EnSiTa - A Trilingual Multi-Domain Parallel Dataset and Benchmark for Domain-Specific Machine Translation

# A Trilingual Multi-Domain Parallel Dataset and Benchmark for Domain-Specific Machine Translation
Source: [https://arxiv.org/html/2609.29511](https://arxiv.org/html/2609.29511)
Surangika Ranathunga⋄, Nisansa de Silva†, Aloka Fernando†, Kavindu Warnakulasuriya‡Isuru Wijesiri⋆,Menan Velayuthan,Charitha Rathnayaka,Thivaharan Varatharajan,Affiliation:Utrecht UniversityAffiliation:Rowan UniversityEmail:[NisansaDdS@cse\.mrt\.ac\.lk](mailto:)Sajeevi SilvaPiumi KandanaarachchiAffiliation:District General Hospital, HambantotaUthayasanker Thayasivam†⋄Massey University†University of Moratuwa‡National University of Singapore⋆WSO2

###### Abstract

Machine Translation \(MT\) for low\-resource languages remains far behind that of high\-resource languages, and the gap is widest in specialised domains, where parallel data is scarce or entirely absent\. We presentEnSiTa, a trilingual multi\-domain parallel dataset and benchmark for English, Sinhala and Tamil\.EnSiTaprovides human post\-edited training data for seven domains, plus manually translated test sets for those and one additional domain, all produced by professional translators under a multi\-year, rigorously quality\-controlled process\. Using this dataset, we conduct an extensive study of domain\-specific MT for all six language directions, fine\-tuning a from\-scratch Transformer, a pre\-trained translation model \(NLLB\-600M\), and decoder\-only LLMs \(Gemma 3 family, 1B–12B, and TranslateGemma\) across training\-data sizes, model scales, and in\-domain, cross\-domain, multilingual and multi\-domain settings\. To the best of our knowledge, this is the most extensive systematically documented multi\-domain parallel data creation and benchmarking effort for low\-resource MT\. Our data and models will be publicly released\.

## 1Introduction

Solomonic ability shown by Large Language Models \(LLMs\) in solving language\-related problems in the context of high\-resource languages is not replicated when it comes to Machine Translation \(MT\) in low\-resource languages \(LRLs\)[Pang et al\. \(2025\)](https://arxiv.org/html/2609.29511#bib.bib1)\. Although LLMs, just like their predecessors[Lample et al\. \(2018\)](https://arxiv.org/html/2609.29511#bib.bib2), exhibit emergent cross\-lingual capabilities[Wang et al\. \(2024\)](https://arxiv.org/html/2609.29511#bib.bib3), they are still constrained by limited representation of LRLs in pre\-training corpora and insufficient exposure to parallel data\. This challenge is exacerbated in domain\-specific contexts, where parallel corpora for LRLs are often sparse or entirely absent[Ranathunga et al\. \(2024b\)](https://arxiv.org/html/2609.29511#bib.bib4)\. This highlights the need for continuous investing on parallel dataset creation efforts targetting LRLs\.

We present a parallel dataset that contains train/test data across seven domains \(movie subtitles, Mathematics, Health, News, History \(Wikipedia\), Literature and open\-domain\), plus test data for one additional domain \(LLM Jailbreak\) for three language pairs: English\-Tamil, English\-Sinhala and Sinhala\-Tamil\. Parallel corpora, obtained from noisy web\-mined corpora or by translation via MT systems, were manually post\-edited by professional translators to compile a training dataset of 200k\+ sentences pairs, across the three language pairs\. All the test data \(10k\+ sentences pairs\) was manually created by the same translators\.

Using these datasets along with existing train/test data for two other domains \(Government and Bible\), as well as test data from FLORES\+[NLLB Team et al\. \(2024\)](https://arxiv.org/html/2609.29511#bib.bib5), we carry out an extensive set of experiments on domain\-specific NMT across 11 test sets, for the three language pairs\. Specifically, we evaluate the performance of a vanilla seq\-seq translation model using Fairseq toolkit[Ott et al\. \(2019\)](https://arxiv.org/html/2609.29511#bib.bib6), a pre\-trained encoder\-decoder translation model \(NLLB\-600\)[NLLB Team et al\. \(2022\)](https://arxiv.org/html/2609.29511#bib.bib7)and an LLM \(Gemma 3\-1B instruct\)[Gemma Team et al\. \(2025\)](https://arxiv.org/html/2609.29511#bib.bib8)\. We further investigate how training data size differences, as well as model scaling, impact NMT results\. We quantify the impact of divergence across domains when using existing data to build NMT systems for unseen domains, as well as the benefit of multi\-domain, multilingual, as well as multi\-domain \+ multilingual fine\-tuning\.

To the best of our knowledge, this is the first study in LRL\-MT to release per\-domain training data and manually create test sets across this many domains at scale, together with systematic in\-domain, cross\-domain, multilingual and multi\-domain experiments\.Our data and models will be publicly released\.

## 2Related Work

### 2\.1MT Datasets for LRLs

There have been many NMT dataset creation efforts for LRLs in the past\. These efforts can be broadly categorised as parallel corpus mining from the web[El\-Kishky et al\. \(2020\)](https://arxiv.org/html/2609.29511#bib.bib9);[Batextasciitilde nón et al\. \(2020\)](https://arxiv.org/html/2609.29511#bib.bib10);[Schwenk et al\. \(2021b\)](https://arxiv.org/html/2609.29511#bib.bib11), manually correcting Machine Translated text[Bhattacharjee et al\. \(2025\)](https://arxiv.org/html/2609.29511#bib.bib12), or full manual translation[Winata et al\. \(2023\)](https://arxiv.org/html/2609.29511#bib.bib13);[Singh et al\. \(2025\)](https://arxiv.org/html/2609.29511#bib.bib14)\.

However, most of them focused on a single domain, or there is no specific domain\. Some research that did create multi\-domain parallel data did not carry out any domain\-related MT experiments[Premjith et al\. \(2019\)](https://arxiv.org/html/2609.29511#bib.bib15)\. An exception is[Bhattacharjee et al\. \(2025\)](https://arxiv.org/html/2609.29511#bib.bib12)\. They created a dataset that covers three domains: Government, health and general, across 11 Indic languages\. They carried out domain\-specific experiments, however, only considered in\-domain experiments \(testing with a domain for which the NMT model was trained\)\. Similarly,[Appicharla et al\. \(2026\)](https://arxiv.org/html/2609.29511#bib.bib16)presented multi\-domain experiments across four domains for Hindi\-Mailithi and[Singh et al\. \(2025\)](https://arxiv.org/html/2609.29511#bib.bib14)presented a dataset across three domains for Bhili\-Hindi\. Experiments with these datasets have been limited to training and testing with a single domain at a time\. Other closest resources to ours are MENYO\-20k[Adelani et al\. \(2021\)](https://arxiv.org/html/2609.29511#bib.bib17), an English–Yorùbá corpus with six test domains used for domain\-adaptation experiments, and IN22[Gala et al\. \(2023\)](https://arxiv.org/html/2609.29511#bib.bib18), human\-translated test data over 13 domains for 22 Indic languages \(∼\\sim80 sentences per domain\)\. However, neither releases per\-domain training data at scale, nor supports cross\-domain experiments\.

### 2\.2Domain\-specific NMT Evaluation

Some other research focused on evaluating various facets of domain\-specific NMT for LRLs, using existing data[Ranathunga et al\. \(2024b\)](https://arxiv.org/html/2609.29511#bib.bib4);[Khiu et al\. \(2024\)](https://arxiv.org/html/2609.29511#bib.bib19);[Nayak et al\. \(2023\)](https://arxiv.org/html/2609.29511#bib.bib20)\. Most of these used fine\-tuned encoder\-decoder pre\-trained language models such as mBART[Tang et al\. \(2021\)](https://arxiv.org/html/2609.29511#bib.bib21)\. Out of these,[Ranathunga et al\. \(2024b\)](https://arxiv.org/html/2609.29511#bib.bib4)is the most comprehensive \- they investigated how fine\-tuned mBART performs across different domains, as well as how domain divergence impacts cross\-domain performance\. However, all these studies considered less than 5 domains\.

### 2\.3Sinhala\-Tamil\-English MT

Sinhala111Sinhala is also referred to asSinhalese,Singhala, andSinghalese[Englebretson and Genetti \(2005\)](https://arxiv.org/html/2609.29511#bib.bib22)\.is the native language of the Sinhalese people \(approximately 17 million\) and the only official language unique to Sri Lanka \([41](https://arxiv.org/html/2609.29511#bib.bib23),[2022](https://arxiv.org/html/2609.29511#bib.bib23)\)\. Sinhala belongs to the Indo\-European language family and the Indo\-Aryan branch[Arangala \(2024\)](https://arxiv.org/html/2609.29511#bib.bib24)\. Sinhala employs its own writing system and is morphologically rich[Sirisoma \(1990\)](https://arxiv.org/html/2609.29511#bib.bib25)\.

Tamil is the other official language in Sri Lanka\. Tamil belongs to the Dravidian language family[Krishnamurti \(2003\)](https://arxiv.org/html/2609.29511#bib.bib26)\. It employs its own Brahmic script[Daniels and Bright \(1996\)](https://arxiv.org/html/2609.29511#bib.bib27)and is an agglutinative language with rich morphology\. Although Tamil has a more mature NLP ecosystem than Sinhala[de Silva \(2026\)](https://arxiv.org/html/2609.29511#bib.bib28), it remains mid\-resourced[Ranathunga and de Silva \(2022\)](https://arxiv.org/html/2609.29511#bib.bib29)\. In particular, publicly available resources for Sri Lankan Tamil remain limited, and dialectal variation continues to hinder robust NLP applications[Mahaganapathy et al\. \(2026\)](https://arxiv.org/html/2609.29511#bib.bib30)\.

According to[Ranathunga and De Silva \(2022\)](https://arxiv.org/html/2609.29511#bib.bib31)’s categorisation, Sinhala is classified as a low\-resource language, and Tamil as a mid\-resource language\.

Being a multilingual country, MT systems are of paramount importance to Sri Lanka, especially in domains such as education, health and government communications\. The dearth of qualified human translators further reinforces the need for MT systems[Farhath et al\. \(2018a\)](https://arxiv.org/html/2609.29511#bib.bib32)\.

As a result, there have been research efforts to build MT systems for these language pairs \(specifically Si\-Ta\), for more than a decade\.[Pushpananda et al\. \(2014\)](https://arxiv.org/html/2609.29511#bib.bib33)trained the first Sinhala\-Tamil SMT systems on 25k parallel sentences, followed by several data augmentation efforts to improve the same[Farhath et al\. \(2018b\)](https://arxiv.org/html/2609.29511#bib.bib34);[Fernando et al\. \(2020\)](https://arxiv.org/html/2609.29511#bib.bib35);[Farhath et al\. \(2018a\)](https://arxiv.org/html/2609.29511#bib.bib32)\. Subsequent research has moved to Vanilla NMT[Tennage et al\. \(2017\)](https://arxiv.org/html/2609.29511#bib.bib36);[Pramodya et al\. \(2020\)](https://arxiv.org/html/2609.29511#bib.bib37), with techniques such as Byte Pair Encoding[Nissanka et al\. \(2020\)](https://arxiv.org/html/2609.29511#bib.bib38), back\-translation[Epaliyana et al\. \(2021\)](https://arxiv.org/html/2609.29511#bib.bib39), data augmentation[Fernando and Ranathunga \(2021\)](https://arxiv.org/html/2609.29511#bib.bib40)and transliteration[Tennage et al\. \(2018\)](https://arxiv.org/html/2609.29511#bib.bib41)techniques improving its performance\.[Thillainathan et al\. \(2021\)](https://arxiv.org/html/2609.29511#bib.bib42);[Lee et al\. \(2022\)](https://arxiv.org/html/2609.29511#bib.bib43)experimented with fine\-tuning pre\-trained seq\-seq language models such as mBART, which showed clear advantage over vanilla NMT\.[Ranathunga et al\. \(2024a\)](https://arxiv.org/html/2609.29511#bib.bib44)ultimately showed that fine\-tuning the translation\-specific NLLB model[NLLB Team et al\. \(2022\)](https://arxiv.org/html/2609.29511#bib.bib7)outperforms a fine\-tuned mBART model\.

None of this research publicly released the MT corpora it developed\. Several publicly available web\-mined corpora, such as CCMatrix\([Schwenk et al\., 2021b](https://arxiv.org/html/2609.29511#bib.bib11)\), CCAlign\([El\-Kishky et al\., 2020](https://arxiv.org/html/2609.29511#bib.bib9)\), WikiMatrix\([Schwenk et al\., 2021a](https://arxiv.org/html/2609.29511#bib.bib45)\)and NLLB\([NLLB Team et al\., 2022](https://arxiv.org/html/2609.29511#bib.bib7)\)cover the languages investigated in this study\. However, as reported by[Fernando et al\. \(2025\)](https://arxiv.org/html/2609.29511#bib.bib46);[Ranathunga et al\. \(2024a\)](https://arxiv.org/html/2609.29511#bib.bib44), such corpora are extremely noisy\. FLORES and MultiMWP corpus[Gamage et al\. \(2025\)](https://arxiv.org/html/2609.29511#bib.bib47)are the only human\-curated publicly available corpora that cover all three language pairs, according to our knowledge\. The former is an open\-domain corpus\. The latter only contains Math word problems\.

## 3Multi\-Domain Dataset Preparation

### 3\.1Data Sources and Preparation

We constructed parallel datasets from two types of sources: publicly available parallel corpora and domain\-specific monolingual corpora\. Table[1](https://arxiv.org/html/2609.29511#S3.T1)shows data source details\. Data selection was constrained by available data sources, translators, as well as funding\.

##### Publicly Available Parallel Corpora:

As mentioned in Section[2](https://arxiv.org/html/2609.29511#S2), publicly available web\-mined corpora are extremely noisy\. Therefore, following[Ranathunga et al\. \(2024a\)](https://arxiv.org/html/2609.29511#bib.bib44), we applied sentence\-level filtering to theNLLBandOpenSubtitlescorpora\. Specifically, we deduplicated each corpus based on the English side and scored each source–target sentence pair using LASER\-3[Heffernan et al\. \(2022\)](https://arxiv.org/html/2609.29511#bib.bib48)\. Then, from the ranked corpus, we selected the top 30,000 sentence pairs, excluding sentences containing fewer than three words\. We also used the Sinhala\-Tamil\-English portion of the MultiMWP corpus \(707 tri\-aligned pairs and 439 EnTa pairs\)[Gamage et al\. \(2025\)](https://arxiv.org/html/2609.29511#bib.bib47)as a training corpus\. Unlike the aforementioned corpora, this is a clean, human curated dataset\.

##### Monolingual Corpora for Machine Translation:

We obtained monolingual corpora from sources where the data can be reused without copyright restrictions\. Whenever possible, we consider documents related to Sri Lankan context\. In most cases, monolingual data was in English; these corpora were translated into Sinhala and Tamil using Google Translate and IndicTrans2[Gala et al\. \(2023\)](https://arxiv.org/html/2609.29511#bib.bib18), respectively\. For part of the news corpus originally available in Sinhala, Sinhala sentences were first translated into English using Google Translate, and the post\-edited English sentences were then translated into Tamil\.

All web\-mined and machine\-translated sentences were post\-edited by professional human translators to ensure translation quality\.

### 3\.2Test Set Creation

For each of the training datasets, we created a test set entirely through manual translation\. An additional test set was created using several corpora relating to LLMJailbreakprompts\. Depending on translator availability, the validation set was created either through manual translation or by post\-editing machine\-translated sentences\. Explicit checks were conducted to avaoid any overlapping between the three data splits\.

In addition, we used three existing parallel corpora in our experiments\. First, we extracted 25,000 parallel verses from theBiblecorpus, using the data sources and scripts reported by[Nayak et al\. \(2023\)](https://arxiv.org/html/2609.29511#bib.bib20)\. The corresponding test and validation sets were extracted separately to ensure that there was no overlap with the training data\. Second, we used 25,000 sentence pairs from a proprietaryGovernment\-domain corpus[Fernando et al\. \(2021\)](https://arxiv.org/html/2609.29511#bib.bib49);[Ranathunga et al\. \(2018\)](https://arxiv.org/html/2609.29511#bib.bib50)\. For this corpus, we used the separately released validation and test sets provided by the original authors\.

CorpusData TypeSourcesNLLBParallel![[Uncaptioned image]](https://arxiv.org/html/2609.29511v1/images/opus.png)[NLLB](https://opus.nlpl.eu/datasets/NLLB)OpenSubtitlesParallel![[Uncaptioned image]](https://arxiv.org/html/2609.29511v1/images/opus.png)[OpenSubtitles](https://opus.nlpl.eu/datasets/OpenSubtitles)WikipediaMonolingualLocal/Asian/European History \- Pages before 1945[https://www\.wikipedia\.org/](https://www.wikipedia.org/)NewsMonolingualLocal News \-[https://esana\.com\.lk/](https://esana.com.lk/)World News \-[https://www\.bbc\.com/news/world](https://www.bbc.com/news/world)LiteratureMonolingual![[Uncaptioned image]](https://arxiv.org/html/2609.29511v1/images/huggingface.png)[sil\-ai/bloom\-lm](https://huggingface.co/datasets/sil-ai/bloom-lm)[https://www\.gutenberg\.org/](https://www.gutenberg.org/)MathsMonolingual![[Uncaptioned image]](https://arxiv.org/html/2609.29511v1/images/huggingface.png)[microsoft/orca\-math\-word\-problems\-200k](https://huggingface.co/datasets/microsoft/orca-math-word-problems-200k)[https://www\.gutenberg\.org/](https://www.gutenberg.org/)HealthMonolingual[https://www\.who\.int/mega\-menu/data/reports](https://www.who.int/mega-menu/data/reports)[https://fhb\.health\.gov\.lk/resources](https://fhb.health.gov.lk/resources)JailbreakMonolingual[Liu et al\. \(2023\)](https://arxiv.org/html/2609.29511#bib.bib51)![[Uncaptioned image]](https://arxiv.org/html/2609.29511v1/images/huggingface.png)[jackhhao/jailbreak\-classification](https://huggingface.co/datasets/jackhhao/jailbreak-classification/blob/main/default/jailbreak_dataset_train.csv)Table 1:The sources from which the parallel and monolingual data were obtained for corpus preparation
### 3\.3Post\-Editing & Review Process

Manual translation and post\-editing were subject to a rigorous quality\-control process\. The translators for the task were selected via a screening process\. Whenever possible, domain\-specific translation experience was given priority\. For example, health professionals/students with translation experience were selected for the health domain\. Selected translators were introduced to the task through a recorded demonstration video\. In addition, we prepared customised translation guidelines for each domain and language pair; Appendix[B](https://arxiv.org/html/2609.29511#A2)provides a sample guideline prepared for English–Sinhala translators\. The well\-experienced translators identified through the selection process were employed as reviewers\.

For each domain, translators first completed post\-editing tasks and subsequently carried out manual translation of the test sets\. Translators’ work was continuously reviewed and necessary feedback was provided\.

This process was applied to both English–Sinhala and English–Tamil translation\. To construct the Sinhala–Tamil corpus, we aligned the corresponding Sinhala and Tamil corpora using English as a pivot language\. If Sinhala–Tamil translators were available222Translators are much scarce for this pair\., the parallel data underwent a trilingual alignment with further refinement\.

Over a period of three years, more than a 110 translators across the three language pairs contributed to the project\. Translators were compensated in one of two ways, depending on their preference\. Most translators received monetary compensation based on the government\-approved translation rates of the relevant country\. Others were offered co\-authorship, following the increasingly adopted participatory research approach in NLP[Chang et al\. \(2025\)](https://arxiv.org/html/2609.29511#bib.bib52), where contributors involved in language\-data creation are recognised as research collaborators\.

### 3\.4Dataset Statistics & Analysis

Table[2](https://arxiv.org/html/2609.29511#S3.T2)shows the statistics of the corpora we used for experiments\. Not shown in this table are the 1000 sentence pairs per domain that we maintain as a hidden set, to be used in a future shared task\. As mentioned earlier,Bible,Government, andFLORES\+corpora are borrowed from prior research\. The corpus we compiled totals up to 200k\+ for training, 8k\+ for validation and 11k\+ for testing\. Altogether, the corpus spans across 10 domains333The notion of a ‘domain’ in MT is rather ambiguous\. It can be interpreted with respect to topic and genre[Van der Wees et al\. \(2015\)](https://arxiv.org/html/2609.29511#bib.bib53)\. However, identifying the exact domain of our datasets is out of scope for our work\. Rather, we adopt[Koehn and Knowles \(2017\)](https://arxiv.org/html/2609.29511#bib.bib54)’s definition of a domain in MT \- ‘a domain is defined by a corpus from a specific source, and may differ from other domains in topic, genre, style, level of formality, etc\.’\.

\#DomainDir\.TrainDevTest3en\-si1,000406890en\-ta2,0006301,000Healthen\-si\-ta––784†5en\-si6,6741,0001,000en\-ta3,500511864si\-ta––831†Literatureen\-si\-ta802†––3en\-si6,607634–en\-ta4,645644–Mathsen\-si\-ta707†–1,0004en\-si16,3371,0001,000en\-ta10,000602996Newsen\-si\-ta10,000†317†992†5NLLBen\-si\-ta25,0371,0001,0006en\-si25,007––en\-ta28,350––si\-ta25,111––OpenSubtitlesen\-si\-ta–1,0001,0007en\-si24,160964–en\-ta23,879––Wikipediaen\-si\-ta20,502†1,0001,0008Jailbreaken\-si\-ta––1,0009en\-si25,0001,0001,000Bibleen\-ta25,0001,0001,00010Governmenten\-si\-ta25,0001,0001,00011FLORES\+en\-si\-ta––1,012Total277,30712,39114,762Table 2:Statistics of corpora used for experiments\. Last three have been borrowed from other sources\. Totals exclude trilingual \(en\-si\-ta\) subsets marked†, which are derived from or largely overlapping with the bilingual pairs listed for the same domain and splitFigure 1:Distribution of English Sentence Lengths by Domain \(Logarithmic Scale\)In Fig[1](https://arxiv.org/html/2609.29511#S3.F1)we visualize the English sentence lengths of each domain using the same colour scheme as Table[2](https://arxiv.org/html/2609.29511#S3.T2)\. It can be noted thatJailbreakdataset shows the most irregular distribution\. This is unsurprising given the specialised purpose of that dataset\.LiteratureandOpenSubtitlesshow similar distributions favouring shorter utterances; however,Literaturedoes have a considerable count of single\-word exclamations or utterances as well as some rather long sentences\.Bible,Government, andHealthshow an affinity to longer sentences\.NLLB,Bible,Wikipedia, andNewsshow the closest resemblance to normal distributions\.

In order to further quantify the differences across domains, we calculated the Jensen–Shannon Divergence \(JSD\)[Lin \(1991\)](https://arxiv.org/html/2609.29511#bib.bib55)\. Data sizes used for this calculation range from∼\\sim16k to 25k \(domains with limited training data were brought into this range by sourcing additional data considering their original distribution\)444FLORES was not considered in this calculation, since we could not identify its exact source\.\. Accordingly, the calculation considered only English data\.

![Refer to caption](https://arxiv.org/html/2609.29511v1/js_divergence_heatmap.png)Figure 2:Pairwise JSD values between domains, computed on English training samplesFigure[2](https://arxiv.org/html/2609.29511#S3.F2)shows the pairwise JSD values between the 10 domains\. Overall,NLLBshows low divergence withOpenSubtitles,Literature,News, andJailbreak, whileLiteraturealso shows low divergence withBibleand as doesNewswithGovernment\. These pairs have JSD values between 0\.12 and 0\.18\.NLLB, being open\-domain, has the lowest divergence against four other domains, which could be due to it being a web\-mined corpus\. In contrast,OpenSubtitlesshows high divergence withHealth,Government, andWikipedia;Healthalso diverges strongly fromBibleandLiterature, whileBibleandGovernmentform another highly divergent pair\. These pairs have JSD values between 0\.34 and 0\.40\.OpenSubtitles, which has shorter, spoken\-style sentences, has the highest divergence against three other domains\. Similarly,Health, with its highly domain\-specific terminology, also shows high divergence against three other domains\.

In Figure[3](https://arxiv.org/html/2609.29511#S3.F3)we show the KL divergence across all language pair directions by dataset\. The reason for using KLD for this analysis is twofold: \(1\) Unlike JSD, KLD is directional and thus allows us to show cases where a pair is relatively highly divergent in one direction and less so in the other\. This is well evident in Figure[3\(a\)](https://arxiv.org/html/2609.29511#S3.F3.sf1)forBible, where we can observe high divergence in the EnSi and EnTa direction attesting the fact that Sinhala and Tamil languages had not evolved alongside Christianity as much as English did and as such, needs more words in a sentence to discuss something English would do so with fewer555One good example for this was the translation of the word ”damned”\. Which is a simple enough concept to explain in a single word in English but in Sinhala, not only did it take 3 to 4 words, but also those phrases were not exactly the same across multiple occurrences\., \(2\) JSD is bounded and thus high divergence saturate\. In Figure[3\(b\)](https://arxiv.org/html/2609.29511#S3.F3.sf2), we can see the intricacies of the the levels of divergence through KLD\. But when the same visualization was created with JSD, all datasets got flattened out\. For example,Mathsindiscriminately showed 0\.62 across all language pairs whileNLLB,News,Wikipedia, andOpenSubtitlesall flattened out to 0\.68\. But with KLD in Figure[3\(b\)](https://arxiv.org/html/2609.29511#S3.F3.sf2)for example we can note the interesting observation ofGovernmenthaving relatively low divergence in EnSi and EnTa direction owing to the fact that government domain words that did not exist in Sinhala or Tamil before the colonial rule were explicitly ‘‘created’’ usually as loanwords and at least as calques, resulting in lower divergence in word frequencies666It should be noted that these loanwords or calques are not always sourced from English\. in fact, a considerable number of Sinhala words in this domain has roots in Portuguese or even Dutch who colonised Sri Lanka before the English\. But this does not dilute the fact that one word in English in this domain can map to one word in Sinhala or Tamil, at a relatively high frequency\.\.

![Refer to caption](https://arxiv.org/html/2609.29511v1/KL-SL.png)\(a\)Sentence Length
![Refer to caption](https://arxiv.org/html/2609.29511v1/KL-WF.png)\(b\)Word Frequencies

Figure 3:KL Divergence Across All Language Pair Directions by Dataset![Refer to caption](https://arxiv.org/html/2609.29511v1/grid.png)Figure 4:Source\-domain \(columns\) vs cross\-domain \(rows\) results for EnSi direction \(chrF\+\+\)Figure 5:Language direction\-wise results of models trained and tested with Government data \(chrF\+\+\)Figure 6:Cross\-domain results of Government\-trained models for EnSi direction \(chrF\+\+\)Figure 7:Strategy comparison across training data \(chrF\+\+\) for EnSi

## 4Experiments

### 4\.1Experiment Plan

We conducted a comprehensive set of experiments to analyse the impact of training\-data size, model architecture, and domain\. We selected three types of models for fine\-tuning: a standard encoder–decoder model implemented in Fairseq, a translation\-specific encoder–decoder model \(NLLB\-600M\)[NLLB Team et al\. \(2022\)](https://arxiv.org/html/2609.29511#bib.bib7), and an LLM \(the instruction\-tuned Gemma 3 1B,gemma\-3\-1b\-it\)\. Owing to computational resource constraints, we focused primarily on models with relatively small parameter counts\. Each model was fine\-tuned under six training\-data settings: 1k, 5k, 10k, 15k, 20k, and 25k sentence pairs\. We evaluated performance in both in\-domain \(models were trained and tested on data from the same domain\) and out\-of\-domain \(models were tested on a domain different to the training domain\) settings\.

We then conducted an ablation study to examine the effect of model choice\. In addition to the models described above, we fine\-tuned larger Gemma variants, namely Gemma 3 4B and 12B, as well as TranslateGemma 4B and 12B\. To keep the experimental space tractable, this analysis was limited to the government domain\.

Third, we investigated the effects of multilingual fine\-tuning, multi\-domain fine\-tuning, and combined multi\-domain multilingual fine\-tuning using Fairseq, Gemma 12B and NLLB\. For this experiment, we sampled 1,000 sentence pairs from each domain for each language pair\.

### 4\.2Experiment Setup

We adapt all LLMs with Low\-Rank Adaptation \(LoRA\)\([Hu et al\., 2022](https://arxiv.org/html/2609.29511#bib.bib56)\), using a minimal translation prompt in each model’s native chat template\. The recipe is held fixed across the Gemma 3 1B/4B/12B variants so that quality differences reflect model scale, whereas TranslateGemma uses a deliberately lighter recipe to limit erosion of its reinforcement\-learned translation ability\([Finkelstein et al\., 2026](https://arxiv.org/html/2609.29511#bib.bib57);[Biderman et al\., 2024](https://arxiv.org/html/2609.29511#bib.bib58)\), and is therefore assessed against its own zero\-shot baseline rather than compared like\-for\-like with base Gemma 3\. Prompt formats, full hyperparameters \(Table[3](https://arxiv.org/html/2609.29511#A1.T3)\), the rationale for the asymmetric recipes, and hardware details are given in Appendix[A](https://arxiv.org/html/2609.29511#A1)\.

##### Combined\-corpus fine\-tuning\.

The multi\-domain experiment trains one adapter per translation direction on a mixture of 1k sentence pairs from each domain’s training portion; the per\-domain multilingual experiment trains, for each of the five domains with data in all three languages, one adapter jointly on all six translation directions; and the combined multilingual \+ multi\-domain experiment trains a single adapter on all six directions of the multi\-domain mixture\. All three settings are run with NLLB, Fairseq, and Gemma 3 12B \(the multi\-domain setting also with TranslateGemma 12B\), with all other settings matching the corresponding single\-domain recipes\.

## 5Automatic Evaluation

Due to space limitations and brevity, we use ChrF\+\+ for our discussion\. Summary results for En\-Si is shown in Figure[4](https://arxiv.org/html/2609.29511#S3.F4)\. In particular, we noticed that COMET scores are inconsistent and unreliable\. For example, fully degenerate Gemma\-1B outputs \(chrF\+\+ at or below 1\.5\) still receive COMET scores of 0\.40 on average \(up to 0\.54\), further affirming the observations in the prior literature[Falctextasciitilde ao et al\. \(2024\)](https://arxiv.org/html/2609.29511#bib.bib59)\.

### 5\.1Impact of training\-data size, model architecture, and domain\.

#### 5\.1\.1In\-domain

##### Individual Domain Complexity:

In\-domain experiments with the vanilla Transformer model helps us understand the domain\-wise translation difficulty\. At 5k,LiteratureandOpenSubtitlesare consistently the hardest domains to learn from scratch;MathsandBiblethe easiest\.LiteratureandOpenSubtitles, along withNLLBare the hardest domains for Gemma\-3\. In contrast, for NLLB\-600M,OpenSubtitlesbecomes the easiest domain, whileLiteratureremains the hardest\. This indicates the bias introduced during the pre\-training stage\. Even after training with 25k, the difference between the domains exists, with the maximum difference being about 10 ChRF\+\+ points\.

In order to identify the reason behind this variation, we calculated the following per\-domain statistics from the actual training/test corpora \(English and target sides\): average sentence length, vocabulary diversity \(type–token ratio, i\.e\., distinct words per 10k running words\), test\-set OOV rate \(share of test words never seen in the training data\), and train–test overlap\. We observe that the dominant factor for the Model training is the token budget, not lexical difficulty\. Domains differ enormously in words per sentence\. For example,OpenSubtitleshas an average of 7\.7 English words per sentence,Literature11\.7,Bible25\.2 andHealthhas 26\.1 words\. This means, for example, the number of training tokens in 5kBiblecorpus is more than three times that ofOpenSubtitles\.

##### Model\-Specific Performance:

In all the experiments, NLLB\-600M always outperforms the other two models\. However, its margin over Gemma\-1B shrinks as the training dataset size increases\. Gemma\-1B outperforms Fairseq when the training dataset size is small, however the latter catches up, when the training dataset size rises to 25k\.

##### Impact of Fine\-tuning:

NLLB is a strong zero\-shot baseline on these domains \(47\.0 chrF\+\+, averaged over the nine trainable domains and all language directions\)\. Fine\-tuning further improves the performance consistently: 1k in\-domain pairs results in a gain of 4\.6 ChRF\+\+ points on average\. When the dataset size goes to 25k, a further gain of 3\.9 points is recorded, on average\. Gemma\-1B starts near zero \(8\.0 average zero\-shot\), however it benefits immensely from fine\-tuning, with a gain of 29\.7 points over the zero\-shot scores and gains up to \+39\.1 at full size, while Fairseq averages 5\.4 at 1k and gains \+7\.3 to \+40\.6 from 1k to each domain’s maximum\. The three model classes therefore differ in where additional data helps: NLLB obtains about half of its total gain from the first 1k pairs, whereas Fairseq and Gemma\-1B need volume before domain\-specific data pays off\. Interestingly, NLLB fine\-tuned on just 1k pairs beats both competitors trained on each domain’s entire corpus, in all 9 domains

#### 5\.1\.2Out\-domain

Out\-domain results are mostly lower than the corresponding in\-domain results, as expected \(e\.g\. result forBibleis less, when a model trained with another domain is used\)\. Some domains seem to be more favourable than others\. For example, for theGovernmentdomain, models trained with NLLB data are better than usingBibledata\. This effect is more evident in the vanilla Transformer model\. To investigate this further, we calculate the correlation between JSD and the result from each model\. For the vanilla Transformer, an r of \-0\.56 suggests that there is a high negative correlation between the divergence between the two domains and the corresponding out\-domain result\. However, the impact of domain divergence becomes less visible in the pre\-trained models\.

We also note that fine\-tuning NLLB with more data from a highly divergent domain steadily erodes its performance on the held\-out domains\. Averaged over all language directions, its performance on FLORES\+ drops by 7\.2 chrF\+\+ when theBibletraining set is increased from 1k to 25k, and by 2\.1 chrF\+\+ forGovernment\. In contrast, less divergent domains are free or even show a gain \(open\-web \+0\.8,News\+0\.1,OpenSubtitles\+0\.2,Wikipedia\+0\.2 on FLORES\+\)\. The erosion rate correlates with the domain’s mean divergence from the other domains \(rank correlation \-0\.71 across the 8 domains\)\. TheJailbreaktest set shows the same pattern \(Bible\-9\.2,Government\-2\.3, broad domains near zero\)\. However, Gemma\-1B does not seem to have this issue\. Every domain improves its FLORES\+ score \(\+2\.5 to \+22\.5\)\. A model that is still learning to translate only benefits from more parallel data; forgetting is a phenomenon of already\-competent translation models\. The same contrast appears withinTranslateGemmaunder an identical light fine\-tuning recipe: the 4B variant loses about 3 FLORES\+ points while the 12B variant loses none; model scale also protects previously learned translation ability\.

The best single “donor” domain is model\-dependent:Wikipediafor the vanilla Transformer \(28\.1 mean chrF\+\+ across the other domains\) and for Gemma\-1B \(33\.8\), andHealthfor NLLB \(50\.4, from only 1k pairs\);BibleorLiteratureis the worst donor for every model \(NLLB withBible: 40\.0\)\. For NLLB the spread between donors is small, because its zero\-shot performance is already strong\. The largest in/out\-domain penalty for NLLB is onOpenSubtitles\(in\-domain 60\.4 vs 56\.9 for the best donor\), so conversational language is the hardest register to reach from any other domain\.

### 5\.2Language\-wise Analysis

Figures[4](https://arxiv.org/html/2609.29511#S3.F4)and[5](https://arxiv.org/html/2609.29511#S3.F5)depict these results\. Three independent lines of evidence identify generating Sinhala as the system\-building bottleneck, beyond scores simply being lower\. First, training collapses concentrate there: counting fine\-tuned models whose output remains degenerate \(chrF\+\+ at or below 1\.5\), the vanilla Transformer has 159 such cases and Gemma\-1B 73, every one of Gemma\-1B’s being into Sinhala \(en\-si 43, ta\-si 30; none into Tamil or English\), while NLLB has none anywhere\. Into\-Sinhala also needs roughly 5\-10x more data to ignite for Gemma\-1B \(open\-web corpus, en\-si: 1\.1 chrF\+\+ at 5k, 6\.2 at 10k, 22\.3 at 25k\)\. Second, a controlled comparison isolates the target side: the Sinhala\-Tamil corpora contain the same sentence pairs in both directions, so si\-ta vs ta\-si compares generating Tamil vs Sinhala on identical content; si\-ta wins in 17 of 19 model\-domain cells, by up to \+15\.9 chrF\+\+\. Third, the gap survives fine\-tuning for small models: at each domain’s maximum size en\-ta exceeds en\-si by \+11\.6 chrF\+\+ on average for NLLB and \+15\.8 for Gemma\-1B, although the 4B/12B\-class models close or reverse this gap on the one domain where we can test them \(Government: Gemma\-4B \-1\.3, Gemma\-12B \-2\.2, TranslateGemma\-12B \+0\.9\)\. Across all trained in\-domain results the target\-language averages are English 42\.7, Tamil 37\.9, Sinhala 31\.3\. The source side shows no comparable asymmetry: zero\-shot NLLB’s ta\-en advantage over si\-en is small and mixed in sign \(\-6\.7 to \+11\.4 across domains\), so the difficulty lies in generating Sinhala rather than in understanding it\.

### 5\.3Effect of model choice

As per Figures[4](https://arxiv.org/html/2609.29511#S3.F4)and[5](https://arxiv.org/html/2609.29511#S3.F5), on theGovernmentdomain at 25k, the model ladder is Fairseq 40\.0 < TranslateGemma\-4B 42\.3 < Gemma\-1B 46\.5 < TranslateGemma\-12B 53\.6 < Gemma\-4B 54\.6 < NLLB\-600M 56\.0 < Gemma\-12B 57\.1 \(chrF\+\+, averaged over all six directions\)\. Base Gemma\-3 scales cleanly with size once fine\-tuned \(46\.5, 54\.6, 57\.1\), yet a 600M\-parameter translation\-specific model beats every LLM except the 12B\. Bigger is not better zero\-shot: Gemma\-12B is worse than Gemma\-1B/4B zero\-shot on the low\-resource directions \(e\.g\. si\-en 7\.6 vs 13\.1/13\.9\), likely due to chat\-style formatting and verbosity; fine\-tuning restores the expected order, so zero\-shot LLM quality is a poor predictor of post\-fine\-tuning quality at these scales\. And TranslateGemma\-12B, already translation\-tuned, scores 43\.2 zero\-shot onGovernmentwhere same\-size base Gemma\-12B is near zero; fine\-tuning adds a further \+10\.4 chrF\+\+, while its 4B variant can degrade \(en\-ta 39\.1 to 29\.8\), consistent with fine\-tuning damaging an already\-optimized translation policy at small scale\.

### 5\.4Multilingual vs Multi\-Domain

We separately compiled a multi\-domain and a multilingual corpus, by sampling out 1k from each domain and language pair\. We trained the three models using these corpora separately, as well as by combining them\. According to Figure[7](https://arxiv.org/html/2609.29511#S3.F7), if a given domain has a dedicated corpus of about 25k, it is generally better than having a multilingual\+multi\-domain corpus of double the size, where each domain is equally represented across language pairs\. On the other hand, if a domain has a small corpus \(less than 5k\), combining it with data from other domains and language pairs is beneficial\. Note that here we merged all the corpora, without considering the domain divergence\. Hand\-picking domains based on the domain divergence might result in better gains, however, exhaustive experiments to verify this is beyond the scope of the current research\.

## 6Conclusion

We presentedEnSiTa, a trilingual multi\-domain parallel dataset and benchmark for English, Sinhala and Tamil, covering all six translation directions and eight domains\. Using this benchmark, we carried out an extensive study of domain\-specific MT for low\-resource languages, spanning training\-data size, model architecture and scale, cross\-domain transfer, and multilingual and multi\-domain fine\-tuning\. Our results translate into concrete guidance for practitioners building domain\-specific MT systems under a limited budget: when starting from a pre\-trained translation model, a small curated in\-domain set is on average worth more than any amount of out\-of\-domain data, most strongly when translating into English, whereas from\-scratch models need volume before domain fit begins to matter; domain divergence reliably predicts transfer loss for from\-scratch models but matters far less once a model is pre\-trained; and continued fine\-tuning on a divergent domain erodes the general translation ability of an already\-competent model, while a model still learning to translate only benefits from more parallel data\. We will release the dataset, the fine\-tuned models, and all evaluation outputs\. We hopeEnSiTaenables the systematic study of domain effects in low\-resource MT that has so far been possible only for high\-resource languages, and we see document\-level translation, additional domains, and human evaluation of multi\-domain systems as natural next steps\.

## Acknowledgements

Data creation and curation was made possible by a Google Diversity and Inclusion grant received by Surangika Ranathunga and Nisansa de Silva\. We thank OpenToken \(opentoken\.global\) for providing the GPU compute for LLM experiments, and LeafCloud \(leaf\.cloud\) for the underlying infrastructure\. We also thank the National Languages Processing Centre \(NLPC\), at the University of Moratuwa for providing the GPUs for Fairseq and NLLB\-600 experiments\. We also acknowledge the inputs from Madhavi Perera and Kengatharaiyar Sarveswaran during the initial phases of the project\. We are grateful to Dhanika Perera from Bhasha Lanka \(Pvt\) Ltd for providing a Sinhala news corpus\.

## References

- Adelaniet al\.\(2021\)D\. I\. Adelani, D\. Ruiter, J\. O\. Alabi, D\. Adebonojo, A\. Ayeni, M\. Adeyemi, A\. E\. Awokoya, and C\. Espatextasciitilde na\-BonetThe effect of domain and diacritics in Yoruba–English neural machine translation\.InProceedings of Machine Translation Summit XVIII: Research Track,pp\. 61–75\.External Links:[Link](https://aclanthology.org/2021.mtsummit-research.6/)Cited by:[§2\.1](https://arxiv.org/html/2609.29511#S2.SS1.p2.1)\.
- Appicharlaet al\.\(2026\)R\. Appicharla, S\. K\. Jha, A\. Ekbal, and P\. BhattacharyyaMaithilimt: developing multi\-domain parallel corpus for hindi\-maithili machine translation\.Language Resources and Evaluation60\(1\),pp\. 12\.Cited by:[§2\.1](https://arxiv.org/html/2609.29511#S2.SS1.p2.1)\.
- Arangala \(2024\)R\. ArangalaLocation of the Sinhala in Regional Linguistic Historicity and the Identity of Sinhala Language\.Journal of Desk Research Review and Analysis2\(1\)\.Cited by:[§2\.3](https://arxiv.org/html/2609.29511#S2.SS3.p1.1)\.
- Batextasciitilde nónet al\.\(2020\)M\. Batextasciitilde nón, P\. Chen, B\. Haddow, K\. Heafield, H\. Hoang, M\. Esplà\-Gomis, M\. L\. Forcada, A\. Kamran, F\. Kirefu, P\. Koehn,et al\.ParaCrawl: web\-scale acquisition of parallel corpora\.InProceedings of the 58th annual meeting of the association for computational linguistics,pp\. 4555–4567\.Cited by:[§2\.1](https://arxiv.org/html/2609.29511#S2.SS1.p1.1)\.
- Bhattacharjeeet al\.\(2025\)S\. Bhattacharjee, M\. K\. Roy, Y\. Poojary, B\. Dave, M\. Raj, V\. Mujadia, B\. Gain, P\. Mishra, A\. Ahsan, P\. Krishnamurthy,et al\.CorIL: towards enriching indian language to indian language parallel corpora and machine translation systems\.arXiv preprint arXiv:2509\.19941\.Cited by:[§2\.1](https://arxiv.org/html/2609.29511#S2.SS1.p1.1),[§2\.1](https://arxiv.org/html/2609.29511#S2.SS1.p2.1)\.
- Bidermanet al\.\(2024\)D\. Biderman, J\. Portes, J\. J\. Gonzalez Ortiz, M\. Paul, P\. Greengard, C\. Jennings, D\. King, S\. Havens, V\. Chiley, J\. Frankle, C\. Blakeney, and J\. P\. CunninghamLoRA learns less and forgets less\.Transactions on Machine Learning Research \(TMLR\)\.Cited by:[Appendix A](https://arxiv.org/html/2609.29511#A1.SS0.SSS0.Px2.p2.1),[§4\.2](https://arxiv.org/html/2609.29511#S4.SS2.p1.1)\.
- Changet al\.\(2025\)T\. A\. Chang, C\. Arnett, A\. Eldesokey, A\. Sadallah, A\. Kashar, A\. Daud, G\. O\. Abosede, L\. M\. Adamu, A\. Praise, M\. S\. Adhikarinayum,et al\.Global piqa: evaluating physical commonsense reasoning across 100\+ languages and cultures\.Cited by:[§3\.3](https://arxiv.org/html/2609.29511#S3.SS3.p4.1)\.
- Daniels and Bright \(1996\)P\. T\. Daniels and W\. BrightThe world’s writing systems\.Oxford University Press on Demand\.Cited by:[§2\.3](https://arxiv.org/html/2609.29511#S2.SS3.p2.1)\.
- de Silva \(2026\)N\. de SilvaSurvey on Publicly Available Sinhala Natural Language Processing Tools and Research\.arXiv preprint arXiv:1906\.02358v27\.Cited by:[§2\.3](https://arxiv.org/html/2609.29511#S2.SS3.p2.1)\.
- Dettmerset al\.\(2022\)T\. Dettmers, M\. Lewis, S\. Shleifer, and L\. Zettlemoyer8\-bit optimizers via block\-wise quantization\.InInternational Conference on Learning Representations \(ICLR\),Cited by:[Appendix A](https://arxiv.org/html/2609.29511#A1.p2.1)\.
- El\-Kishkyet al\.\(2020\)A\. El\-Kishky, V\. Chaudhary, F\. Guzmán, and P\. KoehnCCAligned: a massive collection of cross\-lingual web\-document pairs\.InProceedings of the 2020 Conference on Empirical Methods in Natural Language Processing \(EMNLP\),pp\. 5960–5969\.Cited by:[§2\.1](https://arxiv.org/html/2609.29511#S2.SS1.p1.1),[§2\.3](https://arxiv.org/html/2609.29511#S2.SS3.p6.1)\.
- Englebretson and Genetti \(2005\)R\. Englebretson and C\. GenettiSanta barbara papers in linguistics: proceeding from the workshop on sinhala linguistics\.Santa Barbara, CA: Department of Linguistics at the University of California, Santa Barbara\.Cited by:[footnote 1](https://arxiv.org/html/2609.29511#footnote1)\.
- Epaliyanaet al\.\(2021\)K\. Epaliyana, S\. Ranathunga, and S\. JayasenaImproving back\-translation with iterative filtering and data selection for sinhala\-english nmt\.In2021 Moratuwa Engineering Research Conference \(MERCon\),pp\. 438–443\.Cited by:[§2\.3](https://arxiv.org/html/2609.29511#S2.SS3.p5.1)\.
- Falctextasciitilde aoet al\.\(2024\)J\. Falctextasciitilde ao, C\. Borg, N\. Aranberri, and K\. AbelaCOMET for low\-resource machine translation evaluation: a case study of english\-maltese and spanish\-basque\.InProceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation \(LREC\-COLING 2024\),pp\. 3553–3565\.Cited by:[§5](https://arxiv.org/html/2609.29511#S5.p1.1)\.
- Farhathet al\.\(2018a\)F\. Farhath, S\. Ranathunga, S\. Jayasena, and G\. DiasIntegration of bilingual lists for domain\-specific statistical machine translation for sinhala\-tamil\.In2018 Moratuwa Engineering Research Conference \(MERCon\),pp\. 538–543\.Cited by:[§2\.3](https://arxiv.org/html/2609.29511#S2.SS3.p4.1),[§2\.3](https://arxiv.org/html/2609.29511#S2.SS3.p5.1)\.
- Farhathet al\.\(2018b\)F\. Farhath, P\. Theivendiram, S\. Ranathunga, S\. Jayasena, and G\. DiasImproving domain\-specific smt for low\-resourced languages using data from different domains\.InProceedings of the Eleventh International Conference on Language Resources and Evaluation \(LREC 2018\),Cited by:[§2\.3](https://arxiv.org/html/2609.29511#S2.SS3.p5.1)\.
- Fernandoet al\.\(2025\)A\. Fernando, N\. de Silva, M\. Velayuthan, C\. Rathnayake, and S\. RanathungaImproving the quality of web\-mined parallel corpora of low\-resource languages using debiasing heuristics\.InProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing,pp\. 28264–28281\.External Links:[Document](https://dx.doi.org/10.18653/v1/2025.emnlp-main.1435),ISBN 979\-8\-89176\-332\-6Cited by:[§2\.3](https://arxiv.org/html/2609.29511#S2.SS3.p6.1)\.
- Fernandoet al\.\(2020\)A\. Fernando, S\. Ranathunga, and G\. DiasData augmentation and terminology integration for domain\-specific sinhala\-english\-tamil statistical machine translation\.arXiv preprint arXiv:2011\.02821\.Cited by:[§2\.3](https://arxiv.org/html/2609.29511#S2.SS3.p5.1)\.
- Fernandoet al\.\(2021\)A\. Fernando, S\. Ranathunga, and G\. DiasData augmentation and terminology integration for domain\-specific sinhala\-english\-tamil statistical machine translation\.External Links:2011\.02821,[Link](https://arxiv.org/abs/2011.02821)Cited by:[§3\.2](https://arxiv.org/html/2609.29511#S3.SS2.p2.1)\.
- Fernando and Ranathunga \(2021\)A\. Fernando and S\. RanathungaData augmentation to address out of vocabularyproblem in low resource sinhala english neural machine translation\.InProceedings of the 35th Pacific Asia Conference on Language, Information and Computation,pp\. 61–70\.Cited by:[§2\.3](https://arxiv.org/html/2609.29511#S2.SS3.p5.1)\.
- Finkelsteinet al\.\(2026\)M\. Finkelstein, I\. Caswell, T\. Domhan, J\. Peter, J\. Juraska, P\. Riley, D\. Deutsch, G\. Kovacs, C\. Dilanni, C\. Cherry, E\. Briakou, E\. Nielsen, J\. Luo, K\. Black, R\. Mullins, S\. Agrawal, W\. Xu, E\. Kats, S\. Jaskiewicz, M\. Freitag, and D\. VilarTranslateGemma technical report\.arXiv preprint arXiv:2601\.09012\.Cited by:[Appendix A](https://arxiv.org/html/2609.29511#A1.SS0.SSS0.Px2.p2.1),[§4\.2](https://arxiv.org/html/2609.29511#S4.SS2.p1.1)\.
- Galaet al\.\(2023\)J\. P\. Gala, P\. A\. Chitale, V\. Gumma, S\. Doddapaneni, K\. M\. Aswanth, J\. A\. Nawale, A\. Sujatha, R\. Puduppully, V\. Raghavan, P\. Kumar,et al\.IndicTrans2: towards high\-quality and accessible machine translation models for all 22 scheduled indian languages\.\.Transactions on Machine Learning Research2023\.Cited by:[§2\.1](https://arxiv.org/html/2609.29511#S2.SS1.p2.1),[§3\.1](https://arxiv.org/html/2609.29511#S3.SS1.SSS0.Px2.p1.1)\.
- Gamageet al\.\(2025\)O\. Gamage, S\. Ranathunga, A\. Lee, X\. Sun, A\. Singh, M\. P\. Skenduli, M\. Alam, A\. K\. Nayak, H\. Gao, B\. Deori,et al\.A multilingual dataset \(multimwp\) and benchmark for math word problem generation\.IEEE Transactions on Audio, Speech and Language Processing33,pp\. 1838–1848\.Cited by:[§2\.3](https://arxiv.org/html/2609.29511#S2.SS3.p6.1),[§3\.1](https://arxiv.org/html/2609.29511#S3.SS1.SSS0.Px1.p1.1)\.
- Gemma Teamet al\.\(2025\)Gemma Team, A\. Kamath, J\. Ferret, S\. Pathak, N\. Vieillard, R\. Merhej, S\. Perrin, T\. Matejovicova, A\. Ramé, M\. Rivière,et al\.Gemma 3 technical report\.arXiv preprint arXiv:2503\.19786\.Cited by:[§1](https://arxiv.org/html/2609.29511#S1.p3.1)\.
- Heffernanet al\.\(2022\)K\. Heffernan, O\. Çelebi, and H\. SchwenkBitext mining using distilled sentence representations for low\-resource languages\.InFindings of the Association for Computational Linguistics: EMNLP 2022,pp\. 2101–2112\.Cited by:[§3\.1](https://arxiv.org/html/2609.29511#S3.SS1.SSS0.Px1.p1.1)\.
- Huet al\.\(2022\)E\. J\. Hu, Y\. Shen, P\. Wallis, Z\. Allen\-Zhu, Y\. Li, S\. Wang, L\. Wang, and W\. ChenLoRA: low\-rank adaptation of large language models\.InInternational Conference on Learning Representations \(ICLR\),Cited by:[§4\.2](https://arxiv.org/html/2609.29511#S4.SS2.p1.1)\.
- Khiuet al\.\(2024\)E\. Khiu, H\. Toossi, D\. Anugraha, J\. Liu, J\. Li, J\. Flores, L\. Roman, A\. S\. Doğruöz, and E\. LeePredicting machine translation performance on low\-resource languages: the role of domain similarity\.InFindings of the Association for Computational Linguistics: EACL 2024,pp\. 1474–1486\.Cited by:[§2\.2](https://arxiv.org/html/2609.29511#S2.SS2.p1.1)\.
- Koehn and Knowles \(2017\)P\. Koehn and R\. KnowlesSix challenges for neural machine translation\.InProceedings of the first workshop on neural machine translation,pp\. 28–39\.Cited by:[footnote 3](https://arxiv.org/html/2609.29511#footnote3)\.
- Krishnamurti \(2003\)B\. KrishnamurtiThe dravidian languages\.Cambridge University Press\.Cited by:[§2\.3](https://arxiv.org/html/2609.29511#S2.SS3.p2.1)\.
- Lampleet al\.\(2018\)G\. Lample, A\. Conneau, M\. Ranzato, L\. Denoyer, and H\. JégouWord Translation Without Parallel Data\.InInternational Conference on Learning Representations,Cited by:[§1](https://arxiv.org/html/2609.29511#S1.p1.1)\.
- Leeet al\.\(2022\)E\. A\. Lee, S\. Thillainathan, S\. Nayak, S\. Ranathunga, D\. I\. Adelani, R\. Su, and A\. D\. McCarthyPre\-trained multilingual sequence\-to\-sequence models: a hope for low\-resource language translation?\.InFindings of the Association for Computational Linguistics: ACL 2022,pp\. 58–67\.Cited by:[§2\.3](https://arxiv.org/html/2609.29511#S2.SS3.p5.1)\.
- Lin \(1991\)J\. LinDivergence measures based on the shannon entropy\.IEEE Transactions on Information theory37\(1\),pp\. 145–151\.Cited by:[§3\.4](https://arxiv.org/html/2609.29511#S3.SS4.p3.1)\.
- Liuet al\.\(2023\)Y\. Liu, G\. Deng, Z\. Xu, Y\. Li, Y\. Zheng, Y\. Zhang, L\. Zhao, T\. Zhang, and Y\. LiuJailbreaking chatgpt via prompt engineering: an empirical study\.ArXivabs/2305\.13860\.External Links:[Link](https://api.semanticscholar.org/CorpusID:258841501)Cited by:[Table 1](https://arxiv.org/html/2609.29511#S3.T1.p6.1.2.1.1.1)\.
- Mahaganapathyet al\.\(2026\)A\. Mahaganapathy, S\. Karunakaran, K\. Navakulan, and K\. SarveswaranBridging dialectal variation: a phonetic transcription tool for Tamil\.InProceedings of the 13th Workshop on NLP for Similar Languages, Varieties and Dialects,pp\. 234–241\.External Links:[Document](https://dx.doi.org/10.18653/v1/2026.vardial-1.19)Cited by:[§2\.3](https://arxiv.org/html/2609.29511#S2.SS3.p2.1)\.
- Nayaket al\.\(2023\)S\. Nayak, S\. Ranathunga, S\. Thillainathan, R\. Hung, A\. Rinaldi, Y\. Wang, J\. Mackey, A\. Ho, and E\. A\. LeeLeveraging auxiliary domain parallel data in intermediate task fine\-tuning for low\-resource translation\.arXiv preprint arXiv:2306\.01382\.Cited by:[§2\.2](https://arxiv.org/html/2609.29511#S2.SS2.p1.1),[§3\.2](https://arxiv.org/html/2609.29511#S3.SS2.p2.1)\.
- Nissankaet al\.\(2020\)L\. Nissanka, B\. Pushpananda, and A\. WeerasingheExploring neural machine translation for sinhala\-tamil languages pair\.In2020 20th International Conference on Advances in ICT for Emerging Regions \(ICTer\),pp\. 202–207\.Cited by:[§2\.3](https://arxiv.org/html/2609.29511#S2.SS3.p5.1)\.
- NLLB Teamet al\.\(2022\)NLLB Team, M\. R\. Costa\-jussà, J\. Cross, O\. Çelebi, M\. Elbayad, K\. Heafield, K\. Heffernan, E\. Kalbassi, J\. Lam, D\. Licht, J\. Maillard, A\. Sun, S\. Wang, G\. Wenzek, A\. Youngblood, B\. Akula, L\. Barrault, G\. M\. Gonzalez, P\. Hansanti, J\. Hoffman, S\. Jarrett, K\. R\. Sadagopan, D\. Rowe, S\. Spruit, C\. Tran, P\. Andrews, N\. F\. Ayan, S\. Bhosale, S\. Edunov, A\. Fan, C\. Gao, V\. Goswami, F\. Guzmán, P\. Koehn, A\. Mourachko, C\. Ropers, S\. Saleem, H\. Schwenk, and J\. WangNo language left behind: scaling human\-centered machine translation\.arXiv preprint arXiv:2207\.04672\.Cited by:[§1](https://arxiv.org/html/2609.29511#S1.p3.1),[§2\.3](https://arxiv.org/html/2609.29511#S2.SS3.p5.1),[§2\.3](https://arxiv.org/html/2609.29511#S2.SS3.p6.1),[§4\.1](https://arxiv.org/html/2609.29511#S4.SS1.p1.1)\.
- NLLB Teamet al\.\(2024\)NLLB Team, M\. R\. Costa\-jussà, J\. Cross, O\. Çelebi, M\. Elbayad, K\. Heafield, K\. Heffernan, E\. Kalbassi, J\. Lam, D\. Licht, J\. Maillard, A\. Sun, S\. Wang, G\. Wenzek, A\. Youngblood, B\. Akula, L\. Barrault, G\. M\. Gonzalez, P\. Hansanti, J\. Hoffman, S\. Jarrett, K\. R\. Sadagopan, D\. Rowe, S\. Spruit, C\. Tran, P\. Andrews, N\. F\. Ayan, S\. Bhosale, S\. Edunov, A\. Fan, C\. Gao, V\. Goswami, F\. Guzmán, P\. Koehn, A\. Mourachko, C\. Ropers, S\. Saleem, H\. Schwenk, and J\. WangScaling neural machine translation to 200 languages\.Nature,pp\. 1–6\.Cited by:[§1](https://arxiv.org/html/2609.29511#S1.p3.1)\.
- Ottet al\.\(2019\)M\. Ott, S\. Edunov, A\. Baevski, A\. Fan, S\. Gross, N\. Ng, D\. Grangier, and M\. AuliFairseq: a fast, extensible toolkit for sequence modeling\.InProceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics \(Demonstrations\),pp\. 48–53\.External Links:[Document](https://dx.doi.org/10.18653/v1/N19-4009)Cited by:[§1](https://arxiv.org/html/2609.29511#S1.p3.1)\.
- Panget al\.\(2025\)J\. Pang, F\. Ye, D\. F\. Wong, D\. Yu, S\. Shi, Z\. Tu, and L\. WangSalute the classic: revisiting challenges of machine translation in the age of large language models\.Transactions of the Association for Computational Linguistics13,pp\. 73–95\.Cited by:[§1](https://arxiv.org/html/2609.29511#S1.p1.1)\.
- Parliament of Democratic Socialist Republic of Sri Lanka \(2022\)Parliament of Democratic Socialist Republic of Sri LankaThe constitution of the democratic socialist republic of sri lanka\.Parliament Secretariat\.External Links:[Link](https://www.parliament.lk/files/pdf/constitution.pdf)Cited by:[§2\.3](https://arxiv.org/html/2609.29511#S2.SS3.p1.1)\.
- Pramodyaet al\.\(2020\)A\. Pramodya, R\. Pushpananda, and R\. WeerasingheA comparison of transformer, recurrent neural networks and smt in tamil to sinhala mt\.In2020 20th International Conference on Advances in ICT for Emerging Regions \(ICTer\),pp\. 155–160\.Cited by:[§2\.3](https://arxiv.org/html/2609.29511#S2.SS3.p5.1)\.
- Premjithet al\.\(2019\)B\. Premjith, M\. A\. Kumar, and K\. SomanNeural machine translation system for english to indian language translation using mtil parallel corpus\.Journal of Intelligent Systems28\(3\),pp\. 387–398\.Cited by:[§2\.1](https://arxiv.org/html/2609.29511#S2.SS1.p2.1)\.
- Pushpanandaet al\.\(2014\)R\. Pushpananda, R\. Weerasinghe, and M\. NiranjanSinhala\-tamil machine translation: towards better translation quality\.InProceedings of the Australasian Language Technology Association Workshop 2014,pp\. 129–133\.Cited by:[§2\.3](https://arxiv.org/html/2609.29511#S2.SS3.p5.1)\.
- Ranathungaet al\.\(2024a\)S\. Ranathunga, N\. de Silva, M\. Velayuthan, A\. Fernando, and C\. RathnayakeQuality does matter: a detailed look at the quality and utility of web\-mined parallel corpora\.InProceedings of the 18th Conference of the European Chapter of the Association for Computational Linguistics \(Volume 1: Long Papers\),pp\. 860–880\.External Links:[Document](https://dx.doi.org/10.18653/v1/2024.eacl-long.52)Cited by:[§2\.3](https://arxiv.org/html/2609.29511#S2.SS3.p5.1),[§2\.3](https://arxiv.org/html/2609.29511#S2.SS3.p6.1),[§3\.1](https://arxiv.org/html/2609.29511#S3.SS1.SSS0.Px1.p1.1)\.
- Ranathunga and de Silva \(2022\)S\. Ranathunga and N\. de SilvaSome languages are more equal than others: probing deeper into the linguistic disparity in the NLP world\.InProceedings of the 2nd Conference of the Asia\-Pacific Chapter of the Association for Computational Linguistics and the 12th International Joint Conference on Natural Language Processing \(Volume 1: Long Papers\),pp\. 823–848\.External Links:[Document](https://dx.doi.org/10.18653/v1/2022.aacl-main.62)Cited by:[§2\.3](https://arxiv.org/html/2609.29511#S2.SS3.p2.1)\.
- Ranathunga and De Silva \(2022\)S\. Ranathunga and N\. De SilvaSome languages are more equal than others: probing deeper into the linguistic disparity in the nlp world\.InProceedings of the 2nd Conference of the Asia\-Pacific Chapter of the Association for Computational Linguistics and the 12th International Joint Conference on Natural Language Processing \(Volume 1: Long Papers\),pp\. 823–848\.Cited by:[§2\.3](https://arxiv.org/html/2609.29511#S2.SS3.p3.1)\.
- Ranathungaet al\.\(2018\)S\. Ranathunga, F\. Farhath, U\. Thayasivam, S\. Jayasena, and G\. DiasSi\-ta: machine translation of sinhala and tamil official documents\.In2018 National Information Technology Conference \(NITC\),pp\. 1–6\.Cited by:[§3\.2](https://arxiv.org/html/2609.29511#S3.SS2.p2.1)\.
- Ranathungaet al\.\(2024b\)S\. Ranathunga, S\. Nayak, E\. Lee, X\. Peng, S\. Huang, Y\. Zeng, Y\. Mao, T\. Su, Y\. Chan, S\. Yuan,et al\.Exploiting domain\-specific parallel data on multilingual language models for low\-resource language translation\.ACM Transactions on Asian and Low\-Resource Language Information Processing\.Cited by:[§1](https://arxiv.org/html/2609.29511#S1.p1.1),[§2\.2](https://arxiv.org/html/2609.29511#S2.SS2.p1.1)\.
- Schwenket al\.\(2021a\)H\. Schwenk, V\. Chaudhary, S\. Sun, H\. Gong, and F\. GuzmánWikiMatrix: mining 135m parallel sentences in 1620 language pairs from wikipedia\.InProceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume,pp\. 1351–1361\.Cited by:[§2\.3](https://arxiv.org/html/2609.29511#S2.SS3.p6.1)\.
- Schwenket al\.\(2021b\)H\. Schwenk, G\. Wenzek, S\. Edunov, E\. Grave, A\. Joulin, and A\. FanCCMatrix: mining billions of high\-quality parallel sentences on the web\.InProceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing \(Volume 1: Long Papers\),pp\. 6490–6500\.Cited by:[§2\.1](https://arxiv.org/html/2609.29511#S2.SS1.p1.1),[§2\.3](https://arxiv.org/html/2609.29511#S2.SS3.p6.1)\.
- Singhet al\.\(2025\)P\. Singh, S\. Bhardwaj, V\. Sharma, and S\. KumarLeveraging the cross\-domain & cross\-linguistic corpus for low resource nmt: a case study on bhili\-hindi\-english parallel corpus\.arXiv preprint arXiv:2511\.00486\.Cited by:[§2\.1](https://arxiv.org/html/2609.29511#S2.SS1.p1.1),[§2\.1](https://arxiv.org/html/2609.29511#S2.SS1.p2.1)\.
- Sirisoma \(1990\)M\. H\. SirisomaBrahmi inscriptions of Sri Lanka from 3rd century BC to 65 AD\.Department of Archaeology Colombo\.Cited by:[§2\.3](https://arxiv.org/html/2609.29511#S2.SS3.p1.1)\.
- Tanget al\.\(2021\)Y\. Tang, C\. Tran, X\. Li, P\. Chen, N\. Goyal, V\. Chaudhary, J\. Gu, and A\. FanMultilingual translation from denoising pre\-training\.InFindings of the Association for Computational Linguistics: ACL\-IJCNLP 2021,pp\. 3450–3466\.Cited by:[§2\.2](https://arxiv.org/html/2609.29511#S2.SS2.p1.1)\.
- Tennageet al\.\(2018\)P\. Tennage, A\. Herath, M\. Thilakarathne, P\. Sandaruwan, and S\. RanathungaTransliteration and byte pair encoding to improve tamil to sinhala neural machine translation\.In2018 Moratuwa Engineering Research Conference \(MERCon\),pp\. 390–395\.Cited by:[§2\.3](https://arxiv.org/html/2609.29511#S2.SS3.p5.1)\.
- Tennageet al\.\(2017\)P\. Tennage, P\. Sandaruwan, M\. Thilakarathne, A\. Herath, S\. Ranathunga, S\. Jayasena, and G\. DiasNeural machine translation for sinhala and tamil languages\.In2017 International Conference on Asian Language Processing \(IALP\),pp\. 189–192\.Cited by:[§2\.3](https://arxiv.org/html/2609.29511#S2.SS3.p5.1)\.
- Thillainathanet al\.\(2021\)S\. Thillainathan, S\. Ranathunga, and S\. JayasenaFine\-tuning self\-supervised multilingual sequence\-to\-sequence models for extremely low\-resource nmt\.In2021 Moratuwa Engineering Research Conference \(MERCon\),pp\. 432–437\.Cited by:[§2\.3](https://arxiv.org/html/2609.29511#S2.SS3.p5.1)\.
- Van der Weeset al\.\(2015\)M\. Van der Wees, A\. Bisazza, W\. Weerkamp, and C\. MonzWhat’s in a domain? analyzing genre and topic differences in statistical machine translation\.InProceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing \(Volume 2: Short Papers\),pp\. 560–566\.Cited by:[footnote 3](https://arxiv.org/html/2609.29511#footnote3)\.
- Wanget al\.\(2024\)H\. Wang, P\. Minervini, and E\. PontiProbing the emergence of cross\-lingual alignment during llm training\.InFindings of the Association for Computational Linguistics: ACL 2024,pp\. 12159–12173\.Cited by:[§1](https://arxiv.org/html/2609.29511#S1.p1.1)\.
- Winataet al\.\(2023\)G\. I\. Winata, A\. F\. Aji, S\. Cahyawijaya, R\. Mahendra, F\. Koto, A\. Romadhony, K\. Kurniawan, D\. Moeljadi, R\. E\. Prasojo, P\. Fung,et al\.NusaX: multilingual parallel sentiment dataset for 10 indonesian local languages\.InProceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics,pp\. 815–834\.Cited by:[§2\.1](https://arxiv.org/html/2609.29511#S2.SS1.p1.1)\.
- Zhanget al\.\(2023\)B\. Zhang, B\. Haddow, and A\. BirchPrompting large language model for machine translation: a case study\.InProceedings of the 40th International Conference on Machine Learning \(ICML\),Cited by:[Appendix A](https://arxiv.org/html/2609.29511#A1.SS0.SSS0.Px1.p3.1)\.

## Appendix ATraining Configuration Details

Table[3](https://arxiv.org/html/2609.29511#A1.T3)lists the full LoRA fine\-tuning hyperparameters for both model families\.

In LoRA fine\-tuning, the pre\-trained weights are frozen and only low\-rank adapters injected into the attention and feed\-forward projections are trained, using the Unsloth library over Hugging Face Transformers and PEFT\. For the multimodal Gemma 3 checkpoints \(4B/12B\) the vision encoder is frozen and adapters are applied only to the language\-model layers\. All training usesbfloat16with no weight quantization, and the effective batch size is held constant across model sizes\. The one choice essential for stability was the 8\-bit AdamW optimizer\([Dettmers et al\., 2022](https://arxiv.org/html/2609.29511#bib.bib60)\): 32\-bit AdamW produced NaN gradients early in training\. Decoding is greedy \(beam size11\) with a budget of 256 new tokens; spBLEU uses the FLORES\-200 tokenizer and COMET thewmt22\-comet\-damodel\. All fine\-tuning and evaluation run on a single NVIDIA A100 80 GB GPU, a persistent VM provided free of charge under a compute grant \(PyTorch 2\.10, CUDA 12\.8, Transformers 4\.57, Unsloth, bitsandbytes\), with TF32 and the SDPA attention backend enabled; the full campaign \(355 LoRA adapters and on the order of 5,000 translate\-and\-score evaluations\) used 720 A100 GPU\-hours in total, as measured by provider telemetry, including failed and exploratory runs\.

The NLLB\-600M and Fairseq experiments were run separately, on an institutional GPU cluster

##### Prompt format\.

For the base Gemma 3 models, each training example is rendered in the Gemma chat template as a single user turn followed by the model turn:

> <start\_of\_turn\>user \{SourceLanguage\}: \{source text\} \{TargetLanguage\}:<end\_of\_turn\> <start\_of\_turn\>model \{target text\}<end\_of\_turn\>

where the language names are written out in English \(English, Sinhala, Tamil\) and the loss is masked so that only the target\-text tokens contribute to the gradient\([Zhang et al\., 2023](https://arxiv.org/html/2609.29511#bib.bib61)\)\. At inference the prompt ends after<start\_of\_turn\>modeland the continuation is decoded greedily\. TranslateGemma does not use this string prompt: it keeps its native structured translation message, a single user turn carryingsource\_lang\_codeandtarget\_lang\_codefields \(ISO 639\-1 codes: en, si, ta\) together with the source text, rendered by its own chat template with the generation prompt appended, since overwriting its instruction\- and reinforcement\-tuned prompt format degrades quality\. In both families, training and evaluation use identical prompts\.

##### Combined\-corpus training sizes\.

The multi\-domain mixture samples 1k sentence pairs from each domain’s training portion, giving 9,000 pairs for each English\-Sinhala and English\-Tamil direction and 6,509 for Sinhala\-Tamil, where only five domains \(Government,News, open\-web,OpenSubtitles,Wikipedia\) have data\. The per\-domain multilingual adapters are trained on those same five domains, with 1k pairs per direction \(6,000 per domain\); the combined multilingual \+ multi\-domain adapter is trained on all six directions of the multi\-domain mixture, 49,018 examples in total\.

The two families are studied under*different*research questions, namely the effect of model scale for Gemma 3 versus the cost of domain\-adapting an already\-translation\-tuned model for TranslateGemma, so their recipes are intentionally not matched\. Unlike the base Gemma 3 checkpoints, TranslateGemma has already undergone supervised fine\-tuning followed by reinforcement learning that optimizes translation quality against an ensemble of automatic reward models, including MetricX\-QE and AutoMQM\([Finkelstein et al\., 2026](https://arxiv.org/html/2609.29511#bib.bib57)\)\. Adapting it as aggressively as a general\-purpose model would overwrite this reinforcement\-learned policy and induce catastrophic forgetting of precisely the translation competence that makes the checkpoint valuable\. Reducing the adapter rank, learning rate, and number of epochs \(and adding a small amount of dropout\) is an established, capability\-preserving domain\-adaptation strategy\([Biderman et al\., 2024](https://arxiv.org/html/2609.29511#bib.bib58)\): low\-rank, smaller updates constrain how far the adapted policy can drift from the pre\-trained one, bounding the erosion of general translation ability so that the resulting trade\-off \(in\-domain gain versus loss of general ability\) can be*measured*rather than masked\. We consequently do not claim a controlled, like\-for\-like comparison*between*the two families: a difference in score could reflect the recipe \(rank, learning rate, epochs\) as much as the base model\. Instead, each TranslateGemma adapter is assessed against its own zero\-shot baseline, with FLORES held out as an out\-of\-domain probe of general translation ability\. This is a within\-model comparison that uses the same model and recipe on both sides and is thus unaffected by the recipe asymmetry\.

Gemma 3Translate\(1B/4B/12B\)Gemma \(4B/12B\)LoRA rankrr3216LoRAα\\alpha3216LoRA dropout0\.00\.05Adapted modulesattn \+ MLPattn \+ MLPLearning rate2×10−52\\times 10^\{\-5\}1×10−51\\times 10^\{\-5\}Epochs32Optimizer8\-bit AdamW8\-bit AdamWLR schedulecosinecosineWarmup10%10\\%10%10\\%Weight decay0\.010\.01Gradient clipping1\.01\.0Effective batch size3232Max sequence length10241024Precisionbfloat16bfloat16Random seed4242Table 3:Training hyperparameters for LoRA fine\-tuning\. The Gemma 3 recipe is held fixed across the 1B/4B/12B variants for the model\-scaling study; TranslateGemma uses a lighter recipe to limit erosion of its pre\-existing translation tuning\. The effective batch size \(32\) is kept constant across model sizes by trading per\-device batch against gradient accumulation\.

## Appendix BTraining Materials

We have provided the translators with video recordings and guideline documents to familiarize with the task and to reference during the task\. The guidleline document given forWikipediadomain English\-Sinhala tranlsators is shown in Figures[8](https://arxiv.org/html/2609.29511#A2.F8)\-[12](https://arxiv.org/html/2609.29511#A2.F12)\.

![Refer to caption](https://arxiv.org/html/2609.29511v1/EnSi-Wikipedia_Parallel_Sentences_Cleaning-1.png)Figure 8:Guideline document given to the translators for the Wikipedia English\-Sinhala task\. \- Page 1![Refer to caption](https://arxiv.org/html/2609.29511v1/EnSi-Wikipedia_Parallel_Sentences_Cleaning-2.png)Figure 9:Guideline document given to the translators for the Wikipedia English\-Sinhala task\. \- Page 2![Refer to caption](https://arxiv.org/html/2609.29511v1/EnSi-Wikipedia_Parallel_Sentences_Cleaning-3.png)Figure 10:Guideline document given to the translators for the Wikipedia English\-Sinhala task\. \- Page 3Figure 11:Guideline document given to the translators for the Wikipedia English\-Sinhala task\. \- Page 4Figure 12:Guideline document given to the translators for the Wikipedia English\-Sinhala task\. \- Page 5