- notebooks/fa_lg_trf_training.ipynb: trains fa_dep_news_lg/fa_core_news_lg
(floret vectors, byte-identical recipe to md, only --paths.vectors differs)
and fa_dep_news_trf/fa_core_news_trf (HooshvareLab/roberta-fa-zwnj-base,
fine-tuned via spacy-transformers; not ParsBERT, whose model card carries
no licence). Self-contained: downloads UD_Persian-PerDT from public GitHub
URLs, expects the repo source (git archive) and an lg floret wheel uploaded
by hand since Gitea is LAN-only and floret training happens elsewhere.
- scripts/finalize_pipeline.py: extend --size to lg/trf. lg's vectors note is
generated from the trained model's actual vector table shape rather than
hardcoded, since that table is still being iterated on; trf records the
transformer source instead of a floret one.
- configs/fa_dep_news_md.cfg, configs/fa_ner_md.cfg, scripts/unpack_vectors.py:
existing dependencies the notebook calls directly, committed so a
git-archive-based upload is self-sufficient.
