Dataset Curator

Overview We're building a labeled evaluation corpus for a PII detection system that scans files across many formats. We need a meticulous data researcher to find, verify, and organize files from open-licensed public datasets — images of identity documents, tax forms, medical records, structured data exports, code files, and logs. This is a research and organization role. You are not generating, creating, or modifying files. You are finding files that already exist under open licenses and delivering them in a structured, documented format. What You'll Do - Search open data repositories (Zenodo, HuggingFace Datasets, Kaggle, data.gov, SEC EDGAR, PhysioNet, IRS.gov, USCIS.gov, GitHub) for specific document types - Verify each file has an acceptable open license (CC0, CC-BY, or public domain) — document the source URL for every file - Rename files descriptively and organize into a structured folder hierarchy - Fill out a sources.csv for each file: path, format, document type, polarity (positive/hard negative), source URL, license, notes - Flag gaps when no licensed file can be found for a deliverable category — document what you searched What You'll Be Finding Organized into 12 deliverable categories across 4 phases: - Images (Phase 1 — highest priority): US driver's licenses (IDNet/Zenodo CC0), passports (MIDV-500/CC0), insurance cards, scanned financial documents, hard negative blanks - PDFs (Phase 2): Financial filings, medical records, blank government forms (IRS/USCIS — public domain) - Structured data (Phase 2): CSV/parquet datasets with personal identifiers - Code & logs (Phase 3): Open-licensed source files and log samples with technical identifiers We provide starting sources for each category — you're extending and filling gaps, not starting from scratch. Required Skills - Experience navigating academic and open data repositories (Zenodo, HuggingFace, Kaggle, government portals) - Solid understanding of Creative Commons licensing — you must verify licenses, not assume them - Strong organizational discipline: consistent file naming, complete CSV documentation, no gaps left undocumented - Comfort working with diverse file types: images, PDFs, CSVs, JSON, code, log files - Clear written English for README notes and gap reports Nice to Have - Familiarity with document datasets used for ML/AI (DocVQA, RVL-CDIP, IDNet, MIDV-500) - Basic Python or command-line experience (useful for batch downloading Zenodo ZIPs) - Understanding of privacy/PII concepts — helps judge whether a file is a positive or hard negative Deliverables Per Batch - ZIP of files with descriptive names organized by format - sources.csv — one row per file with all required columns - README.txt — batch notes, new sources discovered, gap report Scope: ~480–560 files across 12 deliverable categories, delivered in phased batches (batch 1 after S-01 through S-05 are complete). Ongoing engagement.