scripts/dedup/README.md

How MaterialsSaddles v2 was produced from v1

These are the scripts that were run, with one change. We removed the 2 broken triplets of step 7 in a separate step after the write. Here, steps 7 to 9 do this removal. The scripts write the same rows and tables as the release. Absolute paths inside them are those of the original run (ROOT = the v1 files, OUT = a work directory) and must be adapted. Dependencies: numpy, scipy, pyarrow, ase, ase_db_backends, lmdb, huggingface_hub; fairchem-core only for one check in v2_verify.py.

Duplicate detection and split (computed on the v1 files)

1. dedup_pass1.py — reads every row once. A triplet is the three consecutive ms_ids 3k, 3k+1, 3k+2 (asserted against the stored side where present). Per triplet it stores the reactant, saddle and product positions, atomic numbers, FixAtoms mask and cell, bucketed by a hash of the exact composition, plus the chemical system, reduced formula and upstream source identifier.

2. dedup_pass2.py — within each exact composition (all subsets together), triplets in ms_id order are leader-clustered: a triplet joins the first cluster whose representative has

   • the same lattice: cell lengths within 2 %, angles within 1°, every cell-matrix entry within 2 % of the longest cell vector;
   • a saddle within 0.2 Å max-atom distance;
   • reactant and product both within 0.2 Å, in the same or the swapped orientation.

   Max-atom distance = largest atomic displacement after per-species Hungarian matching (minimum summed squared distance) under the minimum-image convention, fractional coordinates mapped into the mean cell. The representative of each cluster (its lowest ms_id) is kept.

3. dedup_pass3.py — drops non-representatives and assigns every chemical system (element set of the cell) as a whole to train / val / test with probabilities 0.90 / 0.05 / 0.05 (numpy.random.default_rng(0)), one draw per chemical system, globally across subsets. Writes the per-split ms_id lists and duplicates.parquet (every triplet → its representative).

4. Audits: verify_integrity.py (every v1 row: ms_ids are exactly 0..3N−1, the three rows of a triplet share atomic numbers, cell and pbc, side = ms_id % 3 − 1) and verify_dedup.py (an independent distance implementation re-checks removed triplets against their representative, searches kept triplets against all same-composition triplets for missed duplicates, and checks the split files).

Rewrite into the v2 layout (from the SciLM v1 files, which carry energies and forces)

5. v2_sha.py — checks every downloaded v1 file against the SHA-256 recorded by Hugging Face.
6. v2_index.py — indexes every row of a set of files: ms_id, file, LMDB key, SHA-1 of the stored bytes, and hashes of the geometry and of data.info. Run on the v1 files the deduplication was computed from and on the SciLM v1 files; v2_compare.py then shows both hold the same ms_ids with identical geometry and info, and that every row has finite energy and forces.
7. v2_write.py — writes <subset>/<split>/<subset>_<split>_<NNNN>.aselmdb, at most 50,000 triplets per file, in ascending ms_id order. Rows are not re-serialised: the stored (compressed JSON) value of each row is copied unchanged under the new ASE row id, and nextid is written as ase_db_backends does. v2_write.py cuts the files on the full split lists. Then it skips the 2 broken triplets (saddle ms_id 35584423 and 79800988, forces at the saddle above 100 million eV/Å). As a result, lemat_train_0213 and lemat_train_0479 hold 49,999 triplets each. The other files do not change.
8. v2_verify.py — reads every row of every output file and checks: keys 1..N + nextid; complete, adjacent R, S, P triplets in ascending ms_id order; every stored value byte-identical to the v1 row with the same ms_id; files of a split named consecutively, in ms_id order, all but the last full (except lemat_train_0213 and lemat_train_0479, each one broken triplet short); each split directory holds exactly the ms_ids of its split, less the 2 broken triplets; the union is exactly the kept set, less the 2 broken triplets; no removed or broken triplet present. It also compares a random sample of rows read through ase.db against the v1 rows and opens files with fairchem's AseDBDataset.
9. v2_metadata.py — builds metadata/triplets.parquet and metadata/removed_duplicates.parquet from what v2_verify.py observed in the written files (refuses to run if verification failed). removed_duplicates.parquet lists each removed duplicate with its kept triplet (reason = duplicate) and the 2 broken triplets with no kept triplet (reason = broken). A later commit removed 2 triplets whose cell has 1 atom and added them with no kept triplet (reason = one_atom).