small-models-for-glam/index-card-detection-v5

Dataset Card for Archival Index Card Detection — v5 (ensemble-relabelled navy)

Refined version of small-models-for-glam/index-card-detection-v3. All NLS / BPL / Rubenstein rows are passed through unchanged. The 25 navy-nurse-corps rows have their bounding boxes re-labelled via a v3+v4 model ensemble plus human review, replacing the SAM3-only bootstrap from v3.

Dataset Details

Dataset Description

Same 1,425-row mixed-collection composition as v3. The only difference: the 25 navy multi-card scans now use higher-quality bounding boxes produced by:

1. Running both small-models-for-glam/index-card-detector-v3 and small-models-for-glam/index-card-detector-v4 on each navy image (per-model conf threshold 0.10).
2. Greedy IoU-NMS at 0.5 to merge near-duplicate detections from the two models, recording which models contributed to each surviving box (box_source = "v3+v4", "v3", or "v4").
3. Human review in a custom HTML bbox editor, prioritising rows where the two models disagreed (rows with v3-only or v4-only boxes appeared first in the review queue).

Outcome: 86 raw ensemble boxes (83 unanimous + 3 disagreement), reduced to 84 after one envelope-style v4 artefact was removed and 2 boxes were structurally adjusted.

• Curated by: Daniel van Strien, Machine Learning Librarian, Hugging Face
• Language(s): en
• License: Mixed — see Source Data table below

Dataset Sources

• Repository: https://huggingface.co/datasets/small-models-for-glam/index-card-detection-v5
• Predecessor: https://huggingface.co/datasets/small-models-for-glam/index-card-detection-v3
• Trained model: https://huggingface.co/small-models-for-glam/index-card-detector-v5

Uses

Direct Use

Train an object detector to find archival index cards on full pages, multi-card sheets, or pre-cropped single-card images. Schema matches v3.

Out-of-Scope Use

Same as v3: detection only (not OCR), single class, English-only, US/UK archival conventions only.

Dataset Structure

Identical schema to v3:

Column	Type	Description
image	Image	RGB scan, original resolution
source_collection	string	nls-advocates / navy-nurse-corps / bpl / rubenstein
source_repo	string	Source HF dataset id
source_row_id	int64	Index in source dataset
source_url	string	Original item link (when present)
text	string	Source OCR / filename
objects.bbox	list[[float, 4]]	xywh in original pixel coordinates
objects.category	ClassLabel(["card"])	Always 0
objects.box_source	list[string]	nls-original / v3+v4 / v3 / v4 / human / auto-full

Per-collection: NLS 100 / navy 25 / bpl 800 / rubenstein 500 = 1,425 rows.

Dataset Creation

Curation Rationale

v3's navy boxes were SAM3-bootstrap + human review at the start of the project. By the time v3 and v4 were trained, both models had developed their own implicit understanding of where cards sit in navy scans. Comparing the two models' predictions surfaced a small number of cases where the original human labels were under- or over-inclusive — and ensemble + targeted review fixed these without re-labelling from scratch.

Source Data

Collection	Source	Rows	Label provenance (v5)	Change vs v3
NLS Advocates	NationalLibraryOfScotland/nls-index-cards-object-detection	100	Original NLS bboxes (10 has_card=False negatives)	unchanged
Navy Nurse Corps	biglam/index-cards-navy-nurse-corps	25	v3+v4 ensemble + human review of disagreements	re-labelled
BPL Catalog	biglam/bpl-card-catalog	800	Auto bbox = whole image	unchanged
Rubenstein Manuscript	biglam/rubenstein-manuscript-catalog	500	Auto bbox = whole image	unchanged

Annotations

Annotation process (navy re-labelling)

1. Both index-card-detector-v3 and index-card-detector-v4 ran on the 25 navy images on GPU (HF Jobs l4x1) at confidence threshold 0.10 (low enough to surface weak signals; fusion handles noise).
2. All detections from both models were merged with IoU-NMS at threshold 0.5. Each surviving box was tagged with the set of contributing models.
3. The 25 images were sorted in the review UI: rows with the most single-model boxes (= model disagreement = ambiguity) appeared first.
4. Human reviewer (Daniel van Strien) corrected disagreements: removing one envelope-style v4 artefact, refining 2 boxes. Most boxes (unanimous from v3+v4) passed through unchanged.

Final box mix: 56 v3+v4 unanimous, 27 human-touched (small adjustments to position/size), 1 v3-only kept.

Bias, Risks, and Limitations

Same as v3:

• English-only
• US/UK archival conventions only
• Multi-card variety concentrated in 25 navy images
• No true non-card negatives beyond 10 NLS background pages
• BPL + Rubenstein labels are rule-based, not human-verified

The ensemble re-labelling does not address any of these — it only refined within-distribution navy labels.

Citation

@dataset{vanstrien_archival_index_cards_v5_2026,
  author    = {van Strien, Daniel},
  title     = {Archival Index Card Detection — mixed collections (v5)},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/small-models-for-glam/index-card-detection-v5}
}

Dataset Card Authors

Daniel van Strien