mpilhlt/canine-salamanca-boundary-classifier

Canine Nonbreaking Line Boundary Classifier

This model identifies nonbreaking line boundaries in early modern Spanish and Latin printed texts — that is, line breaks where a word continues on the next line without a hyphen. It is the first stage in a two-model pipeline for automatic abbreviation expansion in historical texts from the Salamanca School corpus.

Model description

The model is a fine-tuned CANINE-s encoder with a binary classification head. CANINE operates directly on Unicode codepoints without tokenization, making it well suited to historical text where orthographic variation, diacritics, and special characters are meaningful signals. The local attention downsampling in CANINE-s makes it efficient for the short input sequences used here.

Input: the last 40 characters of line N concatenated with the first 40 characters of line N+1, separated by the boundary marker ↵ (U+21B5).

Output: a binary label — 1 (nonbreaking, word continues across the boundary) or 0 (genuine line break).

Training

The model was trained on approximately 1.6 million line boundary examples extracted from TEI XML transcriptions of early modern legal and theological texts in Latin and Spanish, published as part of the School of Salamanca Digital Collection.

• Positive examples (nonbreaking): ~684,000 boundaries identified by editors as word-continuations across line breaks
• Negative examples (breaking): ~918,000 genuine line boundaries within the same paragraphs
• Base model: google/canine-s
• Training hardware: NVIDIA T4 (HuggingFace Jobs)
• Training duration: approximately 5.5 hours
• Optimizer: AdamW, learning rate 2e-5, weight decay 0.01
• Batch size: 32
• Epochs: 5
• Class imbalance handling: BCEWithLogitsLoss with pos_weight capped at 8.0

Training results

Epoch	Loss	Precision	Recall	F1
1	0.4343	0.949	0.838	0.890
2	0.1771	0.932	0.934	0.933
3	0.1135	0.925	0.939	0.932
4	0.0761	0.957	0.916	0.936
5	0.0536	0.949	0.925	0.937

Test set evaluation

Evaluated on a held-out document-level test set (documents not seen during training):

	Precision	Recall	F1	Support
breaking	0.94	0.97	0.96	918,084
nonbreaking	0.96	0.92	0.94	684,090
accuracy			0.95	1,602,174
macro avg	0.95	0.95	0.95	1,602,174
weighted avg	0.95	0.95	0.95	1,602,174

Threshold selection

A classification threshold of 0.0017 was selected from the precision-recall curve to achieve ≥0.90 precision on the validation set, yielding precision 0.900 and recall 0.965 on nonbreaking boundaries. This threshold is saved as threshold.json in the model repository and loaded automatically by the inference code below.

The low raw threshold value (well below 0.5) reflects the class imbalance handling via pos_weight in training — the model's raw logits are shifted, so the sigmoid output should be interpreted relative to this calibrated threshold rather than the naive 0.5 default.

Intended use

This model is intended as a preprocessing step for abbreviation expansion in early modern Latin and Spanish printed texts. It identifies which consecutive line pairs should be concatenated before being passed to a downstream abbreviation expansion model (such as mpilhlt/byt5-salamanca-abbr).

It is not intended as a general-purpose line boundary detector and has not been evaluated on other corpora or languages.

How to use

Installation

pip install transformers torch
# or, to install the full pipeline:
pip install git+https://github.com/digicademy/svsal-poco

Direct use with transformers

import json
import torch
from transformers import CanineTokenizer
from huggingface_hub import hf_hub_download

# If using the svsal-poco package:
from boundary_classifier.boundary_classifier import BoundaryClassifier

# Load model and threshold
model_repo = "mpilhlt/canine-salamanca-boundary-classifier"
tokenizer  = CanineTokenizer.from_pretrained("google/canine-s")

model = BoundaryClassifier(use_lexicon=False)
weights_path = hf_hub_download(model_repo, "best_model.pt")
model.load_state_dict(torch.load(weights_path, map_location="cpu"))
model.eval()

threshold_path = hf_hub_download(model_repo, "threshold.json")
threshold = json.loads(open(threshold_path).read())["threshold"]

# Classify a single boundary
line_end   = "rum prudentium consultum. &c̃. legi tan"   # end of line N
line_start = "tùm adscribitur humanę. Sicuti & illa"    # start of line N+1

text = line_end[-40:] + "↵" + line_start[:40]
enc  = tokenizer(text, return_tensors="pt", max_length=128,
                 truncation=True, padding="max_length")

with torch.no_grad():
    out  = model(input_ids=enc["input_ids"],
                 attention_mask=enc["attention_mask"])
    prob = torch.sigmoid(out["logits"]).item()

is_nonbreaking = prob >= threshold
print(f"Probability: {prob:.4f} | Nonbreaking: {is_nonbreaking}")

Use in the full pipeline

When using the full svsal-poco pipeline, the boundary classifier runs automatically as the first stage of infer.py. See the svsal-poco repository for the complete pipeline documentation.

python infer.py \
  --input             texts.jsonl \
  --output            expanded.jsonl \
  --boundary_model_dir ./canine-salamanca-boundary-classifier \
  --byt5_model_dir    mpilhlt/byt5-salamanca-abbr \
  --batch_size        32

Limitations

• Trained exclusively on texts from the School of Salamanca corpus; performance on other early modern corpora is unknown and may be lower
• Cross-page boundaries (where a word straddles a page break) are included in training data but may show lower performance than within-page boundaries
• The model has no explicit language identification; Latin and Spanish boundaries are handled jointly
• Marginal notes and main text are treated identically; boundaries between different text zones are excluded from training

Citation

If you use this model, please cite the School of Salamanca project and this repository:

@misc{svsal-poco,
  author    = {Wagner, Andreas and others},
  title     = {svsal-poco: Abbreviation Expansion Pipeline for Early Modern
               Spanish and Latin Printed Texts},
  year      = {2025},
  publisher = {GitHub},
  url       = {https://github.com/digicademy/svsal-poco}
}