aschachner/vacua_vault

CY Vacua Vault

A community-curated dataset of flux vacuum solutions for Type IIB string theory compactifications on Calabi–Yau threefolds. Solutions are stored per-model and cross-referenced against the geometry catalogue aschachner/cy-database.

Scope. This repository stores flux-sector vacua — solutions of the complex-structure + axio-dilaton flux equations (flux vectors, stabilised moduli, tadpole values, classification flags, metadata). It does not store the underlying geometries; those live in the geometry repo above.

KKLT / de Sitter constructions are out of scope here. Vacua that additionally stabilise the Kähler moduli — KKLT-type non-perturbative superpotentials, de Sitter uplifts — are maintained in a separate dedicated repository. This vault is limited to the flux (complex-structure + dilaton) sector; the overall Calabi–Yau volume is left unfixed (see the mass-units note in §3.2).

Maintained for use with the stringforge package (database/vault I/O) together with jaxvacua (flux-vacuum physics), but the parquet schema is library-agnostic — any tool that can read Apache Arrow can consume the data.

---

1. Dataset summary

Each row in the vault is a single flux vacuum, i.e. a choice of integer fluxes $f,h\in H^3(Y,\mathbb{Z})f,\; h\; \backslash in\; H^3(Y,\; \backslash mathbb\{Z\})$ together with the stabilised complex structure moduli $z\in {\mathbb{C}}^{h^{1,2}}z\; \backslash in\; \backslash mathbb\{C\}^\{h^\{1,2\}\}$ and axio-dilaton $\tau \in \mathbb{C}\backslash tau\; \backslash in\; \backslash mathbb\{C\}$ satisfying

$$D_{A}W(z,\tau ;f,h)=0\text{(SUSY)}\text{or}{\mathrm{\partial }}_{A}V(z,\tau ;f,h)=0\text{(non-SUSY)},D\_A\; W(z,\; \backslash tau;\; f,\; h)\; \backslash ;=\backslash ;\; 0\; \backslash quad\; \backslash text\{(SUSY)\}\; \backslash qquad\; \backslash text\{or\}\; \backslash qquad\; \backslash partial\_A\; V(z,\; \backslash tau;\; f,\; h)\; \backslash ;=\backslash ;\; 0\; \backslash quad\; \backslash text\{(non-SUSY)\},$$

with the tadpole constraint $\mathrm{\mid }{f}^T\mathrm{\Sigma }h\mathrm{\mid }\le N_{\max }|f^T\; \backslash Sigma\; h|\; \backslash le\; N\_\{\backslash max\}$. Solutions are grouped by the Calabi–Yau model they belong to. Curated (reviewed) datasets live directly in the model directory; pending contributions are staged in a community/ sub-folder of the same model directory. Once a maintainer curates a community submission it is promoted out of community/ into the model directory (see §6 for the lifecycle).

---

2. Repository layout

Curated datasets sit directly in each model directory. community/ is a submission queue for pending (unreviewed) contributions; once reviewed, files move out of community/ up to the model level.

Homogeneity invariant. Each parquet file in the vault contains rows describing exactly one model — every row of a given file shares the same (ks_id, triang_id) (TDF), cicy_id (CICY), or model_ID (local jaxvacua models). The directory layout (<namespace>/h12_N/<identifier>/<label>.parquet) makes one-file-per-model the natural unit, and the catalogue, server-side validation, and the validate_parquet_file auto-load path all rely on this assumption. Mixed-model files are rejected by validate_parquet_file with a clear top-level error.

vacua_vault/
├── catalog.parquet                    # Auto-generated index of all datasets
├── schema.json                        # Schema version + column spec
├── retractions.parquet                # Audit trail of retracted entries
├── README.md                          # This file
├── tdf/                               # Trilayer, double-favourable Kreuzer–Skarke models
│   ├── h12_2/
│   │   ├── ks_00029_tri_0000/
│   │   │   ├── SUSY_Nmax34.parquet                # curated, lives at model level
│   │   │   ├── nonSUSY_ISD-_Nmax200.parquet       # curated
│   │   │   ├── community/                          # pending submissions queue
│   │   │   │   ├── alice-hf_racetrack_W0_small.parquet
│   │   │   │   └── bob-hf_nonSUSY_critpts.parquet
│   │   │   └── _rejected/                          # (optional, split-mode only)
│   │   │       └── alice-hf_racetrack_W0_small.rejected.parquet
│   │   └── ...
│   └── h12_3/
│       └── ...
├── cicy/                              # Complete Intersection CY models
│   └── cicy_7807/
│       ├── SUSY_Nmax20.parquet                    # curated
│       └── community/
│           └── ...
└── local/                            # Local jaxvacua models (addressed by h12 + model_ID,
    └── h12_2/                         #   not a cy-database id — see §2.1)
        └── model_1/
            ├── ISDplus_demo.parquet               # curated
            └── community/
                └── alice-hf_extra_ISDplus.parquet

2.1 Path semantics

Segment	Meaning
tdf / cicy	Model rooted in the geometry repo (aschachner/cy-database).
local	Local jaxvacua model, addressed by its bundled-model index (h12, model_ID) rather than a cy-database id.
h12_{h12}	Hodge number $h^{1,2}h^\{1,2\}$ — fixes the dimension of the complex structure sector.
ks_{ks_id}_tri_{triang_id}	KS polytope + triangulation index (TDF models).
cicy_{cicy_id}	CICY configuration matrix index.
model_{model_ID}	jaxvacua-internal model index (the local/ namespace).
Curated files at model level	Reviewed, canonical datasets. Downloaded by default.
community/	Pending, unreviewed contributions. Filename prefixed with the contributor's HF username.
_rejected/	(Opt-in, split-mode only.) Split-off rows that failed row-level validation. Excluded from default queries.

2.2 Filename conventions

Filenames are free-form human-friendly labels. Classification (SUSY status, sampling method, $N_{\mathrm{m}\mathrm{a}\mathrm{x}}N\_\{\backslash mathrm\{max\}\}$, post-selection qualifiers, …) lives in the catalogue and in parquet-level key-value metadata, not in the filename.

• Curated files: {label}[_v{n}].parquet where {label} matches the slug regex ^[A-Za-z0-9][A-Za-z0-9_-]*$. e.g. SUSY_Nmax34.parquet, racetrack_small_W0.parquet, nonSUSY_ISD-_Nmax200.parquet.
• Community files (inside community/): {hf_username}_{label}[_v{n}].parquet e.g. community/alice-hf_racetrack_candidates.parquet.
• Rejected files (inside _rejected/, split-mode): {original_filename}.rejected.parquet.
• Versioned files: _v{n} is auto-appended by push_vacua_to_hub(if_exists="new_version") on collisions — contributors don't write it manually.

Reserved names (CI rejects them): catalog.parquet, retractions.parquet, schema.json, README.md, and any filename starting with _.

The {hf_username} prefix on community files avoids collisions between independent contributions. When a community submission is curated, the file is moved up to the model directory and the {hf_username}_ prefix is stripped; the original contributor is recorded in the catalogue as original_contributor so attribution survives the rename.

Collision policy (push_vacua_to_hub(if_exists=...)):

if_exists=	Behaviour
"error" (default)	Refuse upload; pick a different label or use "new_version".
"append"	Load existing + concatenate + dedupe + rewrite (single-writer only).
"new_version"	Auto-suffix with _v2, _v3, ... — always safe.

---

3. Schema

Every parquet file has the same column layout (see schema.json for the machine-readable spec).

3.1 Required columns

Column	Type	Description
flux	list<int> (length 2·(h12+1))	Integer flux vector $(f,h)(f,\; h)$ stacked.
moduli_re	list<float> (length h12)	Real part of stabilised $zz$.
moduli_im	list<float> (length h12)	Imaginary part of stabilised $zz$.
tau_re	float	Real part of $\tau \backslash tau$ (axion).
tau_im	float	Imaginary part of $\tau \backslash tau$ (dilaton, $>0>0$).
label	string	Human-readable dataset label.
committed_by	string	Contributor identifier (ORCID preferred, HF username accepted).
commit_date	date	Date of commit (ISO 8601).

3.2 Optional derived columns (recommended)

Column	Type	Description
W_re, W_im	float	Superpotential $W(z,\tau )W(z,\; \backslash tau)$.
F_terms_re, F_terms_im	list<float>	$D_{A}WD\_A\; W$ components.
N_flux	float	Tadpole $\mathrm{\mid }{f}^T\mathrm{\Sigma }h\mathrm{\mid }|f^T\; \backslash Sigma\; h|$.
residual	float	Newton residual of the solution.
is_susy	bool	Whether .
is_minimum	bool	Whether the Hessian is positive-definite.
V	float	Scalar potential at the solution.
g_s	float	String coupling $g_s=1\mathrm{/}\mathrm{I}\mathrm{m}(\tau )g\_s\; =\; 1/\backslash mathrm\{Im\}(\backslash tau)$.
mass2	list<float>	Sorted real scalar mass-squared spectrum — eigenvalues of the canonically-normalised no-scale mass matrix (sign-carrying; negative entries are tachyonic). FluxEFT normalisation — see note.
m_gravitino	float	Gravitino mass $m_{3\mathrm{/}2}=e^{K\mathrm{/}2}\mid W\mid m\_\{3/2\}\; =\; e^\{K/2\}\backslash lvert\; W\backslash rvert$. Same normalisation.

Mass units. mass2 and m_gravitino are in the FluxEFT no-scale normalisation, not $M_{\mathrm{P}\mathrm{l}}M\_\{\backslash mathrm\{Pl\}\}$. At the flux-EFT level the overall Calabi–Yau (Einstein-frame) volume ${\mathcal{V}}_E\backslash mathcal\{V\}\_E$ — a function of the Kähler moduli — is not stabilised, so the FluxEFT Kähler potential omits the $-2\log ({\mathcal{V}}_E)-2\backslash log(\backslash mathcal\{V\}\_E)$ term and both quantities carry an unfixed overall volume factor: $m_{\mathrm{p}\mathrm{h}\mathrm{y}\mathrm{s}}^2={\mathcal{V}}_E^{-2}m^2\_\{\backslash mathrm\{phys\}\}\; =\; \backslash mathcal\{V\}\_E^\{-2\}\backslash ,$mass2 and $m_{3\mathrm{/}2,\mathrm{p}\mathrm{h}\mathrm{y}\mathrm{s}}={\mathcal{V}}_E^{-1}m\_\{3/2,\backslash mathrm\{phys\}\}\; =\; \backslash mathcal\{V\}\_E^\{-1\}\backslash ,$m_gravitino. Only the dimensionless ratio $m\mathrm{/}{m}_{3\mathrm{/}2}m/m\_\{3/2\}$ (in which ${\mathcal{V}}_E\backslash mathcal\{V\}\_E$ cancels) is volume-independent and physical. These columns can be (re)computed from flux + stabilised moduli with db.complete_missing(df, model=...).

3.3 Metadata / provenance columns

Column	Type	Description
tags	string (JSON array)	Free-form tags for filtering (e.g. ["SUSY", "ISD-", "Nmax200"]).
extra_data	string (JSON object)	Per-solution metadata (e.g. {"moduli_max": 100, "solver": "hybrid"}).
designated_id	int64	Monotonic ID assigned on upload.
model_ID	int64 (nullable)	jaxvacua-internal model index for local/ models (catalogue-derived from the path); NULL for tdf / cicy.
status	string	Lifecycle state: "pending" (in community/), "curated" (at model level), "rejected" (split-mode failure), "retracted" (post-merge retraction).
original_contributor	string	HF username of the original contributor, preserved when a community submission is curated and the filename prefix is stripped.
retracted	bool	True if the entry has been retracted.
retracted_reason	string	Reason (required if retracted=True).
retracted_at	timestamp	When the retraction occurred.

Classification columns (auto-populated from parquet-level key-value metadata by rebuild_catalog; see §2.2):

Column	Type	Description
susy	string	"SUSY", "nonSUSY", "mixed", or "unknown".
method	string	"enumerate", "sample", "ISD+", "ISD-", "F", "H", "newton", "hybrid", or a custom method string.
nmax	int (nullable)	Tadpole bound. NULL means unbounded or unknown.
qualifiers	string (JSON object)	Free-form dict of extra classifiers (e.g. {"moduli_max": 100, "post_select": "racetrack"}).
classification_source	string	"parquet_metadata" (contributor set it), "filename_guess" (best-effort fallback), or "manual" (maintainer override).
version	int	Version number from _v{n} suffix; 1 if absent.
superseded_by	string	Relative path of a compacted file that replaced this one; empty string if current.

Schema evolution is tracked in schema.json. Breaking changes bump a major version; additive changes bump a minor version.

---

4. Relationship to aschachner/cy-database

cy-database (the geometry repo) and vacua_vault (this repo) are complementary:

	aschachner/cy-database	aschachner/vacua_vault (this repo)
Contents	CY geometries: intersection numbers, $c_{2}c\_2$, Kähler cone data, GV/GW invariants, conifold data	Flux vacuum solutions: fluxes, stabilised moduli, tadpole, classification
Update cadence	Static / slow (append-only as new triangulations are added)	Fast (new solutions continuously produced)
Size	Could reach millions of models	Grows with sampling effort per model
Key identifiers	(ks_id, triang_id) for TDF, cicy_id for CICY	Same keys (plus (h12, model_ID) for local jaxvacua models) + label + committed_by

A row in vacua_vault is only meaningful relative to a row in cy-database: the flux vector lives in the integer lattice $H^3(Y,\mathbb{Z})H^3(Y,\; \backslash mathbb\{Z\})$ defined by the model's period data. The recommended workflow is:

1. Load the model via stringforge.LCSDatabase(dataset="tdf").load_model(ks_id=29, triang_id=0).
2. Load solutions via db.fetch_vacua_from_hub(ks_id=29, triang_id=0, ...).
3. Reconstruct the moduli physics using the loaded model + stored fluxes.

---

5. Using the dataset

5.1 With stringforge (recommended)

from stringforge import LCSDatabase

db = LCSDatabase(dataset="tdf")

# Browse what's available
catalog = db.list_hub_vacua(h12=2)
print(catalog)

# Fetch curated solutions for a specific model (default)
df = db.fetch_vacua_from_hub(
    ks_id=29, triang_id=0,
    label="SUSY_Nmax34",       # optional filter
    include_community=False,    # default: only curated data
    cache=True,                 # cache locally
)

# Include pending community submissions (use with care)
df_all = db.fetch_vacua_from_hub(
    ks_id=29, triang_id=0,
    include_community=True,
)
print(df[["flux", "moduli_im", "tau_im", "N_flux", "is_susy"]].head())

5.2 With datasets or raw parquet

from datasets import load_dataset

# Curated split: reviewed datasets only
ds_curated = load_dataset("aschachner/vacua_vault", split="curated")

# Community split: pending submissions
ds_pending = load_dataset("aschachner/vacua_vault", split="community")

import pandas as pd
# Curated file — sits at model level
df = pd.read_parquet(
    "hf://datasets/aschachner/vacua_vault/tdf/h12_2/"
    "ks_00029_tri_0000/SUSY_Nmax34.parquet"
)

# Community file — inside community/ with HF-username prefix
df_pending = pd.read_parquet(
    "hf://datasets/aschachner/vacua_vault/tdf/h12_2/"
    "ks_00029_tri_0000/community/alice-hf_racetrack_W0_small.parquet"
)

5.3 Validating downloaded data

from stringforge import LCSDatabase
db = LCSDatabase(dataset="tdf")
model = db.load_model(ks_id=29, triang_id=0)
report = db.validate_vacua(df, model=model, F_term_tol=1e-6)

---

6. How to contribute

We welcome contributions from the community! Contributions go through a pull-request workflow on HuggingFace with automated validation on the PR branch.

6.1 Prerequisites

• A HuggingFace account (free).
• A contributor identifier — ORCID preferred; HF username accepted.
• Your solutions must reference a geometry that is already in aschachner/cy-database. If your model is not yet catalogued, contribute it there first (see the geometry repo's contributor guide).

6.2 Workflow

1. Prepare your dataset locally. Use jaxvacua to generate and designate solutions into your local vacua_vault/:

   # After running flux searches...
   db.designate_vacua(
       results_df,
       label="my_new_dataset",
       committed_by="Your Name (ORCID:0000-...)",
       model=model,
       tags=["SUSY", "Nmax200"],
   )
   

2. Push as a pull request:

   db.push_vacua_to_hub(
       results_df,
       label="my_new_dataset",
       committed_by="Your Name (ORCID:0000-...)",
       model=model,
       tags=["SUSY", "Nmax200"],
       create_pr=True,                   # opens a PR instead of pushing to main
       partial_failure="strict",          # or "split" (see §7)
   )
   # Prints a PR URL — share with reviewers.
   

   This creates a branch on your HF fork, uploads tdf/h12_{N}/ks_{X}_tri_{Y}/community/{your_hf_username}_my_new_dataset.parquet, and opens a PR against main.

3. Wait for CI validation. The PR-branch CI (see §7 for full details) validates your submission and posts two comments:

   • A validation report (pass/fail per file, error counts, row indices).
   • A catalogue diff ("this PR adds 1,247 rows across 1 dataset").

4. Respond to feedback. If CI fails, fix the issues locally and push again — CI re-runs automatically. If a reviewer requests changes, update your dataset and repush.

5. Merge. After CI is green and a maintainer approves, the PR is merged. Your dataset goes live under community/. The repository-wide catalogue is regenerated automatically on main (see §8).

6. (Optional) Promotion to curated. After review, a maintainer may promote your contribution out of community/ up to the model directory. The {hf_username}_ prefix is stripped but your attribution is preserved in the catalogue's original_contributor column.

6.3 What a good contribution looks like

• Reproducible. Include sufficient metadata in extra_data to reproduce the solution (solver, tolerance, sampling seed, pipeline parameters).
• Deduplicated. Run db.load_local_vacua(model=model) locally and deduplicate against your existing solutions before pushing.
• Focused label. One label per physical question (e.g. racetrack_small_W0, nonSUSY_dS_candidates), not a generic dump.
• Tagged. Tags should describe filters a user might query on (mode, Nmax, solver type, post-selection criteria).
• Within tadpole bound. All solutions must satisfy the D3-brane tadpole. Solutions that saturate it should be flagged in extra_data.

6.4 What is not a valid contribution

• Solutions for a geometry that is not in aschachner/cy-database (contribute the geometry first).
• Solutions that fail client-side validation (fix them first).
• Duplicate solutions already in the catalogue (check with db.list_hub_vacua(...) before pushing).
• Solutions whose provenance cannot be verified (no committed_by).

6.5 Retraction policy

If a solution in the vault is later found to be incorrect, the contributor or maintainer can retract it:

db.retract_designated(
    designated_id=12345,
    reason="F-term residual exceeds 1e-6 with corrected period data",
)

Retraction is reversible: the file stays on disk, the catalogue entry is flagged retracted=True, and a row is appended to retractions.parquet at the repo root. Default queries skip retracted entries.

Irreversible purges (file deletion) are reserved for maintainer action under explicit policy (see CONTRIBUTING.md on the repo). They require both dry_run=False and an explicit confirm=True safeguard.

---

7. Validation and failure handling

Every contribution goes through two layers of validation:

Layer 1 — Client-side, before push. push_vacua_to_hub() runs locally:

• Schema version compatibility (matches schema.json).
• Flux vectors are integers of the expected length.
• Tadpole constraint $\mathrm{\mid }{f}^T\mathrm{\Sigma }h\mathrm{\mid }\le N_{\max }|f^T\; \backslash Sigma\; h|\; \backslash le\; N\_\{\backslash max\}$ (when $N_{\mathrm{m}\mathrm{a}\mathrm{x}}N\_\{\backslash mathrm\{max\}\}$ is known).
• F-terms for SUSY-flagged rows: tol.
• No duplicates against the local vault (and, if check_remote=True, the remote catalogue).
• committed_by is non-empty.

Failures raise ValidationError locally — nothing is uploaded.

Layer 2 — Server-side, on the PR branch. A CI workflow runs python -m stringforge.vacuavault validate on every PR. It repeats all Layer 1 checks (no trust in client-side validation), reconstructs the model from identity fields, verifies F-terms to a tight tolerance, and cross-checks against the existing catalogue. Output:

• A per-file summary comment on the PR.
• A machine-readable validation_report.json attached as a PR artifact.
• A blocking status check — merges are blocked until CI is green.

7.1 Failure modes and remediation

Case	What happens	How to fix
Whole-dataset failure (schema error, malformed parquet, identity mismatch)	CI comments the root-cause error. PR merge blocked.	Fix locally, re-push. CI auto-reruns.
Partial failure, partial_failure="strict" (default)	CI lists the failing row indices with errors. PR merge blocked.	Fix or drop the bad rows, re-push.
Partial failure, partial_failure="split"	Valid rows go to the main file; failing rows move to _rejected/{label}.rejected.parquet with a per-row error column. Main file passes CI; rejected file is catalogued with status="rejected".	No fix required — but review the rejected file and consider retracting/removing it in a follow-up PR.
Duplicate against existing catalogue	CI reports which existing entries your rows match.	Remove the duplicates locally or include only novel rows.
Post-merge discovery of bad data	Use retract_designated() (see §6.5).	File stays on disk (audit); catalogue entry flagged; retractions.parquet updated.

7.2 Communication channels

Channel	Purpose	Who reads
PR comment (CI bot)	Primary feedback, immediate, inline	Contributor + reviewers
validation_report.json (PR artifact)	Machine-readable diagnostics	Tooling, scripts
Catalogue status column	Per-dataset state: "pending", "curated", "rejected", "retracted"	Downloaders (via query filters)
retractions.parquet (repo root)	Permanent audit trail of retractions	Maintainers, auditors
HF Community tab	Open-ended questions (not blocking)	Anyone

---

8. Catalogue automation

The repository-wide catalogue (catalog.parquet at the repo root) is never hand-edited. It is a derived index — regenerated automatically from the parquet files in the repo.

8.1 PR preview (non-destructive)

During PR CI (§7), after validation passes, a preview of the post-merge catalogue is computed and posted as a PR comment. Example:

This PR adds 2 datasets (1,247 rows total) to tdf/h12_2/ks_00029_tri_0000/community/ and proposes 0 retractions.

Reviewers see the impact at a glance before merging.

8.2 Post-merge rebuild (automatic)

On every push to main (i.e. after a PR merge), a separate CI workflow triggers and runs:

python -m stringforge.vacuavault rebuild_catalog

This scans all **/*.parquet files, extracts per-file metadata (path, label, contributor, row count, schema version, checksum, status), and commits a refreshed catalog.parquet under the [catalog-bot] author. No human intervention is required.

8.3 Manual fallback

If the automated workflow fails, the maintainer can rebuild the catalogue locally:

git clone https://huggingface.co/datasets/aschachner/vacua_vault
cd vacua_vault
python -m stringforge.vacuavault rebuild_catalog --repo-path .
git commit -am "rebuild catalog" && git push

Users browsing the catalogue use db.list_hub_vacua(...) which reads catalog.parquet directly — they never need to know how it was produced.

---

9. Related datasets & code

• Geometry catalogue: aschachner/cy-database — intersection numbers, $c_{2}c\_2$, Kähler cone, GV/GW invariants, conifold data.
• Code (database/vault I/O): aschachner/stringforge — LCSDatabase, the vault writer, and the server-side vacuavault validation / catalogue tooling.
• Code (physics): aschachner/jaxvacua — flux vacuum generation, period integration, and post-selection.

---

10. Citation

If you use this dataset in a publication, please cite both the dataset and the paper(s) that produced the contribution you used. Dataset citation:

@misc{vacua_vault_2026,
  author       = {Schachner, A. and community contributors},
  title        = {{CY Vacua Vault: a community-curated database of flux
                  vacuum solutions for Calabi--Yau compactifications}},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/datasets/aschachner/vacua_vault}},
}

For individual contributions, the committed_by, label, and optional extra_data.paper fields identify the source work.

---

11. Licence

Content of this dataset is released under CC-BY-4.0. You are free to use, share, and adapt the data with attribution to the contributor(s) (as recorded in committed_by) and to this dataset.

Individual contributions may additionally reference a paper under a separate licence; cite the paper where applicable.

---

12. Changelog

See CHANGELOG.md for a per-release summary of schema changes, new curated datasets, and notable community contributions.

---

Dataset maintainer: @aschachner. Issues and questions: open a discussion on this repository's Community tab.