An open archive of LibRPA materials calculations — band structures, large-system GW, RPA energies, and reader conversions from both ABACUS and FHI-aims front ends — each with its original inputs, exact basis bindings, software identity, and validation scope. Nothing to install: the link above searches, inspects, previews, and downloads every archived file. The companion ABACUS + LibRPA GW basis skill covers the ABACUS/NAO workflow itself.
| Layer | Location |
|---|---|
| Catalog, UI, validators, basis library | this repository |
| Calculation payload bytes (inputs, results, logs, figures) | external archive |
- Primary archive: https://huggingface.co/datasets/AroundPeking/LibRPA-database
- Mirror (reachable from mainland China): https://hf-mirror.com/datasets/AroundPeking/LibRPA-database
Every payload is declared in its record.json with the repository-relative
path as identity, size_bytes, sha256 as the integrity anchor, and direct
plus mirror URLs (tier: "referenced"). Validation gates fetch the bytes and
verify the hash (.external-cache/, never committed), so acceptance strength
is the same as when the bytes sat in git. A merged data PR uploads its
payloads to the archive automatically; contributors never need a HuggingFace
account. Basis payloads (data/elements) and species bindings stay in git.
data/calculations/{2d,3d}/<Material>/<case-id>/ record.json + inputs/ + results/ + logs/ + figures/
data/elements/<Element>/ basis library payloads
database/ catalog UI; data.json is the machine-readable catalog
data/log-index.json exactly one owner per retained log
curl -s https://aroundpeking.github.io/LibRPA-database/database/data.json -o data.jsonEach declared file carries its hash and download URLs — verify the SHA-256
after downloading. For AI agents, the
plugin/ directory
packages a zero-dependency read-only MCP server and two skills: contributing a
calculation as a pull request, and reading the data without MCP.
Data checks use the Python standard library; frontend tests need Node.js 22+.
python3 scripts/library.py --write-catalog
python3 scripts/build_indexes.py --write
python3 scripts/public_scan.py .
python3 -m unittest discover -s tests -v
node --test tests/*.mjsCI accepts only the historical software-identity pairs frozen in
data/historical-software-identity-exceptions.json; a new record needs a full
source commit or the executable's SHA-256. Band figures are rebuilt with
scripts/generate_band_figures.py (display and evidence contract in
references/plotting.md). Contribution rules:
CONTRIBUTING.md.
Pushing to main publishes through the Publish database workflow; the
deployed commit is recorded in build-info.json. To preview a build locally:
python3 scripts/build_site.py --output _site
python3 -m http.server 8766 --bind 127.0.0.1 --directory _siteThe scripts, schema notes, and catalog UI are MIT licensed. Third-party software and scientific data retain their own licenses and reuse conditions; check the source record before redistributing a payload.