YOLO26n object detector for archival index cards. Extends the NLS v1 detector to handle multi-card scans and cards from more institutions, with multi-scale training for input-size robustness.
Model Details
Model Description
Fine-tuned from NationalLibraryOfScotland/archival-index-card-detector/model.pt (YOLO26n) on a mixed-collection dataset of 1,425 archival scans spanning four institutions. Compared to v1, this checkpoint:
handles multi-card scans (2–9 cards arranged on a sheet, e.g. Navy biographical record series)
generalises across institutional card styles (NLS, BPL printed, Rubenstein handwritten, Navy typed)
is trained at random imgsz ±50% per batch (multi_scale=True), so it tolerates input scans from ~640px up to ~1500px without re-tuning the inference imgsz
Single class: card.
Developed by:Daniel van Strien, Machine Learning Librarian, Hugging Face
Model type: Object detection (YOLO26n, single class, 2.5M params, ~5.5 MB)
Language(s): en (cards are English; model is language-agnostic visually but evaluated on English archives)
License: AGPL-3.0 (inherits from upstream Ultralytics / NLS baseline)
Run the model on any archival scan to locate index cards. Returns one bounding box per detected card. Pair with a downstream OCR/VLM model for content extraction.
Triage — quickly identify which pages in a digitised collection contain card-like content vs other materials
Card counting — automated tallies for collection-level metrics
Out-of-Scope Use
OCR / content extraction — this is a detection model. It says where the card is, not what it says.
Content classification — single class. Does not distinguish blank from content, manuscript from typed, etc. For blank/content filtering see small-models-for-glam/index-card-blank-detector.
Non-English cards — training data is English-only.
Card-style traditions outside US/UK archives — not validated on continental European library catalogs, East Asian indices, etc.
Bias, Risks, and Limitations
No true non-card negatives in training beyond the 10 NLS background pages. The model may over-predict on newspaper clippings, photographs, or book pages when those appear in mixed archival scans.
Multi-card variety concentrated in 25 Navy images. Very different multi-card layouts (12+ card grids, severely overlapping or rotated cards) are out-of-distribution.
BPL + Rubenstein training labels are auto-generated (bbox = whole image for pre-cropped sources). Boxes on cropped-card inputs are loose by construction. If you need tight per-card pixel-accuracy, validate on a held-out set first.
Multi-scale training trades 1–3% in-distribution mAP@50:95 for scale robustness. v4's bboxes are slightly looser than v3's (single-scale checkpoint also published at index-card-detector-v3 for comparison). For downstream OCR this is acceptable; for tight cropping you may prefer v3.
Recommendations
For mixed-content archives (cards + photos + clippings), apply a downstream content classifier or train v5 with explicit negatives.
For tight pixel-accurate cropping, evaluate v3 vs v4 on your data and pick.
For non-English collections, fine-tune with additional samples from that tradition.
The ONNX export is dynamic on the spatial axes (height, width), so it accepts variable input sizes without re-export. Use best.pt if you need GPU acceleration or further fine-tuning; use best.onnx for CPU-only deployment.
285 held-out images (20% of training set, stratified per collection). No external test set used.
Metrics
mAP@50 and mAP@50:95 (standard COCO-style detection metrics), reported per collection.
Results
Collection
n val
mAP@50
mAP@50:95
NLS Advocates
20
0.995
0.857
Navy Nurse Corps
5
0.986
0.764
BPL Catalog
160
0.995
0.991
Rubenstein Manuscript
100
0.995
0.993
mAP@50 is essentially saturated across all collections — the model finds every card. mAP@50:95 is lower on NLS and Navy because multi-scale training trades bbox tightness for scale robustness (see Bias/Risks above).
For comparison, the single-scale index-card-detector-v3 checkpoint trained on the same data scores higher mAP@50:95 (NLS 0.980, Navy 0.929) but is less robust on out-of-distribution input sizes.
Summary
v4 is the recommended checkpoint for general use across varied input sizes. Use v3 if you need maximum bbox tightness on archives that match the training distribution.
Environmental Impact
Training carbon footprint is small — ~30 min on a single A100. Approximate emissions: ~0.05 kgCO₂eq (based on Lacoste et al. 2019 calculator, averaged grid mix). Inference on CPU is ~0 emissions.
Technical Specifications
Model Architecture and Objective
YOLO26n — single-stage anchor-free detector. 260 layers, 2.5M parameters, 5.8 GFLOPs. Three detection heads at strides 8, 16, 32 over a CSP-Darknet-style backbone.
Compute Infrastructure
Hardware: HF Jobs a100-large (1× A100-SXM4-80GB)
Image: vllm/vllm-openai:latest (used as host for CUDA-12-compatible PyTorch — see training script for context)
