romansh-idiom-classifier 1.0.0

Romansh idiom classifier — Sursilvan, Sutsilvan, Surmiran, Puter, Vallader, Rumantsch Grischun, German, French, Italian, English

romansh-idiom-classifier

Python Romansh idiom classifier. Given a text, it identifies which of the six Romansh idioms it is written in — Sursilvan, Sutsilvan, Surmiran, Puter, Vallader, or Rumantsch Grischun — and also distinguishes German, French, Italian, and English.

Label	Language
rm-sursilv	Sursilvan
rm-sutsilv	Sutsilvan
rm-surmiran	Surmiran
rm-puter	Puter
rm-vallader	Vallader
rm-rumgr	Rumantsch Grischun
de	German
fr	French
it	Italian
en	English

Pure Python inference — no scikit-learn required at runtime. All four trained models are bundled. Trained in farscrl/romansh-idiom-identification.

Installation

pip install romansh-idiom-classifier

Usage

from romansh_idiom_classifier import RomanshIdiomClassifier

classifier = RomanshIdiomClassifier()  # default: full LR model (best overall accuracy)

idiom = classifier.predict("L'uolp era puspei inagada fomentada. Cheu ha ella viu sin in pégn in tgaper che teneva in toc caschiel en siu bec.")
print(idiom)  # "rm-sursilv"

Choosing a model

classifier = RomanshIdiomClassifier()              # "lr" — full LR, best overall (default)
classifier = RomanshIdiomClassifier(model="svm")   # full SVM, best on schoolbook text
classifier = RomanshIdiomClassifier(model="lr-lite")   # smaller LR (10k/5k vocab)
classifier = RomanshIdiomClassifier(model="svm-lite")  # smallest and fastest

Model	Avg accuracy	Avg macro-F1	JSON size
lr (default)	0.948	0.937	75 MB
svm	0.949	0.939	14 MB
lr-lite	0.943	0.930	4.0 MB
svm-lite	0.944	0.930	2.1 MB

Accuracy averaged over test sets A–D (Romansh idioms). All four models are bundled (~95 MB total).

With raw scores

score() returns unbounded real numbers — positive means evidence for that idiom, negative means evidence against. The gap between the highest and lowest score reflects confidence.

scores = classifier.score("L'uolp era puspei inagada fomentada...")
# { "rm-sursilv": 14.57, "rm-surmiran": 1.31, "rm-sutsilv": -0.79, ... }

With soft scores (for confidence bars)

soft_scores() applies softmax and returns values between 0 and 1 summing to 1. Useful for displaying a confidence bar chart. Not calibrated probabilities — use score() when you need to reason about model certainty.

soft = classifier.soft_scores("L'uolp era puspei inagada fomentada...")
# { "rm-sursilv": ~1.0, "rm-surmiran": ~0.0, ... }

From a custom model file

classifier = RomanshIdiomClassifier(model="path/to/lr_export.json")

API

RomanshIdiomClassifier(model=None)

Creates a classifier. model can be:

• None or "lr" — bundled full LR model (default)
• "lr-lite", "svm", "svm-lite" — other bundled models
• A file path string to a JSON export
• A pre-parsed dict

classifier.predict(text: str) -> str

Returns the single most likely idiom label.

classifier.score(text: str) -> dict[str, float]

Returns a raw decision score per class. Unbounded reals — positive = evidence for, negative = evidence against.

classifier.soft_scores(text: str) -> dict[str, float]

Returns softmax-normalised scores between 0 and 1, summing to 1. For display use. Not calibrated probabilities.

Verification

The Python inference is a faithful reimplementation of the sklearn pipeline. Predictions should match the sklearn reference results within the float32/float64 precision gap (~0.001 threshold), and are bit-for-bit identical to the TypeScript package on every test set.

Run the classifier on the test sets

cd packages/romansh-idiom-classifier-py

# Default model (LR) — save results to JSON
python scripts/verify.py --model lr \
  --output ../../data/04_evaluation/results_LR-py.json

# LR-lite
python scripts/verify.py --model lr-lite \
  --output ../../data/04_evaluation/results_LR-lite-py.json

You can restrict to specific test sets:

python scripts/verify.py --model lr --tests test_a test_b

Compare Python package vs sklearn reference

# From the repository root
python tools/compare_results.py \
  data/04_evaluation/results_LR.json \
  data/04_evaluation/results_LR-py.json

All differences should show ≈ (below the 0.001 threshold). Any larger delta indicates a discrepancy in the inference implementation.

Compare Python package vs TypeScript package

python tools/compare_results.py \
  data/04_evaluation/results_LR-lite-ts.json \
  data/04_evaluation/results_LR-lite-py.json

Results should be identical — both implementations share the same tokenisation logic.

License

MIT