DibaAi/DiLLM.nano

DiLLM Phase 1 — Persian Corpora

The training corpora for DiLLM (دیبا), a Persian-first language model built in Iran by Dibachain.

Publisher	Dibachain — dibachain.ir
Country	Iran 🇮🇷
Model	DibaAi/DiLLM.nano
Source code	github.com/AliAkrami1375/DibaLLM
Language	Persian (فارسی)
Total	~1.02M documents · ~379M tokens

This dataset was assembled from scratch for DiLLM. Every corpus was downloaded from its primary source, filtered, and cleaned by the project's own pipeline — nothing here is a repackaged third-party dataset. The build is fully reproducible: scripts/download_data.py fetches the raw sources and scripts/build_corpus.py produces exactly these files.

Why this dataset exists

Persian is under-served by multilingual corpora. Web-scraped Persian data is noisy, and general-purpose tokenizers split Persian words into three or more tokens. DiLLM starts from the opposite end: a corpus assembled for Persian, with the domains that matter to an Iranian audience weighted deliberately — classical literature, Iranian history, Persian grammar, and Twelver Shia heritage.

Configurations

Config	Documents	Tokens (est.)	Contents
wiki_general	860,495	344.5M	Persian Wikipedia — the general language base
poetry	110,179	21.0M	Classical Persian poetry from 195 poets
history_iran	5,835	7.0M	Contemporary Iranian history, Qajar era to present
grammar	41,922	3.5M	Persian lexicon and language articles
religion_shia	3,717	3.3M	Quran with Persian translations, Twelver Shia heritage
identity_sft	91	—	Supervised pairs defining the model's identity and behavior

poetry — classical Persian literature

Complete Shahnameh (49,609 verses by Ferdowsi), plus Hafez, Saadi, Rumi, Khayyam, Nezami, Attar, Jami, Sa'eb Tabrizi and 187 other poets. Sourced from Ganjoor, the reference digital archive of Persian poetry. ZWNJ (half-space) is preserved throughout — critical for correct Persian orthography.

grammar — everything about the Persian language

• ~40,000 lexicon entries from Persian Wiktionary, each with part of speech (اسم / صفت / فعل / قید / مصدر), etymology (ریشه‌شناسی), pronunciation (آوایش), synonyms and derived forms.
• 1,405 Wikipedia articles on grammar, syntax, morphology, phonology, orthography, dialects, prosody (عروض) and rhetoric — selected by real MediaWiki category membership, not title matching.
• 45 Wikibooks chapters on teaching Persian.

religion_shia — Twelver Shia heritage

The complete Quran in Arabic (Uthmani script) with three Persian translations (Makarem Shirazi, Fooladvand, Ansarian), articles on the Fourteen Infallibles, Mahdaviat, Nahj al-Balagha, Ashura and Karbala, Shia theology and jurisprudence, plus religious texts from Persian Wikisource.

identity_sft — behavioral fine-tuning pairs

Supervised pairs that teach the model who it is and how to behave, in five categories: identity, boundary (saying "I don't know"), phase, citation (quote-then-interpret), and initiative.

Schema

Corpus configs:

{"source": str,   # origin: fawiki, ganjoor, tanzil, fawiktionary, ...
 "title":  str,   # article title, poem title, or sura reference
 "text":   str}   # the document body

identity_sft:

{"instruction": str, "response": str, "category": str}

Usage

from datasets import load_dataset

poetry = load_dataset("DibaAi/DiLLM.nano", "poetry", split="train")
print(poetry[0]["title"])   # فردوسی — بخش ۱

Sources and licensing

Corpus	Primary source	Source license
wiki_general, history_iran	Persian Wikipedia (2025-07 dump)	CC BY-SA 4.0
grammar	Persian Wiktionary, Wikipedia, Wikibooks	CC BY-SA 4.0
poetry	Ganjoor	Public domain (classical works)
religion_shia	Tanzil, Persian Wikisource, Wikipedia	Tanzil terms; CC BY-SA 4.0
identity_sft	Written for this project	MIT

This dataset is released under CC BY-SA 4.0, inherited from its Wikimedia components. Attribute Wikimedia projects and Ganjoor when redistributing. Quran text follows Tanzil's terms of use: the text must not be modified and must be attributed to Tanzil.

Processing pipeline

1. Download — each source fetched from its primary origin.
2. Extract — MediaWiki dumps parsed; templates, refs, category and file links stripped; HTML entities decoded.
3. Filter — documents under 250 characters and disambiguation pages dropped (189,373 removed). Lexicon entries use a 40-character threshold.
4. Classify — domain corpora selected by MediaWiki category membership with an explicit denylist for categories that match a topic rather than the language (loanword categories were pulling in articles about chromium and opera, for example).
5. Emit — one JSON object per document, converted to zstd-compressed Parquet.

Known limitations

• Wikipedia dominance. 91% of tokens come from Persian Wikipedia, carrying its coverage gaps and editorial perspective.
• No deduplication yet. Document-level and paragraph-level dedup is planned.
• No tafsir. The Quran appears with translation but without exegesis; no open Persian tafsir corpus was available at build time.
• Perspective. history_iran and religion_shia reflect their Persian Wikipedia and Iranian sources. Treat them as documents of a viewpoint.
• Not safety-filtered. The corpus has not been screened for toxic content beyond the quality filters described above.

Citation

@dataset{dillm_corpora_2026,
  title     = {DiLLM Phase 1: Persian Corpora},
  author    = {Dibachain},
  year      = {2026},
  url       = {https://huggingface.co/datasets/DibaAi/DiLLM.nano},
  publisher = {Dibachain, Iran}
}

---

Built in Iran by Dibachain.