dhivehi-tiny-latin2thaana-keymap-v1

📄 Paper: Small Enough for the Browser: A 9M-Parameter Dhivehi Transliterator That Matches Its 300M Teacher

A 9M-parameter byte-level encoder–decoder for Latin-script Dhivehi → Thaana transliteration, trained from scratch and shipped as ONNX for the browser. It runs entirely on the device: 35 ms per sentence on a laptop CPU in WebAssembly, from a 9.6 MB download, with no server and no text leaving the page.

On the public news test split it matches the 300M ByT5 fine-tune it was distilled from on character accuracy (0.919 vs 0.919) and beats it on exact match (0.286 vs 0.247), and it passes the same one-word checks. On a held-out set of Dhivehi Latin chat text it leads on word accuracy (0.804 vs 0.797) and trails on character accuracy (0.920 vs 0.955).

Code, demo page and evaluation harness: https://github.com/streetsolider/div-transliteration-web

Quick start

Requirements

Python 3.9+ (only to serve the folder with the right headers; nothing is installed)
A browser with WebAssembly: any current Chrome, Edge, Firefox or Safari, desktop or phone
Optional: Node 18+ for the evaluation harness, a WebGPU-capable browser for the fp16 path

Browser demo

Live at https://streetsolider.github.io/div-transliteration-web/ — or run it yourself:

git clone https://github.com/streetsolider/div-transliteration-web
cd div-transliteration-web
python serve.py            # port 8788; add a number for another port

Visit http://localhost:8788. The page fetches the model from the Hub on first load (9.6 MB) and the runtime caches it in the browser, so after that the only thing still coming from elsewhere is Transformers.js itself, from jsDelivr. Every font is served from this repo. ?model=byt5-tiny-distill loads the copy under models/ instead of the Hub, leaving jsDelivr as the only remote fetch.

serve.py is a plain static server that adds Cache-Control: no-cache and the cross-origin isolation headers (COOP/COEP). Without those headers the runtime falls back to single-threaded WASM, about twice as slow per sentence, so any other static server works but is slower. To try it from a phone, open http://<your-LAN-IP>:8788 (the server binds 0.0.0.0).

Query parameters on the page: ?device=webgpu uses the fp16 model on the GPU, ?dtype=fp16 picks the fp16 files, ?model=<hub id or folder> loads another model.

Node harness

Scores the model on a TSV of latin<TAB>keymap rows through the same chunking as the page, and times it:

npm install
node harness.mjs models/byt5-tiny-distill rows.tsv --dtype q8 --n 500

rows.tsv is a header line latin\tkeymap followed by one pair per line, the target in the Segha keymap (ASCII), not Thaana. No test file ships with the repo; the news test split is on the Hub at alakxender/dhivehi-transliteration-pairs.

Python

The same ONNX files load in Python with optimum (the file names are not optimum's defaults, so they are passed explicitly); the tokenizer is the stock ByT5 one from any ByT5 checkpoint:

pip install "optimum[onnxruntime]" transformers

from optimum.onnxruntime import ORTModelForSeq2SeqLM
from transformers import AutoTokenizer
import json

model = ORTModelForSeq2SeqLM.from_pretrained(
    "str33t/dhivehi-tiny-latin2thaana-keymap-v1", subfolder="onnx",
    encoder_file_name="encoder_model_quantized.onnx",
    decoder_file_name="decoder_model_merged_quantized.onnx",
    use_merged=True, use_cache=False)   # the merged decoder carries its own cache
tok = AutoTokenizer.from_pretrained("google/byt5-small")
keymap = json.load(open("keymap.json", encoding="utf-8"))

ids = tok(["miadhu male ah dhaan jehey"], return_tensors="pt")
out = tok.batch_decode(model.generate(**ids, max_new_tokens=128, num_beams=4), skip_special_tokens=True)
print("".join(keymap.get(c, c) for c in out[0]))   # މިއަދު މާލެއަށް ދާންޖެހޭ

Results (beam 4)

System	params	news test: exact / char acc	chat test: char / word acc	one-word inputs	per sentence
this model (stage 3, distilled)	9M	0.286 / 0.919	0.920 / 0.804	4/4	35 ms, browser CPU
this architecture, stage 1 (news only)	9M	0.302 / 0.925	0.884 / 0.784	1/4	40 ms, browser CPU
this architecture, stage 2 (+augmentation)	9M	0.282 / 0.912	0.880 / 0.768	4/4	43 ms, browser CPU
str33t/dhivehi-byt5-latin2thaana-keymap-v1 (300M, served)	300M	0.247 / 0.919	0.955 / 0.797	4/4	356 ms, GPU

News test split: the 37,582-row test split of alakxender/dhivehi-transliteration-pairs. Chat test: 102 human-verified sentences from the Dhivehi Latin chat text corpus described under Training. One-word inputs: the four short inputs the 300M fine-tune exists to fix (bohkuraa, kuru, karu, bas).

Exact match has a low ceiling on this test split: many reference rows are a different headline from their source rather than a transliteration of it, so character accuracy is the number to trust.

Why a 9M model

The served model is a fine-tune of a 300M ByT5-small: 1.2 GB of weights, a GPU, a server round trip. Transliteration is a character-level mapping with short inputs, and the grapheme-to-phoneme literature had already shown a ByT5 with d_model 256 serving a hundred languages in 7–20M parameters. The bet was that a model that small, trained from scratch on the task alone, would match a general model fine-tuned for it, and be fast enough for a phone. It did, with one honest caveat: on Dhivehi Latin chat text it needed the 300M model as a teacher to get there.

Architecture

T5 architecture, ByT5-style byte tokenizer: input id = UTF-8 byte + 3, no vocabulary file.
6 encoder + 2 decoder layers, d_model 256, d_ff 1024, 4 heads, tied embeddings: 9.0M parameters.
Output is the Segha phonetic keymap, one ASCII byte per Thaana codepoint, converted to Thaana by a 55-entry table after decoding. Thaana codepoints are three UTF-8 bytes each, so emitting the keymap costs a third of the decoder steps. The table is in keymap.json; the layout is jawish/jtk's phonetic map.
Exported with optimum to ONNX: encoder 6.6 MB and decoder 3.0 MB at int8 for WASM, fp16 variants for WebGPU. The merged decoder keeps its matrix multiplies inside If-subgraphs, which the dynamic quantizer skips, so the two decoder halves are quantized first and merged afterwards.

Training

Three stages, each checked on the news test split, the chat test set and the four one-word inputs before the next was allowed:

News only. 150,326 pairs from alakxender/dhivehi-transliteration-pairs, from scratch: lr 1e-3, 5% warmup, linear decay, batch 64, 20 epochs, bf16, 49 minutes on one RTX 5070 Ti. Already ahead of the 300M model on the news split, but one- to three-word inputs duplicate the word (bas → basq, basq): the corpus has no short rows.
Augmentation. The same length-diverse augmentations the 300M fine-tune uses (single words, short phrases, long concatenations, place names with chat misspellings, English loanwords), 173,353 pairs. Fixes the short inputs, costs a little on the news split.
Distillation. 73,265 sentence pairs produced by the 300M model over a Dhivehi Latin chat text corpus (Dhivehi typed in Latin letters in messages; not redistributed), with its known systematic errors corrected, added to the stage 2 corpus: 246,618 pairs, 20 epochs, 1 h 53 min. This is what closes the chat-text gap. A warm-started variant, continuing from stage 2, scored about a point lower than training from scratch on the full corpus.

Before release the model was probed for memorisation of that corpus: free-running generations from short prompts, and continuations from two-word prefixes of corpus sentences, were checked against every word 4-gram and character 15-gram unique to it. No corpus text was emitted and no continuation matched.

Usage

In the browser or Node with Transformers.js 3.x. There is no tokenizer.json: ByT5 tokenization is bytes + 3, done in a few lines.

import { env, Tensor, AutoModelForSeq2SeqLM } from '@huggingface/transformers';

const model = await AutoModelForSeq2SeqLM.from_pretrained('str33t/dhivehi-tiny-latin2thaana-keymap-v1', { dtype: 'q8' });
const keymap = await (await fetch('keymap.json')).json();   // ASCII -> Thaana table, in this repo

const enc = new TextEncoder(), dec = new TextDecoder();
function encode(texts) {                                    // ByT5: id = byte + 3, eos = 1, pad = 0
  const rows = texts.map((t) => [...Array.from(enc.encode(t)).map((b) => b + 3), 1]);
  const L = Math.max(...rows.map((r) => r.length));
  const ids = new BigInt64Array(rows.length * L), mask = new BigInt64Array(rows.length * L);
  rows.forEach((r, i) => r.forEach((v, j) => { ids[i * L + j] = BigInt(v); mask[i * L + j] = 1n; }));
  return { input_ids: new Tensor('int64', ids, [rows.length, L]), attention_mask: new Tensor('int64', mask, [rows.length, L]) };
}
function decode(out) {                                      // bytes back to text, stop at eos
  const [n, L] = out.dims, res = [];
  for (let i = 0; i < n; i++) { const b = []; for (let j = 0; j < L; j++) { const v = Number(out.data[i * L + j]); if (v === 1) break; if (v >= 3) b.push(v - 3); } res.push(dec.decode(Uint8Array.from(b))); }
  return res;
}

const out = await model.generate({ ...encode(['miadhu male ah dhaan jehey']), max_new_tokens: 128 });
const thaana = decode(out)[0].split('').map((c) => keymap[c] ?? c).join('');
console.log(thaana);   // މިއަދު މާލެއަށް ދާންޖެހޭ

For inputs longer than about ten words, split as the demo page does (paragraphs → sentences → phrases → 10-word chunks, one batched generate per paragraph, then stitch); pipeline.js is that logic and index.html is a working page. See Quick start for running the page, the Node harness and the Python loader.

Intended use and limitations

Latin-script Dhivehi as people type it in messages and search boxes, to Thaana. Trained on news and on a Dhivehi Latin chat text corpus; evaluated on both. Expect weaker output on rare proper names and on registers far from either. Single training seed. The decoder has no length limit of its own; the demo caps generation at 2.5× the input bytes to stop the rare runaway on an unseen name.

License

MIT, for the code and the model files alike. See LICENSE.

The bundled fonts keep their own terms: IBM Plex is under the SIL Open Font License 1.1 (fonts/IBM-Plex-OFL.txt). Faruma and Mv MAG Round XBold are long-standing free Thaana fonts redistributed without a licence file of their own; they are listed as free for personal and commercial use, but that is not a licence this repo can vouch for.

Downloads last month: -

str33t
/

dhivehi-tiny-latin2thaana-keymap-v1