Add card + splits + metadata

README.md

@@ -0,0 +1,65 @@
+---
+license: other
+tags:
+  - chemistry
+  - electron-density
+  - qm9
+  - molecular-dynamics
+size_categories:
+  - 100K<n<1M
+---
+
+# OrbFlow data
+
+Preprocessed training and evaluation data for **OrbFlow** (equivariant flow matching for atomic-orbital
+coefficients and electron density).
+
+- Code: https://github.com/TODO/OrbFlow
+- Checkpoints: [divelab/OrbFlow](https://huggingface.co/divelab/OrbFlow)
+
+## Contents
+
+The layout matches the `DATAPATH` folder expected by the OrbFlow scripts.
+
+| Path | Size | Description |
+|------|------|-------------|
+| `datasplits.json` | 1 MB | QM9 train / validation / test split |
+| `lmdb/` | 1.4 TB | QM9 LMDB: geometries and density probes, 134 shards (+ `metadata.json`) |
+| `lmdb_gt_ridge/` | 12 GB | QM9 target orbital coefficients (β = 1.3, ridge 1e-6), one shard per `lmdb/` shard |
+| `MD/scdp_lmdb/<mol>/` | 24 GB | MD LMDBs for benzene, ethane, ethanol, malonaldehyde, phenol, resorcinol (each with its own `datasplits.json`) |
+| `MD/scdp_lmdb_gt_ridge_beta1.3/<mol>/` | 0.6 GB | MD target orbital coefficients (β = 1.3, ridge 1e-6) |
+| `MD/flow_prior_ridge1e6_beta1.3/<mol>/` | < 1 MB | Per-molecule flow-prior µ/σ used by the MD training recipe |
+
+## Download
+
+Everything (≈ 1.46 TB):
+
+```bash
+hf download divelab/OrbFlow-data --repo-type dataset --local-dir "$DATAPATH"
+```
+
+Only part of it, e.g. the MD data (≈ 25 GB) or the QM9 target coefficients:
+
+```bash
+hf download divelab/OrbFlow-data --repo-type dataset --local-dir "$DATAPATH" --include "MD/*"
+hf download divelab/OrbFlow-data --repo-type dataset --local-dir "$DATAPATH" --include "lmdb_gt_ridge/*" "datasplits.json"
+```
+
+## How the data was produced
+
+- **QM9 LMDB** — built from the [DTU QM9 VASP charge densities](https://data.dtu.dk/articles/dataset/QM9_Charge_Densities_and_Energies_Calculated_with_VASP/16794500)
+  with the [SCDP](https://github.com/kyonofx/scdp) preprocessing (bond-midpoint virtual nodes, atom cutoff 6 Å):
+  `vision/preprocess_qm9_lmdb.slurm` in the OrbFlow code. It can be rebuilt from the DTU tarballs instead of downloaded.
+- **Target coefficients** — per-molecule ridge fit of the DFT density on the def2-QZVPPD-derived Gaussian basis
+  (β = 1.3, ridge 1e-6, full probe grid, float64): `vision/prepare_flow_gt_coeffs_ridge1e6_gpu.slurm` (QM9) and
+  `vision/prepare_md_gt_coeffs_ridge1e6_beta13_gpu.slurm` (MD).
+- **MD prior** — mean / std of the train-split target coefficients: `vision/compute_md_flow_prior_stats_beta13.slurm`.
+- **MD densities** — TODO: source and citation.
+
+> **Alignment:** `lmdb_gt_ridge/` is matched to `lmdb/` by shard file name and entry order, not by molecule ID.
+> Use it with the `lmdb/` from this repository, or with an LMDB rebuilt by the unmodified script from all 134
+> original DTU tarballs. A mismatched LMDB makes training stop with a clear error.
+
+## License and citation
+
+TODO: license (must be compatible with the DTU QM9 dataset and the MD source data) and citation.

lmdb/metadata.json

@@ -0,0 +1 @@
+{"target_mean": 0.07815587520599365, "target_var": 0.10440606623888016, "avg_num_neighbors": 17, "unique_atom_types": [0, 1, 6, 7, 8, 9]}