NaxiAI is a digital infrastructure project dedicated to preserving and revitalising the Naxi language (纳西语), a Tibeto-Burman language of the Lijiang valleys of Yunnan, China. The ethnic Naxi population is about 300,000–350,000 (UNESCO's Memory of the World nomination gives approximately 300,000; the 2010 census counted 326,000); the number of fluent speakers is much smaller, and this project does not publish its own survey — see how many people speak Naxi for the sources. The Naxi are home to Dongba, the world's last living hieroglyphic script.
纳西语属藏缅语族，约30万人使用，主要分布于云南省丽江。纳西族拥有世界上唯一仍在使用的象形文字——东巴文。
3,800+ Naxi lexical entries with Pinyin, IPA transcription, tone markers, and parallel Chinese/English definitions. Data sourced from Webonary, Pangloss, and original field recordings.
3,800+纳西语词条，含拼音、IPA转写、声调标记及中英文释义。
A growing collection of Naxi–Chinese–English sentence pairs used for training and evaluating translation models. Includes conversational, proverbial, and narrative registers.
纳西语—中文—英文三语平行句对，用于训练和评估翻译模型。
LoRA-adapted language models trained on 60,000+ Naxi examples. Our best model achieves 82.3 NTQS — a 41-point improvement over zero-shot frontier models. Built on Gemma 2 9B, published with full provenance.
基于60,000+纳西语样本训练的LoRA微调模型，最佳模型达82.3 NTQS。
The Dongba script is the world's last living hieroglyphic writing system. We digitise Naxi cultural and linguistic books, OCR them in Chinese, and provide English translations with full-text concordance search.
东巴文是世界上唯一仍在使用的象形文字。我们数字化纳西文化典籍，OCR识别中文，并提供英文翻译与全文检索。
NaxiAI is built and maintained by Andrew Peters, a graduate student in the Department of Anthropology at Harvard University, whose research concerns the Naxi Dongba manuscript tradition and how it is catalogued, read and translated. The corpus here grew out of that work: the dictionary is the reference material the translation pipeline needed, the library is the source material behind the manuscript research, and the gazetteer is the place-name index that reading the manuscripts requires.
Peer-reviewed research behind this project:
Li, Xiaoliang, and Andrew Peters. 2024. “The origin and era of Harvard's collection of ‘double red’ circle Naxi Dongba manuscripts.” Journal of Chinese Writing Systems. doi:10.1177/25138502241300182.
That study traces 598 volumes of Naxi Dongba manuscripts held by the Harvard-Yenching Library to a single village in Yulong County, Lijiang. Two volumes from the same collection are digitised in this library — vol. 1 (2011) and vol. 5 (2018) — and the paper is the citable source for the collection's size and origin. There is a reference page on the collection and its provenance.
The models, benchmarks and evaluation harnesses behind the translation work are published openly under huggingface.co/Apeters247 — see the models page for what each one is.
Contact: anp8875@g.harvard.edu
NaxiAI is sponsored by the Peters Northeastern Foundation, which supports non-profit work in education, culture, and community. Model weights, corpus provenance, and evaluation methods are published openly — not merely described.
What is published here is published on purpose, and what is not is stated rather than implied. Three different things sit on this site:
Citation. NaxiAI. Naxi Digital Dictionary / Naxi Digital Library. https://naxiai.com — dictionary entry pages and place pages print their own citation line, and each library work offers TEI XML for scholarly citation.
For machines. /llms.txt describes the corpus, its caveats and its machine-readable endpoints; /sitemap.xml indexes it.