The dataset is currently empty. Upload or create new data files. Then, you will be able to explore them in the Dataset Viewer.

Archival index-card blank/content pre-filter

🧪 Draft — hf-model-spec-0.1 · Personal working draft, not an HF convention. Feedback via Community tab.

Purpose

A tiny CPU-runnable binary image classifier that labels an archival index card as blank or content — a cheap pre-filter that skips empty cards before expensive VLM metadata extraction in card-catalogue digitisation pipelines. Companion to the NLS card detector (which crops cards from page scans).

Data

The training dataset has 3 ClassLabel values (blank, content, divider) and 3 splits (train / validation / test, 392 / 69 / 75 rows). For v1, drop the divider class — sample count (38 total) is too small to support a third class reliably, and downstream pipelines treat dividers and blanks the same (both = skip the VLM step).

Two source_collection values: boston_public_library (BPL FRC shelf-list cards) and national_library_of_scotland (NLS Advocates Library cards). Styles differ noticeably (paper colour, hole-punch pattern, typewriter vs handwritten content) — per-collection eval required, not optional.

Attempts to date

Date Approach Result Repo
2026-05-21 MobileViT-XX-Small backbone, transformers Trainer SFT. Backbone sweep also tried MobileNetV2 and ViT-tiny — all hit 100% on gold sets; smallest chosen. Shipped (public). 955.8K params, ~26 ms/card CPU. 100% per-collection on gold (BPL blank recall 1.00, NLS content 1.00); 28/28 on punch-hole/smudge slice; 98/98 OOD on unseen BPL cards across 3 styles. small-models-for-glam/index-card-blank-detector (public)
Downloads last month
-