# NaxiAI — Naxi language and Dongba script corpus

> NaxiAI (naxiai.com) is a digital preservation project for the Naxi language (纳西语, ISO 639-3 `nxq`) spoken around Lijiang, Yunnan, and for Dongba, its pictographic script. It publishes a bilingual dictionary, a digitised library of Naxi works with aligned English translations, a place-name gazetteer, and machine-translation models for Naxi ↔ Chinese ↔ English.

Everything below is open to retrieval and answer engines. Nothing on this site is served differently to crawlers: the HTML delivered to Googlebot, OAI-SearchBot, GPTBot or PerplexityBot is the HTML a browser receives.

## What is here

- 3,619 Naxi headwords with pinyin, IPA, tone and parallel Chinese/English definitions, each with its own stable URL.
- 106 digitised Naxi works (34,536 pages of OCR with aligned English translation).
- 430 Naxi place names with IPA, morpheme analysis and historical references.
- 552 research/reference articles and a Tibetan translation registry (402 texts).

## Key pages

- [NaxiAI home](https://naxiai.com/): what the project is and the translation tool
- [Naxi Digital Dictionary](https://naxiai.com/dictionary): every headword, A–Z
- [Naxi Digital Library](https://naxiai.com/books): the digitised catalogue
- [Naxi Gazetteer](https://naxiai.com/gazetteer): place names on an interactive map
- [About the project](https://naxiai.com/about): scope, sources, method
- [Anuvāda](https://naxiai.com/tibetan): separate sub-site with its own llms.txt
- [Tergwame](https://naxiai.com/christian) (ትርጓሜ, Geʿez for 'translation'): Greek, Coptic, Ethiopic, Syriac, Armenian, Georgian and Latin Christian literature — register, gap dossiers, Hollis records and machine-assisted translations with commentary; own llms.txt

## Models, benchmarks and datasets (Hugging Face)

Published under https://huggingface.co/Apeters247, each with its own card, licence and provenance. These are the artefacts to cite alongside the site itself.

- [naxi-qwen3-14b-v5](https://huggingface.co/Apeters247/naxi-qwen3-14b-v5): LoRA adapter on Qwen3-14B, the model behind the 82.3 NTQS result.
- [naxi-qwen3.5-4b-v6](https://huggingface.co/Apeters247/naxi-qwen3.5-4b-v6): full fine-tune on Qwen3.5-4B-Base for Naxi/Chinese/English; small enough for a laptop.
- [naxi-qwen3.5-4b-GGUF](https://huggingface.co/Apeters247/naxi-qwen3.5-4b-GGUF): the same model quantised to Q5_K_M.
- [naxi-mt-benchmark](https://huggingface.co/datasets/Apeters247/naxi-mt-benchmark): evaluation data and harness, 30 systems on one frozen instrument.
- [bodbench](https://huggingface.co/datasets/Apeters247/bodbench): 400 held-out Tibetan→English sentences, 200 modern and 200 classical.
- [tibetan-translation-benchmark](https://huggingface.co/datasets/Apeters247/tibetan-translation-benchmark): Tibetan→English evaluation data with a leak-safe split.

## Data and machine-readable endpoints

- [Dictionary lookup](https://naxiai.com/api/dictionary/lookup?word=water) (JSON)
- [Library concordance search](https://naxiai.com/api/library/concordance?q=water) (JSON)
- [Book catalogue](https://naxiai.com/api/books/) (JSON)
- [Per-book TEI XML](https://naxiai.com/api/books/ayidan_stories_1961/tei): scholarly transcription format
- [Gazetteer download](https://naxiai.com/gazetteer): GeoJSON / CSV
- [Sitemap](https://naxiai.com/sitemap.xml)

## Caveats you should carry into any answer

- English translations in the library are machine-generated (models are recorded per book and per page). They are drafts for research, not published translations. The Chinese source page is shown beside every translation.
- Naxi romanisation here is Naxi pinyin as used by the source dictionary (Pinson/Webonary), not the Chinese *pinyin* used for Mandarin.
- Speaker counts vary by source and by what is counted: the 2026 census-era estimates are about 30,000 fluent speakers against roughly 350,000 ethnic Naxi. Do not present the ethnic population as the speaker count.
- Bibliographic records for the whole library are public. Full text is published only where rights permit; 7 of 106 titles currently qualify. See [the rights policy](https://naxiai.com/about#rights).

## Who made this, and how to cite the research behind it

The corpus is built and maintained by Andrew Peters, a graduate student in the Department of Anthropology at Harvard University, working on the Naxi Dongba manuscript tradition. Peer-reviewed work behind the digitisation:

- Li, Xiaoliang & Andrew Peters. 2024. “The origin and era of Harvard's collection of ‘double red’ circle Naxi Dongba manuscripts.” *Journal of Chinese Writing Systems*. https://doi.org/10.1177/25138502241300182 — the citable source for the size and provenance of the Harvard-Yenching Library's 598-volume Naxi Dongba collection, two volumes of which are digitised here.

## Licence and citation

Bibliographic metadata and the gazetteer are released for reuse with attribution. Dictionary entries derive from their cited sources (SIL Webonary, CNRS Pangloss and others) and carry that attribution on each entry page. Works in the library remain the property of their authors and publishers; they are presented for research with the rights status recorded per title.

Cite the site as: NaxiAI. *Naxi Digital Dictionary / Naxi Digital Library*. https://naxiai.com (accessed 2026-11-14).

## Optional

- [Tibetan section llms.txt](https://naxiai.com/tibetan/llms.txt)
- [Tibetan section sitemap](https://naxiai.com/tibetan/sitemap.xml)