📰 Institutional Newspapers Crop Classifier (Text)

A text-based classifier that categorizes crops extracted from historical newspaper scans into high-level categories. This model is a Model2Vec fine-tune of minishlab/potion-base-32m with a classifier head, which makes it light and efficient.

We recommend using this model alongside institutional/institutional-newspapers-crop-classifier-image-yolo26m-cls, as neither visual nor textual signal alone is sufficient to classify newspaper crops accurately.

More information:

📄 Full analysis available in our technical report

See also:

The Institutional Data Initiative at the Harvard Law School Library works with knowledge institutions—from libraries and museums to cultural groups and government agencies—to refine and publish their collections as data. Reach out to collaborate on your collections.

Evaluation results

Evaluated on a held-out test set of 31,015 crops:

Class	Precision	Recall	F1-Score	Support
Advertisement	0.95	0.94	0.95	11,244
Cartoon	0.85	0.57	0.68	430
Content	0.90	0.96	0.93	11,245
Masthead, nameplate or running head	0.93	0.86	0.89	2,250
Photograph or illustration	0.63	0.38	0.48	551
Section heading	0.94	0.93	0.94	5,295

Metric	Score
Accuracy	0.92
Macro avg F1	0.81
Weighted avg F1	0.92

Training data

This model was trained on part of Boston Public Library's newspapers collection.

Total annotated crops: 185,900
Classes: 6
Split: ~67% train / ~17% val / ~17% test

Usage

from model2vec.inference import StaticModelPipeline

model_name = "institutional/institutional-newspapers-crop-classifier-text-model2vec"
model = StaticModelPipeline.from_pretrained(model_name)

texts = ["CLASSIFIED ADVERTISING — Rooms for rent, furnished apartments..."]
predictions = model.predict(texts, max_length=None)

print(predictions)
# ['Advertisement']

Recommended inference parameters

Parameter	Value	Note
`max_length`	`None`	Do not truncate input text

Citation

@misc{cargnelutti2026institutionalnewspaperspipelinederiving,
      title={Institutional Newspapers Pipeline: Deriving billions of high quality tokens from historical newspapers}, 
      author={Matteo Cargnelutti and Catherine Brobston and Eben English and Jake Sadow and Kacie Bailey and Greg Leppert and Amanda Watson and Jessica Chapel and Jonathan Zittrain},
      year={2026},
      eprint={2608.18972},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2608.18972}, 
}