Reference rows carry a formula string but no composition dictionary, so they all fell into other and never appeared beside the submissions they are meant to calibrate. The formula is now parsed with pymatgen, which handles bracketed notation that a regular expression gets wrong.
Keep reference materials in their own division instead of falling to other
Masking previously copied the whole record and deleted a few keys, so any field added later left through the same door. The public row is now assembled from a fixed list of permitted keys, with exact composition and computed energies attached only when the entrant chose to publish.
Remove every field that can reconstruct a private composition from the leaderboard response
The masked formula hid the element counts but the composition object beside it carried them verbatim, and the measured energies would have allowed the same reconstruction. Both, along with the atom count, are now stripped after the division is assigned and before the response is returned.
link the LEADBOARD benchmark alongside the sister challenge
The migration cell was still being drawn after that axis left the header row, so every score sat one column to the right - the novelty header showed the empty migration cell and read as unscored, while the figures themselves were fine. Cells and headers now match.
An entry missing one axis was previously left with no score at all, discarding figures we already had. The axes that were produced are now recorded and shown against their own maximum. Rank is withheld while any axis is outstanding - a partial total cannot be set beside a complete one - and the row states what the entry needs.
Say what the entrant should do, not that the entry is pending
An entry we could not put a figure on was labelled 'held', which reads as though waiting would resolve it. It does not. The label now names the action instead - structure needed, no reference data, no value - and the note says plainly that waiting changes nothing and that including a structure file clears most cases.
An entry we could not obtain a value for now stays in the table but holds no rank, the same treatment reference materials get. Ranking something that was never measured is not a verdict. The row is marked, the reason opens on click, and the count is reported separately from the scored entries.
Entries are grouped by the anion that frames them - fluoride, chloride/bromide/iodide, sulfide, oxide, nitride, hydride. Scores are unchanged and identical across divisions; the split exists so a candidate can be read against its own chemical family rather than against every family at once. Reference materials stay visible in every view.
Hugging Face sign-in, submitter profile links, and three scored axes
Entries are tied to a Hugging Face account and the submitter's handle links to their profile. Scoring carries three axes: oxidation window 40, lithium-metal stability 35, use novelty 25. The lithium pathway is an entry condition rather than a scored axis. Reference materials and the frozen reference distribution are rebuilt to match.
The guide and the per-model standings now sit on their own tabs instead of stacking
below the board. The season dates, what each axis looks at, and the submission limits
were all available but never shown; the per-row scoring detail was already in the
response and is now revealed by clicking a row.
Participation guide, prompt set, and per-model standings
