Lordkiki/dialingua-bkv2eng-web

Dialingua — Bekwarra → English (browser build)

Quantised ONNX build of a Marian fine-tune for Bekwarra (ISO 639-3 bkv), a Bendi language of Cross River State, Nigeria. Runs entirely in the browser via Transformers.js — no server, no API key.

Base: Helsinki-NLP/opus-mt-mul-en (Apache 2.0).

Size

fp32 export	926 MB
int8 quantised	233 MB

Quantisation targets the unmerged decoders. The merged decoder wraps its whole body in a single ONNX If node, and ONNX Runtime's quantiser does not descend into subgraphs — on that graph it silently compresses nothing.

Honest scope

Trained on 1,095 verse pairs OCR'd from printed Old Testament portions — below the ~3,000 where this stops being a demo.

• chrF++ 28.33 on a held-out book (Ruth).
• Scripture and formal register: works.
• Everyday speech: unreliable. It has never seen conversational Bekwarra.

For comparison, a 6× larger M2M-100 fine-tune on the same data scored 19.35. Smaller was better here, because 418M parameters against 1,095 examples is mostly unused capacity.

Do not rely on this for anything consequential without a native speaker checking the output.

Usage

import {{ pipeline }} from '@huggingface/transformers';

const translate = await pipeline('translation', 'Lordkiki/dialingua-bkv2eng-web');
const out = await translate("Ibang h'uni ang'a de achi i na k'ufo uni irichom?");
console.log(out[0].translation_text);

Orthography

Bekwarra's apostrophe is phonemic — a letter, not punctuation (k'uchu, ng'amin, r'abe). Both the closed and spaced (K' iye) forms occur in current writing; both survive this tokenizer, verified.

Training data

Old Testament portions in Bekwarra, OCR'd from scanned print. The underlying translations are © their rights holders; released here for research and language preservation. Settle permissions before commercial use.