Changelog¶

Unreleased¶

examples/ folder with five complete runnable scripts, and an Examples page in the documentation.
Zenodo DOI (10.5281/zenodo.23832590) added to the README and CITATION.cff.

Documentation site (MkDocs Material): installation, quickstart, SDK, command-line and HTTP API guides, tutorials, configuration and API reference, concepts, results, FAQ and roadmap.
python -m sindhi_nlp.evaluation.evaluate: score any model on any labeled CSV, list the most confident mistakes.
sindhi_nlp.evaluation.error_analysis and sindhi_nlp.evaluation.plots (confusion matrix, model comparison, per-class F1). These modules existed but were empty.
sindhi-nlp --version and python -m sindhi_nlp.
py.typed, so type checkers use the package's type hints.
Optional extras plots and docs.
Docstrings on every public function and class.
CONTRIBUTING.md, SECURITY.md, CODE_OF_CONDUCT.md, issue and pull-request templates, Dependabot, docs-deploy and PyPI release workflows.

sindhi-nlp finalize-model failed after a plain pip install because configs/base.yaml is not installed. It now falls back to the built-in cleaning defaults (identical to configs/base.yaml) and says so in preprocessing.json.
A missing configs/... file now produces an error that explains the folder is part of the GitHub repository and is not installed by pip.

evaluate_external prints a one-line Error: message (exit code 1) instead of a traceback when the one-time test rule refuses a run.
PLAN.md moved to docs/archive/.

The README on PyPI now shows the pip install instructions and working links to the documentation.
The Hugging Face upload tool now tells network problems apart from login problems, instead of reporting every failed login check as "not logged in".
The model is named after its Hugging Face repository (not its cache folder) when loaded by name.

First release: data pipeline, TF-IDF baseline, XLM-RoBERTa training, Python SDK, command-line tool, HTTP API, and publishing tools.