FarhanFarsi/persian_corpus

Persian Corpus (normalized)

Persian web text (blogs, forum posts and Q&A) normalized for language-model training that must match ASR output. The original un-normalized merged_dataset.jsonl (28.7 GB, 12,045,384 rows) has been removed from the repo; only the normalized corpus is published.

	source (removed)	normalized
Rows	12,045,384	11,916,275
Size	28.7 GB JSONL	8.2 GB parquet (zstd)
Columns	label, text	text, label

Rows by source label: blogs 11,631,083 · questions 264,933 · ninisite 20,259.

Rows dropped: 126,591 exact duplicates (after normalization), 1,681 shorter than three words, 837 mostly non-Persian script.

Normalization

Applied to every row by persian_normalizer.py with zwnj_lexicon.json (both in this repo). It is the same normalizer as FarhanFarsi/subkade-persian-subtitles-lm, so the two corpora spell words identically. Steps 8 and 9 below are corpus-specific repairs applied after it.

1. Character unification. NFKC (folds Arabic presentation forms such as ﻻ), then 112 Arabic, Urdu, Kurdish, Pashto and Sindhi letter forms mapped to Persian: ي ى → ی, ك → ک, ة → ه, أ إ ٱ → ا, ؤ → و, ھ ہ → ه, ۀ → ه‌ی, ڤ → ف, ٹ → ت, ۆ → و, ں → ن, and so on. Persian and Arabic-Indic digits are unified.

2. Diacritics, tatweel and bidi marks removed. A word-final damma, used informally for the object marker, becomes و (کارُ → کارو). A detached tatweel suffix is re-attached (پورشم ـه → پورشمه).

3. Numbers spelled out: ۲۵ → بیست و پنج, ۳ام → سوم, 50% → پنجاه درصد, 0047 → صفر صفر چهل و هفت.

4. Bracketed notes removed ([...], (...), {...}, <...>), then all punctuation and Latin or other non-Persian letters removed. The output alphabet is the 32 Persian letters plus ء and ئ, nim-fasele (U+200C) and single spaces.

5. Stretched words collapsed: letter runs are capped at three, then each stretched form is mapped to the more frequent of its one-letter and two-letter collapse: خدااااا → خدا, خیییلی → خیلی, سلااام → سلام, بببین → ببین, مممنون → ممنون. A collapse never produces a single letter or a different word, so interjections stay distinct: هممممم → همم (not هم), اووو → اوو (not او), ششش stays ششش.

6. Nim-fasele: the می/نمی prefix always takes ZWNJ, as do plural های/هایی after a word. ZWNJ is removed where it is invisible (after ا د ذ ر ز ژ و ء آ) so each visible word has a single encoding.

7. Corpus-derived nim-fasele lexicon (zwnj_lexicon.json: 48,347 word fixes, 21,336 two-word fixes), built from how often each spelling (joined / nim-fasele / spaced) occurs in the subtitle corpus:

   • affixes (می، نمی، بی، plural ها، suffixes after silent ه such as ‌ای ‌ام ‌ی) → nim-fasele: میدونم → می‌دونم, کتابها → کتاب‌ها, بچه ها → بچه‌ها, خونه ی → خونه‌ی
   • pronoun suffixes → joined, the standard spelling: اسم‌شون → اسمشون, کمک‌تون → کمکتون
   • other compounds → the most frequent spelling, applied to every variant: هیچ وقت / هیچ‌وقت → هیچوقت, خیلی‌خب → خیلی خب, به‌نظر → به نظر
   • distinct words are left alone: تنها, فرماندهی vs فرمانده‌ی

8. Decimal repair (this corpus only). The source had lost its decimal points before this pass, so 57.2 arrived as two numbers (۵۷ ۲) and was first spelled پنجاه و هفت دو درصد. fix_decimals.py inserts ممیز where two spelled-out numbers stand directly before a unit word (درصد, متر, سانتی متر, کیلوگرم, تومان, میلیون, ...) with no و between them: پنجاه و هفت ممیز دو درصد, صفر ممیز پنج درصد, پنج ممیز چهار سانتی متر. Ascending single digits are left alone because they are an approximation (دو سه متر = two or three metres), as are scale words (دو هزار تومان). 106,995 rows changed.

9. آ restored (this corpus only). The source had flattened آ to ا (ادم, اقای, قران). build_alef_lexicon.py compares every corpus word with ا against the subtitles LM corpus, which kept آ: a spelling with آ is adopted when it is at least twice as frequent there as the flattened form (minimum 5 occurrences), giving alef_lexicon.json with 7,152 words: ادم → آدم, اقای → آقای, قران → قرآن, اب → آب, ایا → آیا, اموزش → آموزش, امریکا → آمریکا, اینده → آینده, درامد → درآمد, می‌امیزند → می‌آمیزند. Words the reference does not know with آ keep their flattened spelling, and genuinely ambiguous forms (اه/آه, اش/آش) are left unchanged. 8,605,978 rows changed (72% of the corpus).

Use the same normalizer on ASR transcripts and hypotheses so ASR and LM spell words identically:

from persian_normalizer import PersianNormalizer
norm = PersianNormalizer("zwnj_lexicon.json")
norm("ميدونم كه ۲۵ تا كتاب ها رو خوندي!")   # -> 'می‌دونم که بیست و پنج تا کتاب‌ها رو خوندی'

Usage

from datasets import load_dataset
ds = load_dataset("FarhanFarsi/persian_corpus", split="train")