Low-resource languages

Low-resource language translation — a fine-tune, or examples in the prompt.

Low-resource language machine translation is for a language that does not have a large parallel corpus. Teams compare a fine-tuned model such as Meta’s NLLB-200, which Scripture Forge trains per project, with few-shot prompting of a frontier model, which is how Aquilla drafts. In Aquilla’s experience that prompting path is the best way we have found to use AI here, and we do not claim it wins a BLEU or chrF score.

The direct answer

What is the best AI for low-resource language translation?

Best depends on the job. NLLB-200 is a translation model trained to cover 200 languages, and Scripture Forge fine-tunes it when a project has completed Scripture to learn from. A frontier language model can draft from a few approved examples without that training step. Aquilla takes the second path. The scores below are other people’s results on frozen test sets, not a claim that one product outranks the other in a translation office.

Meta’s NLLB-200 announcement describes languages without extensive datasets, and reports that one model translates 200 languages, including 55 African languages. Meta’s own figure is a 44 percent average BLEU gain over the previous state of the art on FLORES-101. The 600-million-parameter model card calls NLLB-200 a research model for single sentences in the general domain, not a production system and not a model for domain-specific text.

Robinson, Ogayo, Mortensen, and Neubig (WMT 2023) compared ChatGPT with traditional MT on FLORES-200. NLLB scored higher chrF on 169 of 201 languages (84.1 percent), by 11.9 chrF on average against the better few-shot ChatGPT setting. In the five-shot setting, ChatGPT beat NLLB on 47 percent of the high-resource languages in NLLB’s labels and on 6 percent of the low-resource languages. A frozen test set in 2023 is not the last word. In October 2026, SIL’s Scripture Forge team wrote that newer models had begun to rival and even surpass NLLB for some pairs, including very low data and unusual scripts, and that they were collecting team feedback before exposing those processes in the product. That note is on the SIL community forum.

What the search is actually choosing

Best AI models for low-resource language translation are a job list.

Searchers want a model that can draft when the language is scarce online. The practical choice is who trains, who waits, and who holds the text.

JobFine-tuned NLLBAquilla
Translate a language with little paired textYes — NLLB-200 was built for that coverage.Yes — the draft is prompted, not trained from scratch.
Train a new model before the first draftYes — Scripture Forge fine-tunes NLLB on the project.No — Aquilla does not train a model for the language.
A GPU on the translator's computerNo for Scripture Forge — SIL runs fine-tuning on its servers.No — drafting is in the browser.
Start a Bible draft before a Gospel is finishedNot the New Testament path — Scripture Forge asks for one finished Synoptic Gospel first.Yes — no finished testament is required.
Use an approved rendering on the next draftAfter a training cycle — the next book, or a requested chapter batch.Yes — Living Memory keeps decisions you authorize.
Keep project text off a model providerScripture Forge fine-tunes and runs NLLB on SIL servers.No — a built-in draft sends the prompt to the model provider.
A published BLEU or chrF win over the otherOften ahead on frozen test sets. That is not an Aquilla score.Not claimed. See the benchmark section.

Two ways to use the little data a project has

Fine-tune a translation model

SIL’s Serval API describes translation into a new minority language by fine-tuning NLLB, and back translation into more than 200 languages from the same model. Scripture Forge is one of the interfaces on that API. The model learns by comparing a reference text with the project’s completed translation, then drafts from a source text. The NLLB-200 vs LLM page follows the data, the GPU, and the retraining cycle.

Prompt a frontier model

Few-shot prompting puts examples in the request and does not update the model’s weights. Moslem, Haque, Kelleher, and Way (EAMT 2023) showed a model adapting at translation time to similar approved pairs and to terminology. Aquilla drafts this way: project context, including translations the team has approved, is part of the prompt. Living Memory keeps those decisions for later drafts. Aquilla does not train a checkpoint per language.

Why Aquilla uses prompting

In Aquilla’s experience, in-context learning is the best way we have found to leverage AI for low-resource language translation. A new language does not wait on a training job. The translator does not provision a GPU. A correction the team authorizes can inform the next draft without a retraining cycle. A newer frontier model can be prompted without fitting a new checkpoint for every language. That is our experience from projects. It is not a measured ranking against NLLB.

What neither approach finishes

A draft is not a published translation. Meta’s model card says NLLB output cannot be a certified translation. Aquilla’s agents propose, and a person approves meaning and release. Bible projects still need the team, community, and consultant steps on the workflow page. Aquilla does not assign that consultant.

Scores and the work

A test-set score does not see the approval loop.

BLEU and chrF compare a frozen system’s output with a reference, once. The value of a prompted draft, in Aquilla’s experience, is that a translator approves or corrects and the next draft can use that decision. A static score on a model that never sees those approvals does not measure that loop. That limit is why this page refuses a winner.

The studies that do exist are still worth reading, and they often favor a specialized or fine-tuned NLLB system on the set they chose. Scalvini, Debess, and Simonsen (NoDaLiDa 2025) fine-tuned NLLB for English-to-Faroese. On FLORES-200, their best fine-tuned NLLB 1.3B model ranked second overall and first among open models; Claude 3.5 ranked first. Fine-tuned NLLB on the Sprotin corpus alone beat GPT-4o in their automatic scores. They also note that FLORES was translated from English and is a narrow sample of Faroese life. Kumar, Jyothi, and Bhattacharyya find that BLEU and chrF++ diverge on extremely low-resource languages, and that one metric is not enough. Chen and colleagues find that a large part of the spread in extremely low-resource scores tracks train-test overlap and pretraining exposure, not only the method.

In-context research is closer to the loop, and it is still a one-pass score. A 2025 NAACL Findings paper reports that choosing similar examples can improve low-resource translation. Lu and colleagues (AACL 2026) find that more in-context examples help English-into-low-resource translation, that retrieved examples need fewer shots, and that prompting still adds something on top of fine-tuning. None of those papers scores a team correcting one sentence and feeding that approval into the next. Castilho and Knowles describe the related gap for context: sentence-level BLEU is weakly sensitive to improvements that touch only a few words, such as a pronoun that depends on an earlier sentence.

Which page to use

Stay here for the choice. Use the NLLB page for the training path.

This page is the low-resource language question: fine-tune, or prompt. The NLLB-200 vs LLM page is how Scripture Forge fine-tunes, what data and servers that takes, and how few-shot prompting differs. AI Bible translation is the Scripture product. Expert-in-the-loop is where a person’s judgment sits.

Common questions

Low-resource language translation FAQ

What is low-resource language machine translation?

It is machine translation for a language that does not have a large set of paired sentences to train on. Meta describes these as languages without extensive datasets, and built NLLB-200 to translate among 200 of them. A fine-tuned NLLB model and a prompted frontier model are two different ways to draft. Neither replaces a person who knows the language.

What is the best translation model for low-resource languages?

There is no single published winner. Meta's NLLB-200 is the model most coverage comparisons start from, and a 2023 study found it ahead of ChatGPT on most low-resource languages in the FLORES-200 test set. Scripture Forge fine-tunes NLLB per project. In Aquilla's experience, few-shot prompting with the team's own examples is the best way we have found to use AI for this work. That is our experience, not a BLEU or chrF result, and Aquilla does not claim to beat NLLB on a benchmark.

What are the best AI models for low-resource language translation?

Published comparisons usually include NLLB-200, sometimes fine-tuned, and frontier or open language models used zero-shot or few-shot. Which one scores higher depends on the language, the test set, and whether the model was fine-tuned. Aquilla does not host a menu of those models. It drafts by prompting a frontier model with project context. The NLLB comparison page goes through the training path Scripture Forge uses.

Does Aquilla train a translation model for my language?

No. Aquilla does not fine-tune NLLB or any other model on your language, and it does not ask you to supply a GPU or a finished corpus before drafting. Approved decisions stay in Living Memory and can inform a later draft. If you need a per-language checkpoint trained on your completed books, that is the Scripture Forge path, not this one.

Does a correction change the next draft without retraining?

In Aquilla, a decision you authorize stays in the project and Living Memory can apply it to later drafts. There is no separate training job between the approval and the next draft. Scripture Forge's documented cycle is different: a completed book is added to the training set for the next book, and for the Gospels and Acts a team can ask the Engineer-in-the-Loop team for a new draft after finishing about 8 to 10 chapters.

Where does my text go when Aquilla drafts?

When you use a built-in AI feature, the relevant prompt and project context are sent to the model provider selected for that feature. Aquilla uses business API access intended not to train generally available models on customer content. Scripture Forge's NLLB fine-tuning and inference stay on SIL servers, which SIL cites for data security and cost. Those are different tradeoffs. Read the privacy page before you choose.

Does Aquilla beat NLLB on BLEU or chrF?

No. Aquilla has not published that comparison, and this page does not invent one. Static scores on a frozen model miss the approval loop these projects actually run. Use them to understand a checkpoint. Do not read them as a ranking of Aquilla against Scripture Forge.

How much does low-resource language translation cost in Aquilla?

Anyone can start on the Free plan with a weekly allowance of AI assistance. Bible-translation teams and ETEN affiliates get team-plan access at no cost — contact us to activate your organization’s workspace.

Sources

Related topics

Draft from the work you have already approved.

Open a Free plan and review what the agent proposes. Bible-translation teams and ETEN affiliates can contact us for team-plan access at no cost. Aquilla will not train a model for your language.