Most illustration-mining pipelines filter out pages that contain text and keep image-only plates; this dataset shows why that loses work: 82% of the flagged illustrated pages retain ≥100 OCR words, so illustrations are usually embedded in text columns rather than presented as standalone plates.
What Sets It Apart
- Large, provenance-rich image split: 115,293 pages predicted illustrated (≈9.6 GB of 600px JPEGs) drawn from 975,345 scored pages across 1,160 volumes, enabling vision training at scale while keeping direct links to the source Internet Archive pages.
- Full manifest for reproducibility: the 975,345-row
manifestcarries classifier score, OCR word counts, provenance, edition inference flags and storage/tier metadata — so you can re-threshold, compute per-edition rates, or build custom splits without re-scoring. - OCR-aware labels and repairs: each page includes an OCR word count (
words) and astratum(text/low) used to set different classifier thresholds; 242 volumes had page-order mismatches that were repaired and flagged, preserving correct page–OCR binding. - Classifier provenance and thresholds: pages were scored with an EfficientViT-based detector trained on a 2,573-page labeled set; thresholds were validated separately for
textandlowstrata (precision ~1.0 on small validation samples), so labels are high-precision predictions rather than human annotations.
Who It's For — fit and tradeoffs
Great fit if you need training or evaluation data for document-vision tasks that must handle illustrations embedded in text (classification, detection, retrieval, multimodal historical analysis). The manifest supports dataset curation, per-edition analysis, and streaming workflows.
Look elsewhere if you require human-verified labels for every image (these are classifier predictions with small validation samples), or if you need guaranteed global edition metadata (edition labels are partly inferred from year and some year fields are null). Practical notes: the pages config with images is large (stream recommended), manifest is lightweight for analytics, and licensing depends on source IA items though underlying works are public domain by publication date.
Technical summary: two configs (pages with images, manifest with scores/provenance), coverage 1768–1929, classifier was efficientvit_b1-derived and trained on National Library of Scotland labels, and images point back to Internet Archive source items for attribution and retrieval.
