PseudoPath

PseudoPath trains a named-entity recognition (NER) model, learns a residual line router, and calibrates three recall budgets. Prediction combines the model with fixed rules adapted from EDS-Pseudo and returns spaCy entities.

Three integrated presets keep the common workflow short: camembert, distilcamembert, and tok2vec. You can also supply your own model and adapter without changing the rules or router.

Compare CPU throughput and detection quality across models and routing presets in the interactive benchmark.

Python 3.11 or newer is required. Install from PyPI:

python -m pip install pseudopath
# For CamemBERT or DistilCamemBERT:
python -m pip install 'pseudopath[transformer]'

No separate eds-pseudo installation is needed.

Train, predict, and save in a few lines:

from pseudopath import PseudoPath, Tok2VecTraining, read_jsonl

pipeline = PseudoPath.from_preset("tok2vec")
pipeline.fit(
    read_jsonl("train.jsonl"),
    read_jsonl("validation.jsonl"),
    training=Tok2VecTraining(steps=100, validation_interval=20),
)
entities = pipeline.predict("Luc Martin consulte.", routing="balanced")
pipeline.to_disk("pseudopath-artifact")
restored = PseudoPath.from_disk("pseudopath-artifact")

read_jsonl reads text and annotated character offsets into spaCy documents. See the quickstart for the data format and examples for short workflows with synthetic data. Run them from the repository root, for example:

python examples/train_tok2vec.py
python examples/predict.py

See models and configuration for preset backbones, architecture overrides, and learning controls. For a trained external checkpoint, follow custom models.

fit trains the model and router and calibrates "prudent", "balanced", and "fast" against "all-lines". You can instead set any numeric routing threshold in [0, 1]. A later fit continues the model and, by default, the router. reset_router=True starts a new router. predict(text) and predict(doc) return spaCy spans without changing the input document's annotations.

Read the documentation for installation, training, and inference. The API reference describes public methods, and architecture explains how the pipeline works.

PseudoPath's own code is licensed under GPL-3.0-only. Selected EDS-Pseudo v0.4.0 patterns and derived training behavior retain BSD-3-Clause terms. See LICENSE and THIRD_PARTY_NOTICES.md.