The AIXELO license does not replace an upstream license and does not restrict rights independently obtained from an original data provider.
• Access to this repository does not automatically authorize redistribution of upstream archives. Review the source terms before use or redistribution.

---

Quick Start for AIXELO Team Members

1. Install the Hugging Face client

python -m pip install --upgrade huggingface_hub

2. Authenticate

Use a personal Hugging Face token that has access to this gated repository:

hf auth login

Do not place access tokens in source code, notebooks, Slurm scripts, shell history shared with others, or committed configuration files.

3. Download only the artifact you need

Repository ID:

hermanhugging/qmof_project

Example — QMOF LMDB for CGCNN:

hf download hermanhugging/qmof_project \
  data/lmdb/CGCNN/qmof_cgcnn_lmdb.tar \
  --repo-type dataset \
  --local-dir ./qmof_project

The downloaded file will be placed at:

./qmof_project/data/lmdb/CGCNN/qmof_cgcnn_lmdb.tar

Extract it:

tar -xf ./qmof_project/data/lmdb/CGCNN/qmof_cgcnn_lmdb.tar \
  -C ./qmof_project/data/lmdb/CGCNN/

4. Pin a revision for reproducible runs

For production-like internal experiments, papers, reports, or handovers, use a commit hash rather than the moving main branch:

hf download hermanhugging/qmof_project \
  data/lmdb/CGCNN/qmof_cgcnn_lmdb.tar \
  --repo-type dataset \
  --revision <COMMIT_HASH> \
  --local-dir ./qmof_project

Record the repository revision, artifact path, target, split, preprocessing script revision, and model-code revision in the experiment log.

---

Which File Should I Download?

Ready-to-Train LMDB Archives

Use these when the target model pipeline already expects the corresponding precomputed LMDB representation.

Dataset / task	CGCNN	MGT	PMT
QMOF	qmof_cgcnn_lmdb.tar — 5.69 GB	qmof_mgt_lmdb.tar — 17 GB	qmof_pmt_lmdb.tar — 1.42 GB
ODAC23 CO₂ adsorption	odac_co2_cgcnn_lmdb.tar — 15.1 GB	odac_co2_mgt_lmdb.tar — 38 GB	Not provided
ODAC23 H₂O adsorption	odac_h2o_cgcnn_lmdb.tar — 9.39 GB	odac_h2o_mgt_lmdb.tar — 23.5 GB	Not provided
hMOF	hmof_cgcnn_lmdb.tar — 47.8 GB	Not provided	Not provided
CoRE MOF ASR	asr_cgcnn_lmdb.tar — 905 MB	Not provided	Not provided

Fast choices

• Train CGCNN on QMOF:
  data/lmdb/CGCNN/qmof_cgcnn_lmdb.tar
• Train MGT on QMOF:
  data/lmdb/MGT/qmof_mgt_lmdb.tar
• Train PMT / PMTransformer on QMOF:
  data/lmdb/PMT/qmof_pmt_lmdb.tar
• Train CGCNN on ODAC23 CO₂:
  data/lmdb/CGCNN/odac_co2_cgcnn_lmdb.tar
• Train MGT on ODAC23 CO₂:
  data/lmdb/MGT/odac_co2_mgt_lmdb.tar
• Train CGCNN on ODAC23 H₂O:
  data/lmdb/CGCNN/odac_h2o_cgcnn_lmdb.tar
• Train MGT on ODAC23 H₂O:
  data/lmdb/MGT/odac_h2o_mgt_lmdb.tar

CIF Structures with Target Tables

Use these for direct structure-based input, due diligence, target inspection, custom splits, or re-preprocessing.

Archive	Approx. size	Intended role
data/cif_plus_idprop/QMOF.tar	376 MB	QMOF structures and project-prepared target table
data/cif_plus_idprop/ODAC_init.tar	505 MB	ODAC23 initial / unrelaxed structures and target preparation
data/cif_plus_idprop/HMOF.tar	853 MB	hMOF structures and targets
data/cif_plus_idprop/CoreMOF_CR_Full.tar	162 MB	CoRE MOF CR full project package
data/cif_plus_idprop/CoreMof_SI_NOT_FULL_ASR.tar	26.8 MB	Partial CoRE MOF ASR package; not a full upstream release
data/cif_plus_idprop/MOSAEC_full_and_partial.tar	1.72 GB	MOSAEC full- and partial-structure project packages

Before training, inspect the extracted target table and confirm:

• target column;
• units;
• missing-value handling;
• structure identifier mapping;
• split definition;
• whether structures are relaxed, unrelaxed, full, partial, activated, or otherwise transformed.

Raw or Near-Upstream Source Archives

Use these only when you need to reproduce preprocessing, conduct provenance checks, or compare against the upstream release.

Source family	Repository path	Contents
QMOF	data/original_data/QMOF/	qmof_database.zip — 392 MB
ODAC23	data/original_data/ODAC23/	odac23_is2r.tar.gz — 848 MB; pristine_CO2.tar.gz — 93 MB; pristine_H2O.tar.gz — 72.3 MB
ODAC25	data/original_data/ODAC25/	validation-set project extraction plus conversion script
hMOF / MOFDB	data/original_data/hMof/	bulk-dl-mofdb-version-dc8a0295db.zip — 1.11 GB
CoRE MOF	data/original_data/CoreMof/	CSD-modified package, explicitly partial Zenodo package, and local provenance README
MOSAEC-DB	data/original_data/MOSAEC/	mosaec-db.csv — 57.3 MB and mosaec-db.zip — 903 MB

These archives may have terms that differ from the AIXELO project-specific license. Check the upstream source and the local provenance notes before redistribution.

---

Download Recipes

Download one file with the CLI

hf download hermanhugging/qmof_project \
  data/lmdb/MGT/qmof_mgt_lmdb.tar \
  --repo-type dataset \
  --local-dir ./qmof_project

Download several named files

hf download hermanhugging/qmof_project \
  data/lmdb/CGCNN/qmof_cgcnn_lmdb.tar \
  data/lmdb/CGCNN/odac_co2_cgcnn_lmdb.tar \
  --repo-type dataset \
  --local-dir ./qmof_project

Download a complete model folder

hf download hermanhugging/qmof_project \
  --repo-type dataset \
  --include "data/lmdb/CGCNN/*" \
  --local-dir ./qmof_project

Run a dry check before a large download:

hf download hermanhugging/qmof_project \
  --repo-type dataset \
  --include "data/lmdb/CGCNN/*" \
  --dry-run

Download with Python

One file

from huggingface_hub import hf_hub_download

path = hf_hub_download(
    repo_id="hermanhugging/qmof_project",
    repo_type="dataset",
    filename="data/lmdb/CGCNN/qmof_cgcnn_lmdb.tar",
    local_dir="./qmof_project",
)

print(path)

A selected folder

from huggingface_hub import snapshot_download

root = snapshot_download(
    repo_id="hermanhugging/qmof_project",
    repo_type="dataset",
    allow_patterns=["data/lmdb/MGT/*"],
    local_dir="./qmof_project",
)

print(root)

A pinned revision

from huggingface_hub import hf_hub_download

path = hf_hub_download(
    repo_id="hermanhugging/qmof_project",
    repo_type="dataset",
    filename="data/lmdb/MGT/qmof_mgt_lmdb.tar",
    revision="<COMMIT_HASH>",
    local_dir="./qmof_project",
)

HPC / Slurm Example

Authenticate once on the login node using a securely stored user token, then download to an approved project or work filesystem rather than a small home quota:

#!/usr/bin/env bash
set -euo pipefail

export HF_HOME="${WORK:-$HOME}/.cache/huggingface"
export HF_HUB_DOWNLOAD_TIMEOUT=60

DEST="${WORK:-$PWD}/qmof_project"

hf download hermanhugging/qmof_project \
  data/lmdb/CGCNN/qmof_cgcnn_lmdb.tar \
  --repo-type dataset \
  --local-dir "$DEST"

Do not embed a token directly in a submitted Slurm script.

---

Repository Structure

qmof_project/
├── README.md
├── LICENSE.md
└── data/
    ├── lmdb/
    │   ├── CGCNN/
    │   │   ├── preprocess/
    │   │   ├── qmof_cgcnn_lmdb.tar
    │   │   ├── odac_co2_cgcnn_lmdb.tar
    │   │   ├── odac_h2o_cgcnn_lmdb.tar
    │   │   ├── hmof_cgcnn_lmdb.tar
    │   │   └── asr_cgcnn_lmdb.tar
    │   ├── MGT/
    │   │   ├── preprocess/
    │   │   ├── qmof_mgt_lmdb.tar
    │   │   ├── odac_co2_mgt_lmdb.tar
    │   │   └── odac_h2o_mgt_lmdb.tar
    │   └── PMT/
    │       ├── preprocess/
    │       └── qmof_pmt_lmdb.tar
    ├── cif_plus_idprop/
    │   ├── QMOF.tar
    │   ├── ODAC_init.tar
    │   ├── HMOF.tar
    │   ├── CoreMOF_CR_Full.tar
    │   ├── CoreMof_SI_NOT_FULL_ASR.tar
    │   └── MOSAEC_full_and_partial.tar
    ├── original_data/
    │   ├── QMOF/
    │   ├── ODAC23/
    │   ├── ODAC25/
    │   ├── hMof/
    │   ├── CoreMof/
    │   ├── MOSAEC/
    │   └── analysis notebooks
    ├── util/
    │   ├── atom_init.json
    │   ├── prepare_asr.py
    │   ├── prepare_coremof_cr_datasets.py
    │   ├── prepare_dataset_co2.py
    │   ├── prepare_dataset_h2o.py
    │   ├── prepare_hmof.py
    │   ├── prepare_mosaec_full_structure.py
    │   └── prepare_mosaec_partial_structure.py
    └── cif_models/
        └── DEPRECATED.md

data/cif_models/ is retained only for backward compatibility. New workflows should use data/lmdb/, data/cif_plus_idprop/, and the corresponding preprocessing scripts.

---

Preprocessing Utilities

File	Role
data/util/atom_init.json	Elemental feature initialization used by CGCNN-style graph construction
data/util/prepare_dataset_co2.py	ODAC23 CO₂ dataset preparation helper
data/util/prepare_dataset_h2o.py	ODAC23 H₂O dataset preparation helper
data/util/prepare_hmof.py	hMOF preparation helper
data/util/prepare_asr.py	CoRE MOF ASR preparation helper
data/util/prepare_coremof_cr_datasets.py	CoRE MOF CR dataset preparation helper
data/util/prepare_mosaec_full_structure.py	MOSAEC full-structure preparation helper
data/util/prepare_mosaec_partial_structure.py	MOSAEC partial-structure preparation helper

Model-specific preprocessors are stored below:

data/lmdb/CGCNN/preprocess/
data/lmdb/MGT/preprocess/
data/lmdb/PMT/preprocess/

The preprocessing script and commit used for an artifact are part of its data contract. Do not assume that two archives with similar names contain identical atom selections, target definitions, graph fields, or splits.

---

LMDB Handling Notes

The LMDB packages are precomputed model inputs rather than a single universal schema. Always use the loader belonging to the matching project pipeline and revision.

Before launching a long HPC job:

1. extract the archive in a project or work filesystem;
2. verify that expected train, validation, and test files are present;
3. open a small number of entries with the matching loader;
4. check target units and sample count;
5. run a short CPU or single-GPU smoke test;
6. confirm output and checkpoint paths;
7. record the data and code revisions.

Some project artifacts serialize Python objects. Only deserialize files from a trusted, approved repository revision. Python pickle-compatible data can execute code during loading if it has been maliciously modified.

---

ODAC Workflow Note

ODAC archives represent several stages of the workflow: upstream or near-source structures, prepared CIF/target packages, and model-specific LMDB artifacts.

The exact inclusion of framework and adsorbate atoms, target construction, and sample filtering is controlled by the corresponding preprocessing script and its revision. Treat that script—not a historic README sentence—as the canonical definition of an ODAC artifact.

For a reproducible ODAC result, record:

• CO₂ or H₂O task;
• IS2R / IS2RE or other task formulation;
• initial or relaxed structure choice;
• atom-tag handling;
• target field and unit;
• filtering rules;
• split source;
• preprocessing commit;
• model loader commit.

---

Analysis Notebooks

The data/original_data/ directory also contains project due-diligence and analysis notebooks, including work on:

• QMOF;
• ODAC23 IS2R;
• ODAC25 validation-set extraction;
• CoRE MOF CSD-modified data;
• MOSAEC-DB.

These notebooks document exploratory checks and are not a substitute for the upstream dataset documentation or a finalized production pipeline.

---

Data Provenance and Citations

QMOF

Primary publication:

@article{rosen2021qmof,
  title   = {Machine learning the quantum-chemical properties of
             metal-organic frameworks for accelerated materials discovery},
  author  = {Rosen, Andrew S. and others},
  journal = {Matter},
  volume  = {4},
  number  = {5},
  pages   = {1578--1597},
  year    = {2021},
  doi     = {10.1016/j.matt.2021.02.015}
}

The official QMOF data distribution states that the underlying data are available under CC BY 4.0. Preserve attribution and indicate project-specific transformations.

ODAC23

@article{sriram2024odac23,
  title   = {The Open DAC 2023 Dataset and Challenges for Sorbent Discovery
             in Direct Air Capture},
  author  = {Sriram, Anuroop and others},
  journal = {ACS Central Science},
  volume  = {10},
  pages   = {923--941},
  year    = {2024},
  doi     = {10.1021/acscentsci.3c01629},
  eprint  = {2311.00341},
  archivePrefix = {arXiv}
}

hMOF / MOFDB

@article{boyd2019hmof,
  title   = {Data-driven design of metal-organic frameworks for wet flue gas
             CO2 capture},
  author  = {Boyd, Peter G. and Woo, Tom K.},
  journal = {Nature Chemistry},
  volume  = {11},
  pages   = {1026--1034},
  year    = {2019},
  doi     = {10.1038/s41557-019-0327-5}
}

CoRE MOF

@article{chung2019coremof,
  title   = {Advances, Updates, and Analytics for the Computation-Ready,
             Experimental Metal-Organic Framework Database: CoRE MOF 2019},
  author  = {Chung, Yongchul G. and others},
  journal = {Journal of Chemical and Engineering Data},
  volume  = {64},
  number  = {12},
  pages   = {5985--5998},
  year    = {2019},
  doi     = {10.1021/acs.jced.9b00835}
}

MOSAEC-DB

@article{gibaldi2025mosaec,
  title   = {MOSAEC-DB: a comprehensive database of experimental
             metal-organic frameworks with verified chemical accuracy
             suitable for molecular simulations},
  author  = {Gibaldi, M. and others},
  journal = {Chemical Science},
  volume  = {16},
  pages   = {4085--4100},
  year    = {2025},
  doi     = {10.1039/D4SC07438F}
}

ODAC25

@article{sriram2025odac25,
  title   = {The Open DAC 2025 Dataset for Sorbent Discovery in Direct Air
             Capture},
  author  = {Sriram, Anuroop and others},
  journal = {arXiv preprint arXiv:2508.03162},
  year    = {2025}
}

Dataset and publication licenses are separate. Always check the official dataset source and the exact version used.

---

Access and Permitted Use

Access is intended for:

• AIXELO personnel and collaborators authorized by the Company;
• students continuing related non-commercial study;
• approved academic or non-profit research;
• reproducibility and validation of scientific results.

AIXELO-controlled materials may not be used for commercial or third-party business purposes without prior written permission.

Requesting access does not guarantee approval. Access may be limited or withdrawn. Approved access is personal and must not be shared.

For commercial licensing, redistribution permission, or another use outside the research license, contact:

Aixelo, Inc.
contact@aixelo.com

---

Troubleshooting

401 or 403 when downloading

• confirm that you are logged in with hf auth login;
• confirm that the same Hugging Face account has been approved;
• confirm that your token has read access;
• retry without placing the token directly in the command.

Downloading the entire repository is too large

Download one named archive or use --include. Avoid cloning the full repository when only one model/dataset combination is needed.

Home-directory quota problems on HPC

Set HF_HOME and --local-dir to an approved work or project filesystem.

Dataset Viewer is unavailable

This is expected. Most files are TAR, ZIP, LMDB, notebooks, or source archives, not a viewer-compatible tabular dataset.

A path from an old script no longer exists

Check data/cif_models/DEPRECATED.md, inspect repository history, or pin the revision expected by the script. Current workflows should prefer data/lmdb/ and data/cif_plus_idprop/.

Unsure whether an archive is correct

Do not start a full training job. Inspect the archive, check its preprocessing script and revision, and run a small smoke test first.

---

Repository Status

This repository is an auxiliary research and data-distribution mirror. It is not the complete AIXELO project, not the canonical collaborative code repository, and not an official public product release.

The complete project history and internal components remain in AIXELO’s private GitLab environment.