Feedback wanted on healthcare document AI dataset
Hey all,
This is a synthetic Australian medical document library: 50-doc free sample (CC-BY-NC), modelled on NSW Health practice, with bbox layout labels and 4 scan-quality tiers per doc. It's the first public release and I'd value feedback from anyone using it for layout-aware extraction, OCR robustness training, or AU-specific clinical NLP.
Three specific things I'm trying to decide before the full 5K library finalises:
Bbox format
Currently stored as (x, y, width, height, page) in PDF points, recorded by the generator at render time. Every labelled field has its bbox alongside the structured value (50 docs, 682 total bboxes, average 13.6/doc).
Would (x0, y0, x1, y1) corner format be more useful? Or normalised [0-1000] scale a la LayoutLMv3? I'm happy to ship multiple formats in the manifest if there's a clear preference, or add a small conversion helper.
Scan-quality tier distribution
The sample has 16 clean / 27 scanned / 6 poor / 1 fax. The full 5K library uses the same ratio.
Is that distribution useful for real-world OCR robustness training (heavier on "scanned" because that's what actually shows up in clinical document workflows), or would a balanced 25/25/25/25 split be more useful as an explicit benchmark profile? Should I also ship an "extreme degradation" 5th tier (heavy occlusion, severe skew, low DPI)?
HF Datasets loading script
The sample is currently files-only (PDFs + manifest.json + per-doc JSON). I'm considering adding a loading script so:
from datasets import load_dataset
ds = load_dataset("RootCauseAnalytics/synthetic-australian-medical-documents-sample")
would return a structured Dataset with PDF bytes plus parsed labels in one call.
Would that be useful? Or do most folks prefer raw files for direct PDF processing with their own pipeline (pdfplumber, pdf2image, etc)? Open to either.
Any other format / schema / metadata feedback also very welcome. The full 5K library can incorporate anything that lands in this thread before commercial release.
Cheers,
GhostCanary