Record the duplicate rows upstream leaves in the fs training split

README.md

@@ -162,6 +162,13 @@ pinned revisions. It needs `rdkit` and `pandas`.
 - Reaction subtasks are keyed on the reaction the question names, not on every
   molecule it mentions. Common solvents and reagents recur throughout both
   splits and keying on them would empty the training set for no gain.
+- Duplicates *inside* the training split are left alone, because the rule here
+  is about the test split. One place has many: upstream `fs_major_product.json`
+  holds 2,788 rows but only 1,164 distinct ones, so 1,624 rows of the `fs`
+  training split are byte-identical copies of another row, and `id` is not
+  unique there. Every other file is clean. Drop them with
+  `train.to_pandas().drop_duplicates()` if repeated rows would skew your
+  training mixture.
 - Removal is reported, never silent. If a subtask's training split is too small
   after removal to be useful, the table above says so before you train on it.