FAIR release of rampnet-dataset + rampnet-benchmark: DOI, full cards, Croissant 1.1/GeoCroissant/RAI, evaluation protocol, leaderboard (NSF FAIROS)

Part of the NSF FAIROS work (tracker: ProjectSidewalk/SidewalkWebpage ProjectSidewalk/SidewalkWebpage#5053; award #2623243). Objective 3 names
"RampNet 2.0" — a DOI-versioned curb-ramp detection benchmark with fixed splits, a standard evaluation protocol,
and a leaderboard — as a Year-2 deliverable. Most of the substance already exists on Hugging Face
(projectsidewalk/rampnet-dataset: ~214K panoramas / ~850K labels, fixed train/val/test splits;
projectsidewalk/rampnet-benchmark: 9 cities across two imagery providers; the 1K manual gold set; the
v1.1-corrected-eval erratum; docs/replication.md and docs/data_provenance.md). What's missing is the FAIR
packaging. This issue is that packaging.

Tasks

• Persistent identifier. Mint a DOI for rampnet-dataset and rampnet-benchmark (Hugging Face mints a
  DataCite DOI per revision; the card currently says "DOI: forthcoming"). Decide final repo names before
  minting — a minted repo cannot be renamed or deleted. Cite the ICCV'25 paper and the erratum in the record.
• Dataset cards that meet the Hugging Face template fully: composition, collection process, splits and
  their rationale (city-held-out vs. random), the seam-duplicate note, known biases (city mix, imagery
  provider, capture years), intended uses, citation (dataset DOI + paper), maintenance/versioning policy
  (major version on any file change, minor on metadata-only).
• Croissant 1.1 + GeoCroissant 1.0 + RAI metadata authored by hand (croissant.json in the repo and
  uploaded to the HF record; HF's auto-export is Croissant 1.0 only and has no geo/RAI fields). Validate with
  mlcroissant. This also satisfies the NeurIPS Datasets & Benchmarks submission requirement.
• Evaluation protocol as code: one evaluate.py (or rampnet.eval module) that takes predictions in a
  documented format and outputs AP / P / R with the 1:1 matching rule from the erratum, plus a
  leave-one-city-out runner over rampnet-benchmark. Pin the exact split files by content hash.
• Leaderboard: Hugging Face "private evaluator" pattern — public results dataset + Gradio leaderboard
  Space, hidden test labels for the benchmark cities, submissions via PR/dataset upload; mirror the results
  table as JSON in this repo for permanence. (Papers with Code is gone; Codabench is the fallback if code
  submissions become necessary.)
• Carry the review notes from #127 into the published benchmark card; close #58 (gold set as split) and
  #21/#10 by pointing at the record of truth.
• Baselines: publish the RampNet result and at least one open-weight detector baseline (#39 / #51) on the
  leaderboard at launch.

Acceptance

DOI resolves to a complete card; mlcroissant validates the metadata; evaluate.py reproduces the paper's
corrected numbers from the released predictions; the leaderboard shows ≥2 entries; the benchmark page on the
Project Sidewalk data portal links here.

🤖 Generated with Claude Code (claude-fable-5, effort: xhigh)

Start by reading the existing Hugging Face dataset and benchmark cards, docs/replication.md, docs/data_provenance.md, and the v1.1-corrected-eval erratum. Then inspect the planned croissant.json and evaluation entry point (evaluate.py or rampnet.eval), including the fixed split files. Done means the DOI records, metadata validation, corrected evaluation reproduction, and leaderboard satisfy the stated acceptance checklist.