| # CityShiftBench NeurIPS ED Review Artifact |
|
|
| This file defines the minimum artifact expected for anonymous NeurIPS |
| Evaluations and Datasets review. |
|
|
| ## Required Review Contents |
|
|
| - Paper source: `paper_no150_v4/main.tex` and `paper_no150_v4/sections/*.tex`. |
| - NeurIPS style files: `paper_no150_v4/neurips_2026.sty` and |
| `paper_no150_v4/neurips_2026_formatting/`. |
| - Table and figure builders: |
| `paper_no150_v4/tools/build_neurips_ed_v8_assets.py` and |
| `paper_no150_v4/tools/build_neurips_ed_full_review.py`. |
| - Core data registries: `data/tile_manifest_122.csv`, |
| `data/slices/paper_slice_registry_122.csv`, and active integrity records. |
| - Main result summaries: Scale-122 baseline summaries, SATCA summaries, |
| paired significance CSVs, ablation CSVs, shortcut-control CSVs, and table |
| LaTeX files under `paper_no150_v4/tables/`. |
| - Documentation: `docs/DATA_CARD.md`, `docs/BENCHMARK_CARD.md`, |
| `docs/COMPUTE_RESOURCES.md`, `docs/LICENSES.md`, and |
| `docs/CITYSHIFTBENCH_CROISSANT_METADATA.json`. |
| - Croissant metadata: the review artifact should include core Croissant fields, |
| source/provenance fields, and Responsible-AI fields, including whether the |
| benchmark contains synthetic data. |
|
|
| ## Anonymization |
|
|
| For double-blind review, the public artifact should avoid author names in: |
|
|
| - repository owner or organization name, |
| - README badges or contact blocks, |
| - commit metadata in a submitted archive when feasible, |
| - absolute local paths in documentation. |
|
|
| If an anonymized GitHub repository is used, the OpenReview submission should |
| link to that repository directly. If the repository is not ready, upload an |
| anonymous ZIP artifact through OpenReview and include the same contents listed |
| above. The manuscript PDF itself should remain anonymous; code and data links |
| belong in the OpenReview metadata fields or anonymous supplement. |
|
|
| ## Large Files |
|
|
| GitHub should not be used for raw OSM JSON dumps, large Sentinel rasters, |
| Photoshop/PSB editing files, or generated preview PDFs. Release these through a |
| dataset host such as Hugging Face, Dataverse, OpenML, or Kaggle. If any asset |
| exceeds 4 GB, include a small sample that reviewers can inspect. |
|
|
| ## Rebuild Commands |
|
|
| From the project root: |
|
|
| ```bash |
| python paper_no150_v4/tools/build_neurips_ed_v8_assets.py |
| python paper_no150_v4/tools/build_neurips_ed_full_review.py |
| ``` |
|
|
| For full experiment reruns, use `docs/COMPUTE_RESOURCES.md` and |
| `docs/SERVER_SCALE122_RUNBOOK.md`. |
|
|
|
|
