PortNOIE clean release
Copyright (C) 2021-2026 Bruno Souza Cabral, Marlo Vieira dos Santos e Souza and
Daniela Barreiro Claro

This directory is a clean redistribution prepared from the canonical PortNOIE
source code formerly published as FORMAS/dptoie-neural. It deliberately does
not reuse the original Git history.

The original repository contained an empty file named LICENCE. Its
pyproject.toml declared the malformed license identifier "GPL-3.". On
2026-08-30, maintainer and co-author Bruno Souza Cabral explicitly confirmed
that he is authorized to publish this clean release under GPL-3.0-only,
including the historical checkpoint and the 14 embedded Flair Diário matrices.
This notice records that confirmation; third-party components retain their own
licenses and attribution requirements.

The model.th, model_params.pkl and vocabulary distributed through the public
Hugging Face repository bratao/PortNOIE are historical research
artifacts. The model_params.pkl file is a Python pickle and is used only by the
explicit legacy backend; load only the official, checksum-verified copy. Wheels
and source distributions contain code and the small metadata sidecar, not the
checkpoint. The first extraction downloads the checkpoint at the library's
pinned revision and verifies package-owned hashes; later calls reuse the cache.

The experimental modern backend reconstructs the graph from model.th without
AllenNLP or Flair at runtime. The 287-character index mapping absent from the
checkpoint is preserved in src/portnoie/modern_data/flair_metadata.json. That
sidecar records the source URLs and SHA-256 checksums of the historical
forward-1024.pt and backward-1024.pt assets from which it was derived. The 14
language-model tensors embedded in model.th were validated as byte-for-byte
tensor matches with those source assets. The complete source assets are not
redistributed.

The modern migration has run both a real annotated sentence and raw text with
the real spaCy model, but has not yet been
shown numerically or behaviorally equivalent to the AllenNLP pipeline. The
spaCy pt_core_news_lg model remains an external dependency for raw-text
annotation. The legacy backend additionally requires the complete Flair assets.

The Docker image redistributes the spaCy `pt_core_news_lg` 3.8.0 model. That
model is published by Explosion/spaCy under CC BY-SA 4.0 and includes data
derived from Universal Dependencies Portuguese Bosque. License:
https://creativecommons.org/licenses/by-sa/4.0/ . Model source:
https://github.com/explosion/spacy-models .

No claim is made that the historical LSTM+Flair checkpoint is the best
BERT model reported in the associated thesis.
