Burmese · corpus in collection

မြန်မာဘာသာ

Burmese · 43M speakers · fine-tuning on a community-curated corpus.

I COULD BE BOUNDED IN A NUTSHELL AND COUNT MYSELF A KING OF INFINITE SPACE. [ 2.2.254-255 ]မြန်မာဘာသာ · Bamar & diaspora

The same thought, double the cost.

ENGLISH · EN-US
The river remembers what the map forgets.
words
8
tokens
8
tok/word
1.0
BURMESE · BURMESE
မြေပုံက မေ့လျော့ထားတာတွေကို မြစ်က မှတ်မိနေတယ်
words
4
tokens
17
tok/word
4.3
1.0 tok/wordvs4.3 tok/word· 2.1× the cost to think in your own language

Same thought · 8 tokens in English, 17 in Burmese. Over 4× the cost to think in your own language.

A script that predates the printing press

Burmese has been written continuously for over a thousand years palm-leaf manuscripts, parabaik folding books, monastic libraries whose catalogs pre-date the Domesday. The characters are round because a straight stroke would split a palm leaf; the whole visual grammar carries the memory of its substrate.

In modern LLM vocabularies, this thousand-year record is compressed into an average of 4.2 tokens per word. English gets 1.1. The gap is not linguistic it is bookkeeping.

Zawgyi, and the cost of an encoding war

Until roughly 2019, most Burmese text on the internet was written in Zawgyi a non-standard encoding that looked correct on screen but was gibberish to anything that spoke Unicode. A decade of blog posts, comment threads, and news archives sits in this format. To the tokenizer, they are two different languages.

Normalization is not a preprocessing step. It is a scribal act an emendation, made once, then attested in the colophon.

Fine-tuning, not retraining.

We do not retrain from scratch. We do not need to. We start from an existing pretrained open model Gemma-4, among the candidate bases under community review, its weights frozen and add a small LoRA adapter shaped by community judgment: which texts matter, which spellings are contested, which voices the crawl missed.

The base model is given. What it's add is the language, in its own script, in the community's own hand.

section numbers · toggle blueprint for configs
[00] · plate

We work with open source models. Gemma-4 (and SEA-LION )is being used for Burmese, Karen, and Jinghpaw as well as; Llama-3.1-8B may be used for Rohingya, for byte-level fallback where Hanifi is unrepresented.

Curation
[01] · plate

Community participation and consent for curation and capture. Provenance recorded and fair compensation provided.

Normalization
[02] · plate

Converting legacy encodings and non-standard scripts into universal digital formats (e.g., Zawgyi to Unicode) is an act of cultural and historic preservation. Working alongside local linguists and community members, we systematically standardize text to be machine-readable while ensuring cultural context, orthographic nuances, and historical integrity remain intact.

[03] · plate

Adapting foundational base models to local linguistic realities through targeted, data-efficient techniques calibrated to each language’s structure and context. Training logs and loss curves are published directly to the public archive with complete technical transparency.

[04] · plate

Community-annotated benchmarks. Evals and benchmarks designed with local community and experts. Annotators acknowledgements and compensation. Inter-annotator agreement reported alongside every score.

THE CONTESTED QUESTIONS · KEPT VISIBLE, NOT RESOLVED FOR THE COMMUNITY
  • Q01Which pre-1990 orthographic conventions to preserve alongside the standard.
  • Q02How to weight monastic Pali-inflected registers against contemporary speech.
[ 06 ] THE VOICES
named with consent · faces omitted by choice
Daw Khin Ma Ma
named
monastic librarian, Mandalay
duration · 1:24
Ko Zaw Min
named
journalist, diaspora
duration · 2:11
anonymous
anonymous
poet
duration · 0:58
Ma Thida
named
novelist
duration · 3:02
[ 06 ] SOVEREIGNTY

The corpus belongs to the communities that made it. Consent is per item and revocable. Revocation is honored in the weights, not just in the archive.

Model weights reflect the terms and governance frameworks chosen with the community.

[ 08 ] COLOPHON

Completed at Mae Sot, in 2026, on the Burmese project, rank 32. Released with community consent; evaluation held.

place
Mae Sot
date
2026.07.18
hardware
1× NVIDIA H100 80GB
license
per base terms · corpus CC-BY-SA-4.0
elsinore data co · undertext · v0.5 · 2026.07
next projectS'gaw Karen
contribute