Add pipeline dataset card

README.md

@@ -1,19 +1,47 @@
 ---
-dataset_info:
-  features:
-  - name: smiles
-    dtype: string
-  - name: y
-    list: float64
-  splits:
-  - name: train
-    num_bytes: 12014898
-    num_examples: 176225
-  download_size: 7071406
-  dataset_size: 12014898
-configs:
-- config_name: default
-  data_files:
-  - split: train
-    path: data/train-*
+license: other
+pretty_name: moses_test_scaffolds
+tags:
+- chemistry
+- molecules
+- graph-generation
+- flow-matching
+size_categories:
+- 1M<n<10M
 ---
+
+# nico8771/moses_test_scaffolds — cleaned MOSES (test_scaffolds split)
+
+Each row is a molecule as **canonical SMILES** plus RDKit-recomputed targets. MOSES is
+neutral by construction, so molecules are featurized over **7 atom types** with **no
+formal charges**, and **aromatic bonds are kept as their own class** (no kekulization)
+so the model learns aromaticity directly.
+
+> Source: official MOSES `test_scaffolds.csv.gz` (molecularsets/moses). Code:
+> <https://github.com/Nico-Conti/flow-matching-molecules> (`dataset/`).
+
+## Schema
+
+| column | type | description |
+|---|---|---|
+| `smiles` | string | canonical, single-fragment SMILES (post-sanitize) |
+| `y` | list[float] | RDKit targets, columns = `logP`, `qed`, `SAS` |
+
+## Pipeline
+
+1. **Parse** with RDKit; unparseable dropped.
+2. **Standardize** — remove stereochemistry, sanitize (MOSES is neutral; *no* Uncharger).
+3. **Featurize** over atom vocab (`C`, `N`, `S`, `O`, `F`, `Cl`, `Br`); atoms outside the vocab dropped.
+4. **Round-trip check** — `smiles -> (X, E) -> mol -> smiles`, aromatic bonds preserved.
+
+Bonds use 5 classes (none / single / double / triple / aromatic).
+Targets (`logP`, `qed`, `SAS`) are recomputed from the sanitized SMILES with RDKit.
+
+### Drop / keep counts (this build)
+
+| outcome | count |
+|---|---|
+| `kept` | 155,697 |
+| `kept_no_roundtrip` | 20,528 |
+
+Kept: **176,225** molecules.