jurebes 🐾

Just-sklearn Utility for Reproducible Evaluation of Baselines, Estimators and Solvers.

A classical-ML text classification research framework, built on plain scikit-learn. It uses no NLTK and no padacioso.

Jurebes wires any sklearn featurizer to any sklearn classifier behind a small text-classification API. It also includes a registry of ready-to-use baselines, a benchmark harness, dataset loaders, a CLI, and an OVOS pipeline plugin.

Intent classification is the main use case, but the core works with any single-label text task. Spam, sentiment, topic, and language-ID tasks run on the same API. See the general text classification guide.

Named in memory of Jurebes, the best dog.

Install

pip install jurebes
# optional extras
pip install jurebes[hf]              # HuggingFace dataset loader
pip install jurebes[search-bayes]    # Bayesian hyperparameter search (skopt)
pip install jurebes[search-genetic]  # Genetic-algorithm search (sklearn-genetic-opt)
pip install jurebes[search-all]      # both bayes and genetic
pip install jurebes[bench-plot]      # matplotlib-based comparison plots
pip install jurebes[test]            # pytest stack

Quickstart

from jurebes import IntentClassifier, BASELINES

clf = IntentClassifier(BASELINES.build("logreg"))
clf.add_intent("hello", ["hello", "hi", "hey there"])
clf.add_intent("joke",  ["tell me a joke", "say a joke", "make me laugh"])
clf.fit()

result = clf.predict("hi there")
print(result.intent, result.confidence, result.entities)

Pass any sklearn Pipeline or estimator instead of a baseline name. Jurebes wraps non-probabilistic estimators with CalibratedClassifierCV automatically, so predict_proba always works.

Documentation

Full documentation lives under docs/:

• Getting started: install, first classifier, core concepts, troubleshooting.
• Guides: choosing a baseline, slots, calibration, reproducibility, debugging.
• Theory: classical-ML foundations: featurization, linear models, kernels, ensembles, statistical comparison.
• API reference: every public module, function, dataclass, and CLI flag.

Research

Jurebes is built as a research framework:

• 48 named baselines tagged into groups (linear, kernel, tree, naive_bayes, neural, reduced_dim, online, strategy, feature_engineering, ensemble, discriminant, categorical). This includes neural-bottleneck autoencoder pipelines (autoencoder_*) and dict-input categorical pipelines (categorical_*).
• A hyperparameter search subsystem (jurebes.search) with grid, random, successive-halving, Bayesian (optional), and genetic (optional) backends.
• A benchmark harness with multi-metric scoring (f1_macro, accuracy, log_loss, top_k_accuracy, ...) and pooled tail-latency percentiles.
• See docs/research.md and docs/search.md.

Baselines

BASELINES is a registry of 48 named factories. It covers naive Bayes, logistic regression, linear/RBF SVMs, kNN, online learners (SGD, perceptron, passive-aggressive, ridge), tree ensembles (random forest, extra trees, gradient boosting, HistGBM, bagging), a shallow MLP, voting, stacking, reduced-dim methods (LSA/NMF/LDA/autoencoder), multi-class strategy wrappers (OvR/OvO), discriminant analysis, text-statistics composites, and dict-input categorical pipelines. List them:

jurebes list-baselines

Registry table

Rows total more than 48 because some baselines (for example bagging_logreg, sgd_log, hashing_sgd_hinge) belong to multiple groups. The registry itself holds 48 unique factories.

The categorical_* pipelines accept list[dict[str, str]] instead of raw text. They are available through the programmatic API only, and text-fixture benchmarks skip them.

Register your own:

from jurebes import BASELINES
BASELINES.register("my_pipeline", lambda: build_my_sklearn_pipeline())

Benchmark

jurebes benchmark --dataset data.csv \
    --baselines logreg,nb_multinomial,linear_svc --cv 5 \
    --out report.md

Or programmatically:

from jurebes.benchmark import compare, to_markdown
from jurebes.datasets import load_csv

X, y = load_csv("data.csv")
result = compare(["logreg", "linear_svc", "nb_multinomial"], X, y, k=5)
print(to_markdown(result))

RunResult captures accuracy, macro/micro F1, per-class F1, training time, predict-latency p50/p95/p99, model size, and the confusion matrix.

Slots

Five pluggable taggers cover dictionary, template, sklearn IOB, hybrid cascade, and CRF (optional extra). All are available through the TAGGERS registry and share the same protocol.

from jurebes import IntentClassifier, BASELINES
from jurebes.slots import TAGGERS, SklearnIOBTagger

clf = IntentClassifier(BASELINES.build("logreg"), tagger=SklearnIOBTagger())
# or by name:
clf = IntentClassifier(BASELINES.build("logreg"), tagger="hybrid")

clf.add_entity("name", ["bob", "alice"])
clf.add_intent("name", ["my name is {name}", "call me {name}"])
clf.add_intent("hello", ["hello", "hi"])
clf.fit()
clf.predict("my name is bob").entities  # -> {"name": "bob"}

TAGGERS.names()  # ['dictionary', 'template', 'sklearn_iob', 'hybrid', 'crf']

See docs/ for the research guide, slot tagger details, and the OVOS deployment notes:

• docs/research.md: adding baselines, benchmarks, reports.
• docs/search.md: hyperparameter search backends.
• docs/slots.md: IOB slot tagger and token features.
• docs/opm.md: OVOS pipeline plugin configuration.
• MIGRATION.md: porting from the prior JurebesIntentContainer API.

Examples

Runnable, cell-marked scripts live in examples/notebooks/:

• snips_quickstart.py: fetch SNIPS, train, evaluate.
• banking77_full_research_flow.py: compare baselines, run Friedman+Nemenyi, tune the winner, evaluate on a holdout set.

Both examples need pip install jurebes[hf,bench-plot].