{
 "deposited_utc": "2026-10-07T11:23:10Z",
 "code_commit_at_deposit": "85e8bde61ad4aebd2bbd08c6184f9009fdb9305e",
 "follows": "model/deposit/manifest.json, whose tests were run on 2026-10-07 and are kept in model/results/",
 "why": [
  "In the first deposit a structure type could carry two labels, because the labels used Wyckoff letters and those depend on the choice of origin. Compounds of a held-out type could stay in training under the other label: 15 compounds of the cubic A2MH6 type did, 9 of them labelled hydrides.",
  "A compound stored in the release as two records could have one record on each side of a split: 5 of the 68 held-out A2MH6 compounds had a second record in training.",
  "Both faults were found while the results of the first deposit were being written up, after its tests had been run. Those results are reported beside the results of this deposit."
 ],
 "changes": [
  "Structure types are labelled by space group, stoichiometry and the site symmetry of each occupied orbit (model/structure_types.py), which does not depend on the origin and is coarser than the Wyckoff label.",
  "Only the first record of each compound (reduced formula and structure type) is used; later records are dropped from training and from scoring.",
  "The Hopfield sum S of the whole compound is scored for the graph network on every held-out compound against the training-side geometric mean of S, without a mark.",
  "Per-compound predictions are stored for every split."
 ],
 "unchanged": [
  "The label, the smearing, the three baselines for h, the mark, the extrapolation requirement, and the settings of both estimators (trees: 600 iterations at most, learning rate 0.04, 24 leaves, early stopping; graph: width 96, four layers, 32 distance features to 5 A, 60 epochs, batch 96, learning rate 0.002, one network per split with seed 0, or the fold index for the folds)."
 ],
 "compounds": 27151,
 "compounds_kept": 26058,
 "structure_types": 1740,
 "hydrides_with_h_kept": 4552,
 "structure_types_of_labelled_hydrides": 653,
 "splits": {
  "top": "structure types ranked by their largest h; the leading ones are held out with all their members until a tenth of the labelled hydrides are held out",
  "family": "every compound of the structure type of Mg2IrH6 (225_1[m-3m]_2[-43m]_6[4m.m]) is held out",
  "grouped": "five folds of whole structure types, shuffled with seed 0",
  "random": "five folds of compounds, shuffled with seed 0; reported without a mark"
 },
 "counts": {
  "top": {
   "train_compounds": 25303,
   "train_hydrides": 4022,
   "test_compounds": 755,
   "test_hydrides": 530,
   "test_structure_types": 13
  },
  "family": {
   "train_compounds": 25967,
   "train_hydrides": 4480,
   "test_compounds": 91,
   "test_hydrides": 72,
   "test_structure_types": 1
  },
  "grouped_0": {
   "train_compounds": 21545,
   "train_hydrides": 4114,
   "test_compounds": 4513,
   "test_hydrides": 438,
   "test_structure_types": 137
  },
  "random_0": {
   "train_compounds": 20846,
   "train_hydrides": 3650,
   "test_compounds": 5212,
   "test_hydrides": 902,
   "test_structure_types": 252
  },
  "grouped_1": {
   "train_compounds": 22964,
   "train_hydrides": 3827,
   "test_compounds": 3094,
   "test_hydrides": 725,
   "test_structure_types": 123
  },
  "random_1": {
   "train_compounds": 20846,
   "train_hydrides": 3647,
   "test_compounds": 5212,
   "test_hydrides": 905,
   "test_structure_types": 276
  },
  "grouped_2": {
   "train_compounds": 21013,
   "train_hydrides": 3513,
   "test_compounds": 5045,
   "test_hydrides": 1039,
   "test_structure_types": 139
  },
  "random_2": {
   "train_compounds": 20846,
   "train_hydrides": 3632,
   "test_compounds": 5212,
   "test_hydrides": 920,
   "test_structure_types": 242
  },
  "grouped_3": {
   "train_compounds": 16647,
   "train_hydrides": 3146,
   "test_compounds": 9411,
   "test_hydrides": 1406,
   "test_structure_types": 132
  },
  "random_3": {
   "train_compounds": 20847,
   "train_hydrides": 3648,
   "test_compounds": 5211,
   "test_hydrides": 904,
   "test_structure_types": 286
  },
  "grouped_4": {
   "train_compounds": 22063,
   "train_hydrides": 3608,
   "test_compounds": 3995,
   "test_hydrides": 944,
   "test_structure_types": 122
  },
  "random_4": {
   "train_compounds": 20847,
   "train_hydrides": 3631,
   "test_compounds": 5211,
   "test_hydrides": 921,
   "test_structure_types": 271
  }
 },
 "mark": "mean absolute error in ln h at most half that of the best of the three baselines, with the 95% interval of the ratio, from resampling held-out structure types (compounds for the family split), below one",
 "extrapolation_requirement": "in the top split, at least half of the held-out hydrides whose h exceeds every training value are predicted above every training value, and at most a tenth of the others are",
 "checksums": {
  "labels.csv": "1c16759a7fe5daff5a2b73a1735993613c0664f22c27e7424f4e5c7e91427b9c",
  "structure_types.csv": "25299632a4f72c3ea40fcfc2d9de24aec3633c625f82c3c8bef8989505b73fce",
  "splits.csv": "1bd5bf2c1041b5d071a1ae9c1519e5dd02fdaacd4c8867d9778c14b8a0053f5f",
  "baselines.csv": "ff02955be27c4ba0404e5252eb8d679995fd400cc9e341ff18ce701f2c6e1f67"
 }
}