Datasets:

forzagreen
/

ar-conjugation

data/json/verbs.json        495 records, the canonical form (pretty-printed, sorted as the book was read)
data/jsonl/verbs.jsonl      the same records, one JSON object per line
data/csv/conjugations.csv   one row per inflected form: id,lemma,root,wazn,form,tense,voice,person,variant,text
data/jsonl/conjugations.jsonl  the same rows as JSON objects (the Hugging Face "cells" config)
data/csv/verbs.csv          one row per paradigm: id,lemma,root,wazn,form,past_vowel,nonpast_vowel,full_passive,
                            reduced,root_type,verb_type,model,template,page,printed_page,masdar,
                            active_participle,passive_participle
data/json/errata.json       corrected misprints, deliberately empty cells, printing conventions
scripts/build_dataset.py    rebuilds data/ from a checkout of the private source repository
scripts/validate.py         shape, counts, NFC, csv/json agreement (CI)
scripts/push_hf.py          mirrors data/ + README to Hugging Face
scripts/serve.py            serves site/ locally with data/ mounted as on GitHub Pages
site/                       the browsing page

Fields

Field	Meaning
`id`	Stable identifier, `NNN_MMM`: reading order in the upstream corpus, then the book's model number.
`lemma`	The citation form: 3rd person masculine singular of the active past, as printed.
`root`	The root, bare letters. Hamza is always `ء`; a final ي covers both ي and ى.
`wazn`	The pattern, named the way Arabic grammar names it: form I carries its past and non-past vowels (`فعَل يفعُل`), the other forms their shape (`استفعل`). 22 values.
`form`	Western form number: `I`–`XIII` for triliteral verbs, `Iq`–`IVq` for quadriliteral.
`vowels`	Form I only: the past and non-past stem vowels, `a` / `i` / `u`.
`flags.full_passive`	The page prints a personal passive for persons other than 3ms. Set on 48 paradigms (an engine needs telling, since transitivity cannot be read off the root).
`flags.reduced`	The page prints the assimilated shape of an affix: forms V/VI with the تَ- assimilated into a coronal first radical (اِسَّمَّعَ for تَسَمَّعَ), form VII with the ن assimilated into م (اِمَّدَحَ), form IIIq with the ن assimilated into a weak ل1 (اِهْبَيَّخَ). 10 paradigms.
`classification.root_type`	Root-structure class id (1–55) per the scheme in the upstream `GLOSSARY.md` (Paper 1).
`classification.verb_type`	Augmentation code `[R]+A[-Fn]` per the upstream `GLOSSARY.md` (Paper 2): radicals + added letters, then the form within that group, e.g. `3+2-III`.
`source.model`	The book's own model number for the paradigm (its classification of the verb).
`source.template`	`A`: one of the book's 83 abstract-model pages, five columns (active and passive). `B`: one of the two-verb pages, three columns (active only). `null`: read from the 2nd edition before the page inventory existed and not determinable.
`source.page`	The page's index in the 6th-edition scan (pdfimages numbering).
`source.printed_page`	The page number printed on the page. The offset between the two is not constant (14 to 17), so neither is ever computed from the other; a `null` means it was not recorded.
`source.label`	The upstream fixture label, kept for traceability.
`conjugation.<column>`	`past`, `past_pass`, `ind` (المضارع المرفوع), `ind_pass`, `imp`. `null` = the page has no such column. Otherwise an object keyed by person; a person absent from it is a cell the page leaves blank or misprints (see errata).
`derived`	The block printed under the table: `masdar` (list, in the book's order), `active_participle`, `passive_participle`. `null` for the 28 paradigms read before that block was extracted.
`derived.status.*`	How the derived noun was checked. `agrees-with-module`: one OCR pass that agrees with the independent engine. `read-from-page` is not a status but the flag `read_from_page`: the cell was read by a person from the page image, at 4–14×. `differs-from-module`: the book's reading stands and the engine derives something else (6 cells, all explained upstream). `book-only`: the engine derives no such noun. `none`: the book prints none.

Person codes and cell values

3ms 3md 3mp 3fs 3fd 3fp 2ms 2d 2mp 2fs 2fp 1s 1p — third person first, as the book's tables run. The book prints أنتما twice (masculine and feminine dual, identical forms); the dataset has one gender-neutral 2d. The imperative has the five second persons.

A cell is a string. Ten cells are a list of strings: the book prints two alternatives in one cell (اِئْتِ / اِيتِ, عُشَّ / اُعْشُشْ, يَرَى / يَرْأَى, …), in the book's order. In conjugations.csv these become two rows with variant 1 and 2.

Every string is Unicode NFC (so a vowel precedes a following shadda), contains only Arabic letters and marks, and carries no tatweel or bidi controls. scripts/validate.py enforces all of this.

Errata

errata.json has three lists:

deliberately_empty_cells (17): cells the page prints but the dataset leaves out, each with its reason — six compositor slips where the page sets a neighbouring row's form, three where a page contradicts its own facing page of the same model, three genuinely blank (ink measured).
corrected_misprints (4): cells where the dataset departs from the printed page, with the printed form, the corrected one and why.
printing_conventions (7): spellings the book alternates between that are not differences of form.

Rebuilding

python3 scripts/build_dataset.py --source /path/to/source-checkout   # or AR_CONJUGATION_SOURCE=...
python3 scripts/validate.py
python3 scripts/serve.py          # http://localhost:8000/
HF_TOKEN=... uv run --with huggingface_hub scripts/push_hf.py

The build is deterministic; CI runs validate.py on every push, and pages.yml publishes site/ with data/ mounted beside it, so the site's download links serve the committed files.