Stage v4 tagged release (part 35)

Staging commit for the v4 tagged release on branch v4-tagged, created from main at 4466afa8092fb42f172202ac10b7cbda376e9f00.

DATASHEET.md

@@ -213,7 +213,7 @@ Yes. The release pipeline:
    `<subset>/<split>/<subset>_<split>_<NNNN>.aselmdb` files in `ms_id` order.
    Copy every row byte-for-byte from the files of step 4.
 6. Tag every triplet and sort it (README section Tags and tier2). Remove the 1,841,368 triplets whose
-   reactant and product are the same state or whose barrier is not positive. Move the 13,838,269 doubtful triplets to
+   reactant and product are the same state or whose barrier is not positive. Move the 13,838,269 flagged triplets to
    `tier2/`. This step adds keys to `info`, renames `eigenmode` and `curvature`, and leaves every other field of a row
    unchanged.
 

README.md

@@ -44,7 +44,7 @@ and saddles to 0.05 eV/Å; for the NEB subset (`mp20bat`) see
 [How the data was produced](#how-the-data-was-produced). Every row stores its `uma-s-1p2` **energy and forces**. Each saddle row also stores the approximate eigenmode of its
 search (`approx_eigenmode`). It also stores the three lowest eigenvalues and eigenvectors of the UMA Hessian
 (`eigenvalues3`, `eigenmodes3`). Tags on every row describe the geometry and energies of the triplet. 18,197,618
-triplets sit in the main files, and 13,838,269 doubtful triplets sit in `tier2/` (see [Tags and tier2](#tags-and-tier2)).
+triplets sit in the main files, and 13,838,269 flagged triplets sit in `tier2/` (see [Tags and tier2](#tags-and-tier2)).
 
 The files are already divided into `train/`, `val/` and `test/` directories, with
 **no chemical system (element set) shared between the three splits**. See
@@ -76,7 +76,7 @@ The files are already divided into `train/`, `val/` and `test/` directories, wit
 | `mp20bat/`| Materials Project battery structures        | NEB-CI | 20,017| 18,305| 863| 849| 3|
 | **Total** |                                             |        | **18,197,618**    | **16,393,932**| **896,373**| **907,313**| **685**|
 
-`tier2/` holds the doubtful triplets of each subset, with the same split and file names:
+`tier2/` holds the flagged triplets of each subset, with the same split and file names:
 
 | Subset    | Source                                      | Method | # transition states | train      | val       | test      | Files |
 |-----------|---------------------------------------------|--------|--------------------:|-----------:|----------:|----------:|------:|
@@ -246,7 +246,7 @@ identifiers are collected in the `source_id` column of
 ## Tags and tier2
 
 We tagged every triplet with its geometry, its energies, and the UMA Hessian at its saddle. With these tags, we removed
-the triplets that a fault of our pipeline produced and moved doubtful triplets to `tier2/`. The tags stay on every
+the triplets that a fault of our pipeline produced and moved flagged triplets to `tier2/`. The tags stay on every
 row, so a user can apply other limits.
 
 | Condition | Test | Action |
@@ -260,6 +260,11 @@ row, so a user can apply other limits.
 The push removed 1,841,368 triplets of the previous release (commit 4466afa8); one triplet can meet both removal
 tests. A removal test overrides a `tier2/` test. The tags `pr` and `n_mobile_*` remove or move no triplet.
 
+A triplet in `tier2/` fails at least one of the three `tier2/` tests. It is not an error. The tests mark saddles that
+the current data cannot confirm as a first-order saddle between two distinct minima, for example a saddle with two
+negative modes at the 0.05 eV/Å force limit. We expect many of these saddles to pass after a tighter refinement, and
+we plan to refine them.
+
 The counted eigenvalues are the eigenvalues of the UMA Hessian over the free atoms. In a cell with no fixed atom
 (`lemat`, `mp20bat`), a uniform translation of all atoms changes no energy, so we project out exactly these three
 translations before the count. λ1 ≤ λ2 are the two lowest counted eigenvalues, and `n_neg_mode` counts those below −ε.

croissant.json

@@ -296,7 +296,7 @@
   "prov:wasGeneratedBy": "Data collection: Starting structures were drawn from four public source datasets (LeMat-Bulk, Open Catalyst 2020, Open Catalyst 2022, and a battery-relevant subset of Materials Project). For the three dimer subsets (lemat, oc20, oc22), each search starts from a single source equilibrium structure plus a recipe-based directed displacement that seeds a saddle search; reactant and product are then obtained constructively by double-minimization from the converged saddle. For the mp20bat subset, reactant/product pairs are constructed from working-ion structures and connected by a climbing-image nudged-elastic-band (CI-NEB) band. All energies and forces are produced by Meta's uma-s-1p2 machine-learning interatomic potential via the SaddleMill orchestration package (https://github.com/SciLM-ai/SaddleMill). The released dataset contains no DFT labels; DFT (VASP) was used only as an external evaluation reference, refining the 1,011 of 1,560 sampled triplets across the four subsets that pass three filters (974 saddles converge), not for any released structure or label. Calculations were performed in March 2026 - May 2026 on HPC systems at TACC and NERSC. Preprocessing: Dimer saddles are converged to 0.05 eV/A and their endpoints, obtained by double-minimization, to 0.02 eV/A; only dimer triplets whose three frames converged are kept. mp20bat NEB bands are kept when the climbing image converged (0.05 eV/A) and the band has no intermediate minimum; 189 kept bands have other images unconverged, and 225 triplets have an endpoint whose relaxation stopped above 0.02 eV/A. Each dimer saddle stores the converged dimer direction (approx_eigenmode) and its curvature (approx_curvature); each NEB saddle stores the climbing-image tangent as approx_eigenmode. Each saddle also stores the three lowest eigenvalues and eigenvectors of the UMA Hessian (eigenvalues3, eigenmodes3). Bond-change annotations were derived from connectivity differences between reactant and product. Near-duplicate triplets (saddle and both endpoints within 0.2 A maximum per-atom displacement, same composition and lattice) are removed. Train/validation/test splits are 90/5/5 by element set: every chemical system (the set of elements, adsorbate included) is assigned whole to one split, so no chemical system or reduced formula appears in two splits and sibling triplets from the same parent structure never cross splits; every element in val/test also occurs in train. Annotation: Fully automated, no human annotation. All labels (energies, forces, eigenmodes, curvatures, bond changes, convergence flags) are produced by the MLIP calculations.",
   "rai:dataCollection": "Aggregated transition-state structures derived from four public source datasets: LeMat-Bulk (inorganic crystals), Open Catalyst 2020 (OC20), Open Catalyst 2022 (OC22), and a battery-relevant subset of Materials Project (MP20bat). For the three dimer subsets (lemat, oc20, oc22), each search starts from a single source equilibrium structure plus a recipe-based directed displacement; reactant and product are obtained constructively by double-minimization from the converged saddle. For the mp20bat subset, reactant/product pairs are constructed from Materials-Project working-ion structures and connected by a climbing-image nudged-elastic-band (CI-NEB) band. All energies and forces are produced by Meta's uma-s-1p2 machine-learning interatomic potential via the SaddleMill orchestration package (https://github.com/SciLM-ai/SaddleMill). The released dataset contains no DFT labels. All labels in this dataset (energies, forces, approximate eigenmodes and curvatures, UMA-Hessian eigenvalues and eigenvectors, geometry and energy tags, bond changes, convergence flags) are synthetic in the sense that they are computed from machine-learning-interatomic-potential simulations rather than experimentally measured; atomic structures are likewise simulated configurations, not experimental observations.",
   "rai:dataCollectionTimeframe": "March 2026 - May 2026",
-  "rai:dataPreprocessingProtocol": "Dimer saddles are converged to 0.05 eV/A and their endpoints, obtained by double-minimization, to 0.02 eV/A; only dimer triplets whose three frames converged are kept. mp20bat NEB bands are kept when the climbing image converged (0.05 eV/A) and the band has no intermediate minimum; 189 kept bands have other images unconverged, and 225 triplets have an endpoint whose relaxation stopped above 0.02 eV/A. Each dimer saddle stores the converged dimer direction (approx_eigenmode) and its curvature (approx_curvature); each NEB saddle stores the climbing-image tangent as approx_eigenmode. Each saddle also stores the three lowest eigenvalues and eigenvectors of the UMA Hessian (eigenvalues3, eigenmodes3). Bond-change annotations derived from connectivity differences between reactant and product. Near-duplicate triplets (saddle and both endpoints within 0.2 A maximum per-atom displacement, same composition and lattice) are removed. Train/validation/test splits are 90/5/5 by element set: every chemical system (the set of elements, adsorbate included) is assigned whole to one split, so no chemical system or reduced formula appears in two splits and sibling triplets from the same parent structure never cross splits; every element in val/test also occurs in train. Each triplet carries geometry, energy, and UMA-Hessian tags. Triplets whose reactant and product are the same state or whose barrier is not positive are removed; doubtful triplets are in tier2/.",
+  "rai:dataPreprocessingProtocol": "Dimer saddles are converged to 0.05 eV/A and their endpoints, obtained by double-minimization, to 0.02 eV/A; only dimer triplets whose three frames converged are kept. mp20bat NEB bands are kept when the climbing image converged (0.05 eV/A) and the band has no intermediate minimum; 189 kept bands have other images unconverged, and 225 triplets have an endpoint whose relaxation stopped above 0.02 eV/A. Each dimer saddle stores the converged dimer direction (approx_eigenmode) and its curvature (approx_curvature); each NEB saddle stores the climbing-image tangent as approx_eigenmode. Each saddle also stores the three lowest eigenvalues and eigenvectors of the UMA Hessian (eigenvalues3, eigenmodes3). Bond-change annotations derived from connectivity differences between reactant and product. Near-duplicate triplets (saddle and both endpoints within 0.2 A maximum per-atom displacement, same composition and lattice) are removed. Train/validation/test splits are 90/5/5 by element set: every chemical system (the set of elements, adsorbate included) is assigned whole to one split, so no chemical system or reduced formula appears in two splits and sibling triplets from the same parent structure never cross splits; every element in val/test also occurs in train. Each triplet carries geometry, energy, and UMA-Hessian tags. Triplets whose reactant and product are the same state or whose barrier is not positive are removed; flagged triplets are in tier2/.",
   "rai:dataAnnotationProtocol": "Fully automated. No human annotation. All labels (energies, forces, approximate eigenmodes, UMA-Hessian eigenvalues and eigenvectors, geometry and energy tags, bond changes, convergence flags) are computed from first-principles or ML-potential calculations. No annotator demographics, annotation platform, or inter-annotator agreement scores are applicable.",
   "rai:dataUseCases": "Construct: the dataset is intended to represent a broad sample of converged transition states for solid-state and surface chemistry, with saddle, energy, and eigenmode labels at the uma-s-1p2 level. Demonstrated in the companion paper: (1) training a generative transition-state model (SaddleFlow, https://github.com/SciLM-ai/SaddleFlow), whose predictions cut the force calls of a Sella saddle search 4.6-8.8x; (2) warm-starting DFT saddle searches, where 974 of 1,011 filtered saddles reconverge with DFT in a median of 31-56.5 force calls per subset. Further intended uses, not yet validated: DFT labelling of near-saddle structures for MLIP training, and benchmarking saddle-search and generative methods. Validity has NOT been established for: safety-critical decision making, prediction outside the chemistry covered by the source datasets, or reactions involving long-range solvation, magnetic ordering, or strongly correlated electron systems beyond the approximations of the underlying DFT and MLIP.",
   "rai:dataLimitations": "Coverage is biased toward systems represented in the source datasets (catalytic surfaces from OC20/OC22 and inorganic crystals from LeMat-Bulk and MP20bat); molecular and biochemical reactions are underrepresented. Convergence thresholds may exclude high-barrier or floppy systems with shallow Hessian eigenvalues. All energies and forces in this release are uma-s-1p2 MLIP predictions with no DFT refinement, so residual MLIP errors persist throughout the dataset. On a DFT sample of 974 converged saddles across the four subsets, saddle geometries agree with DFT to a median 0.038 A maximum per-atom displacement and per-atom energies to a mean absolute error of 4 meV/atom, but activation energies have a mean absolute error of 143 meV, as broad scatter rather than the systematic underestimate (softening) reported for earlier universal MLIPs. Energies and forces inherit, through the uma-s-1p2 task embedding, the approximations (functional, basis, pseudopotentials) of the DFT datasets behind each UMA task (OMat24 for lemat and mp20bat, OC20, OC22). Not recommended for: experimental decision-making, safety-critical chemistry, or any setting that requires absolute reaction rates without DFT correction.",