refactor(dataset): organize data and add leaderboard replay
The polynomial-only dataset lacks batch ordering and the separate discriminants needed to replay final scores. This PR supplies those numerical inputs and integrates them into a conventional Hugging Face dataset layout, instead of publishing a separate reproduction bundle alongside top-level shards.
This PR contains the file tree from docs/dataset-layout at 0893dc8cce430bfe5d6cfbc271ad3f034d9074a1. It targets main; main has been restored to the pre-publication contents and must remain unchanged until maintainer approval. No automatic merge or release tag is requested.
Changes
Move the 51 complete polynomial shards to data/records/ and the minimal and missing-pair collections under data/. Put numerical scoring inputs and references in data/scoring/, standalone replay tools in scripts/, detailed instructions in docs/, and export/audit reports in reports/.
Keep the minimal and records Hugging Face configuration names and the minimal default; update only their file paths. Update the dataset card, selective-download and streaming examples, and links. Replace partial manifests with a single root SHA256SUMS covering all 74 other files, including minimal and missing-pair data. Align LFS rules with the relocated files.
All original polynomial payloads and schemas, license terms, access settings, and tags are preserved. Existing commit-pinned downloads still work with their historical paths; consumers using direct file paths on main will need the new paths after merge.
Scoring Scope
The standalone Python replay uses 1,978,899 reduced candidates from the same frozen source snapshot, together with baseline data, stored logs, batch ordering, and tie-breaking metadata. Expected point totals are read only after calculation for comparison. No database, Magma, PARI/GP, or network is needed after download.
Displayed scoreable-pair counts are separately exported primary-data projections, not independently derived from the reduced candidate set. The documentation states this limitation. Member/account details, emails, submission notes, internal build paths, and production extraction scripts are not included.
Validation
Eleven regression tests pass, including 1,000 randomized candidate-reduction cases. Replaying the downloaded review-branch files matches all 163,910 scored pairs and all 140 official leaderboard entries including LMFDB, with zero score/rank differences and zero maximum pair-point difference.
Actual Hugging Face load_dataset streaming tests pass for both configurations: one minimal file and 51 record shards, with coefficient and discriminant strings preserved. Dataset-card validation, document links, file identities, and all checksum entries pass. Main was independently verified to retain the restored contents.
Review Note
README examples currently resolve docs/dataset-layout so reviewers can test the new paths before merge. After approval, they can use main or a pinned release commit. This PR is a snapshot of that branch; subsequent edits must also be applied to the PR ref.