Pith. sign in

REVIEW 4 major objections 6 minor 1 cited by

Mutarjim: Advancing Bidirectional Arabic-English Translation with a Small Language Model

T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A 1.5-billion-parameter Arabic–English translator rivals and sometimes beats models 20–30 times larger on a new balanced benchmark.

desk verdict A useful new Arabic-English benchmark and a trained 1.5B model, but the headline SOTA claim rests entirely on their own benchmark with no contamination check, and external benchmarks contradict it. read the letter →

arxiv 2505.17894 v2 pith:ZOUPCO4R submitted 2025-05-23 cs.CL cs.AI

Khalil Hennara , Muhammad Hreden , Mohamed Motaism Hamed , Zeina Aldallal , Sara Chrouf , Safwan AlModhayan This is my paper
classification cs.CLcs.AI
keywords Arabic-EnglishmachinetranslationsmalllanguagemodelbidirectionalTarjama-25benchmarktwo-phasetrainingdecoder-onlyefficientinference
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper claims that a 1.5-billion-parameter model trained specifically for Arabic–English translation can match or beat much larger general-purpose systems on a carefully designed bidirectional benchmark. The authors build Mutarjim on the Kuwain-1.5B base model, add a translation-oriented pre-training phase on roughly ten billion tokens of parallel data, then fine-tune on six million high-quality sentence pairs with Arabic-heavy sampling. They also release Tarjama-25, a 5,000-pair benchmark with long sentences, balanced directionality, and broad domain coverage, and report that Mutarjim reaches state-of-the-art or near-top scores against models up to twenty or thirty times larger, including GPT-4o mini. If accurate, this shows that a compact, task-specific model can deliver competitive translation quality at a fraction of the compute cost.

What carries the argument

The load-bearing mechanism is a two-phase training recipe on a decoder-only bilingual base. Pre-training exposes the model to 10 billion tokens of Arabic–English parallel pairs formatted with dedicated <|English|> and <|Arabic|> tokens, with sentence order randomized to prevent direction bias. Fine-tuning then uses causal masking so the model only learns to generate the target sentence from the source, over 3 billion tokens of high-quality data. The accompanying Tarjama-25 benchmark supplies the evaluation handle: 5,000 expert-reviewed pairs, 50–100 words long, half originally Arabic and half English, spread across scientific, technical, healthcare, cultural, religious, and general domains.

What would settle it

A simple overlap audit: run exact-match and near-duplicate matching between the 5,000 Tarjama-25 pairs and the pooled OPUS plus proprietary training data (or query the released data for any of the benchmark sentences). Finding even a small fraction of test sentences in training would invalidate the benchmark-based superiority claims; a clean audit would confirm them.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that continued pre-training on bilingual parallel data with a pair of language tokens, followed by fine-tuning on human-curated Arabic–English pairs, gives a 1.5B decoder-only model state-of-the-art performance on the new Tarjama-25 benchmark. Mutarjim leads the Arabic-to-English direction in all three reported metrics and leads or closely matches GPT-4o mini in English-to-Arabic, depending on the table. The authors attribute the result to two factors: an Arabic-centric training mix (a 2:1 ratio of Arabic-source to English-source samples) and a benchmark that removes the usual advantages of large models by testing long, authentic, domain-balanced sentences in both directions.

Load-bearing premise

That no Tarjama-25 test sentence also appears in Mutarjim's training corpus, since all state-of-the-art claims are measured only on this benchmark and the paper does not report an overlap check.

Editorial extensions

If this is right

  • A 1.5B model can serve as a deployable Arabic–English translator in settings where GPU memory or inference latency rules out 8B–30B models.
  • Tarjama-25 provides a reusable evaluation that stresses long, authentic, bidirectional content, potentially exposing weaknesses that short English-centric benchmarks miss.
  • The two-phase recipe of continued pre-training with language tokens plus masked fine-tuning can transfer to other language pairs and to dialectal or low-resource directions.
  • For a fixed compute budget, training a small specialist model may beat prompting a much larger general-purpose LLM for Arabic translation.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A rigorous overlap audit between Tarjama-25 and the training corpora is the most important external check; until it is run, the headline result should be read as conditional on benchmark cleanliness.
  • If benchmark cleanliness holds, the paper suggests that domain- and language-matched small models can outperform huge multilingual models, encouraging more specialist-model development for other low-resource or morphologically rich languages.
  • The reported scores are automatic-metric numbers, so a human evaluation on the same 5,000 pairs would tell whether the lead persists in fluency and adequacy, especially for English-to-Arabic where BLEU's sensitivity to Arabic morphology is known to be imperfect.
  • The paper's own ablations indicate that continued pre-training plus mixed-length fine-tuning carries most of the gain, so those two choices are the most valuable components to replicate independently.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. The paper presents Mutarjim, a 1.5B-parameter decoder-only model for bidirectional Arabic-English translation built on Kuwain-1.5B, trained with continued pre-training on roughly 10B tokens of parallel data followed by fine-tuning on about 6M pairs. It also introduces Tarjama-25, a 5,000-pair benchmark of long, domain-balanced, human-refined Arabic-English sentence pairs, and reports comparisons against larger open and proprietary models on Tarjama-25, WMT24++, and IWSLT2017 using BLEU, chrF++, and COMET. The central claim is that Mutarjim achieves state-of-the-art results on Tarjama-25 and surpasses models such as GPT-4o mini.

Significance. If the claims were supported, a 1.5B model beating much larger proprietary systems on a balanced, long-sentence bidirectional benchmark would be practically significant, and the public release of Tarjama-25 plus the evaluation toolkit is a useful contribution. The paper also deserves credit for testing directional balance, continued pre-training, and context-length effects. However, as reported, the SOTA claims are not supported by the paper's own tables, and the benchmark's independence from the training corpus is not established; the significance is therefore conditional on substantial revision.

major comments (4)
  1. [Section 7, Table 4] The sentence 'Mutarjim achieves state-of-the-art performance on the Tarjama-25 benchmark for the Arabic-to-English direction in all evaluation metrics' is contradicted by Table 4, where for Arabic-to-English the COMET score is 82.63 versus 83.67 for GPT-4o mini and 82.44 for Cohere-32B, and chrF++ is 74.66 versus 76.08 for GPT-4o mini; Mutarjim leads only in BLEU. Moreover, the abstract claims SOTA in the English-to-Arabic task, while Section 7 attributes SOTA to the Arabic-to-English direction; the directions appear to be reversed. The central claim must be corrected.
  2. [Section 3.2, Section 4.1] The fine-tuning corpus 'was designed to align with the domain categories introduced in our benchmark,' yet Section 4.1 reports no overlap or containment check between Tarjama-25's 5,000 pairs and the 10B-token pre-training corpus or the 6M-pair fine-tuning corpus. Since the Introduction explicitly claims contamination-free curation, the absence of any disjointness analysis means the Tarjama-25 results cannot be taken as evidence of generalization over memorization. This is load-bearing because the only SOTA result is measured on Tarjama-25.
  3. [Tables 5 and 6] On the established WMT24++ and IWSLT2017 benchmarks, GPT-4o mini outperforms Mutarjim on nearly all metrics; for example, WMT24++ Arabic-to-English COMET is 83.29 vs 72.99 and English-to-Arabic COMET is 82.32 vs 75.46. The abstract and Section 1 phrases such as 'outperforms models with more than 30 billion parameters, including proprietary systems like GPT-4o mini' are therefore unqualified and overstate the results; they must be restricted to the specific benchmark, direction, and metric where they hold.
  4. [Table 1 vs Table 5] WMT24++ results are inconsistent between tables. Table 1 reports Mutarjim-Bi Arabic-to-English COMET 79.73 and chrF++ 50.27, while Table 5 reports Mutarjim on the same benchmark with COMET 72.99 and chrF++ 52.27; Table 6's Arabic-to-English COMET 82.89 equals the unidirectional value in Table 1, not the bidirectional value in Table 5. The manuscript does not explain which checkpoint or evaluation configuration corresponds to each table, which undermines reproducibility.
minor comments (6)
  1. [Abstract] The abstract contains an incomplete sentence, 'including machine translation, smaller models,' and a duplicated period after 'training corpus'; these should be corrected.
  2. [Table 4] The metric header 'Chrf++' is inconsistent with 'chrF++' used elsewhere; standardize the spelling.
  3. [Figure 1] The caption says the models are sorted by size, but GPT-4o mini's parameter count is not listed; make the size ordering explicit.
  4. [Appendix B] Fine-tuning is reported as 3.3K steps with batch size 4096 and context length 512; with full context this exceeds the stated 3 billion tokens, so please clarify the effective token count.
  5. [Table 6] The NLLB row is entirely blank for IWSLT2017; state whether NLLB was not evaluated on this benchmark and why.
  6. [Appendix D] The AceGPT prompt appears garbled in the current rendering; check that the Arabic prompt displays correctly.

Circularity Check

1 steps flagged · score 6.0 of 10

The headline SOTA claim rests on the authors' own Tarjama-25 benchmark, and the fine-tuning data was explicitly designed to match that benchmark's domain categories, with no reported contamination check.

  1. fitted input called prediction [Section 3.2 and Section 7 (Table 4)]
    "The fine-tuning dataset was designed to align with the domain categories introduced in our benchmark (Section 4), ensuring broad and realistic coverage across cultural, legal, scientific, healthcare, religious, and technical domains. ... Mutarjim achieves state-of-the-art performance on the Tarjama-25 benchmark for the Arabic-to-English direction in all evaluation metrics."

    The benchmark used to support the SOTA claim is introduced in this same paper, and the paper states that the model's fine-tuning corpus was deliberately designed to match that benchmark's domain categories. The benchmark is therefore not an external, independent test of generalization: the model was optimized for the same domain distribution it is then measured on. In addition, Section 4.1's pipeline (collect sentences, MT-translate, select 5k, expert-review) contains no overlap check against the 10B-token pre-training or 6M-pair fine-tuning corpora, so the claim that Tarjama-25 is contamination-free is asserted but not demonstrated. The 'prediction' that Mutarjim leads Tarjama-25 is thus a benchmark-matched result rather than an independent first-principles prediction.

full rationale

Most of the training methodology is not circular: the two-phase training and ablations are evaluated on external WMT24++ and IWSLT2017 benchmarks, and no uniqueness theorem or ansatz is smuggled via self-citation. The base model Kuwain-1.5B is indeed the authors' own prior model, but building on one's own foundation model is standard and not circular by itself. However, the central 'surpassing GPT-4o mini' claim is supported only by Tarjama-25, a benchmark introduced in this same paper, and Section 3.2 explicitly says the fine-tuning dataset was designed to align with the benchmark's domain categories. That makes the evaluation partially constructed from training choices rather than fully independent. Moreover, the pipeline in Section 4.1 does not report any overlap check against the training corpora, despite the Introduction claiming 'careful curation to eliminate contamination from large-scale pre-training corpora.' This is a load-bearing gap. Separately, the Section 7 statement that Mutarjim leads the Arabic-to-English direction on Tarjama-25 'in all evaluation metrics' is inconsistent with Table 4, where GPT-4o mini has higher COMET and chrF++ in that direction; this is a correctness issue, not a circularity issue. Overall, the external benchmark results show a competitive but not SOTA model, while the SOTA headline reduces largely to a self-referential, domain-matched evaluation.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The model's performance claim depends on benchmark independence, which is assumed but not demonstrated; the paper's own external-benchmark results weaken the general claim.

free parameters (4)
  • Arabic-centric sample ratio in fine-tuning = 2:1 (Arabic source : English source)
    Chosen by hand in Section 3.2 to promote Arabic fluency; affects the directional quality balance.
  • Short-sample fraction in context-length experiment = 15% (samples with 2-30 words)
    Added in experiment e2 (Section 6.3) after e1 degraded on short inputs; the value was selected based on WMT24++ results.
  • Fine-tuning epochs = 2 (bidirectional), 3 (unidirectional)
    Selected in Sections 5.2 and 6.1 to balance exposure against overfitting.
  • Hyperparameters (learning rate, batch size, context length) = LR 1e-4/8e-5, batch 1024/4096, context 2048/512
    Standard hand-chosen values from Appendix B; no derivation is given.
assumptions (5)
  • domain assumption OPUS plus proprietary parallel corpora form a sufficient training distribution for Arabic-English translation.
    Section 3.1 assumes corpus quality after filtering; no evaluation of corpus composition is provided.
  • domain assumption COMET, chrF++, and BLEU scores reflect translation quality.
    Section 7 treats these metrics as authoritative without human evaluation or significance testing.
  • ad hoc to paper Randomly ordering source and target during continued pre-training produces bidirectional translation skill.
    Section 5.1 asserts this mechanism; no ablation isolates the random ordering from the parallel-data exposure.
  • ad hoc to paper Tarjama-25 sentence pairs are disjoint from Mutarjim's training data.
    Section 4.1 claims contamination curation, but the paper provides no checks (e.g., n-gram overlap) linking the benchmark to the proprietary training corpus.
  • domain assumption Kuwain-1.5B is an appropriate base model for translation.
    Section 5 builds on the authors' own prior model without external comparison to other bases.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Mutarjim: Advancing Bidirectional Arabic-English Translation with a Small Language Model." pith.science (2026). https://pith.science/paper/ZOUPCO4R

@misc{pith2026250517894,
  author       = {Pith},
  title        = {Pith review of: Mutarjim: Advancing Bidirectional Arabic-English Translation with a Small Language Model},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZOUPCO4R}},
  note         = {Machine review of arXiv:2505.17894}
}
read the original abstract

We introduce Mutarjim, a compact yet powerful language model for bidirectional Arabic-English translation. While large-scale LLMs have shown impressive progress in natural language processing tasks, including machine translation, smaller models. Leveraging this insight, we developed Mutarjim based on Kuwain-1.5B , a language model tailored for both Arabic and English. Despite its modest size, Mutarjim outperforms much larger models on several established benchmarks, achieved through an optimized two-phase training approach and a carefully curated, high-quality training corpus.. Experimental results show that Mutarjim rivals models up to 20 times larger while significantly reducing computational costs and training requirements. We also introduce Tarjama-25, a new benchmark designed to overcome limitations in existing Arabic-English benchmarking datasets, such as domain narrowness, short sentence lengths, and English-source bias. Tarjama-25 comprises 5,000 expert-reviewed sentence pairs and spans a wide range of domains, offering a more comprehensive and balanced evaluation framework. Notably, Mutarjim achieves state-of-the-art performance on the English-to-Arabic task in Tarjama-25, surpassing even significantly larger and proprietary models like GPT-4o mini. We publicly release Tarjama-25 to support future research and advance the evaluation of Arabic-English translation systems.

Figures

Figures reproduced from arXiv: 2505.17894 by the authors.

Figure 1
Figure 1. Performance of various models on Tarjama-25 sorted by model size, our newly intro [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Domain Coverage in Tarjama-25 Benchmark benchmarks, face significant challenges with Tarjama-25. Detailed evaluation results are presented in Section 7. Based on our findings, we recommend: 1. Development of language-specific authentic benchmarks; 2. Greater emphasis on domain-specific translation capability; 3. Integration of cultural and linguistic nuances in evaluation metrics; 4. Regular benchmark updates to ref… view at source ↗
Figure 3
Figure 3. Illustration of the two data formats used in Mutarjim: (Left) pre-training stream data [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. RightNow-Arabic-0.5B-Turbo: An Open Sub-1B Arabic Language Model via Vocabulary Injection and Edge-First Deployment

    cs.CL 2026-04 accept novelty 5.0 of 10

    A fully open 518M Arabic-specialized LLM, built by vocabulary injection and standard post-training on Qwen2.5-0.5B, beats same-class multilingual baselines and ships at 398 MB quantized.

Reference graph

Works this paper leans on

25 extracted references · 5 canonical work pages · cited by 1 Pith paper

  1. [1]

    Gpt-4 technical report

    Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774 ,

    arXiv

  2. [5]

    doi: 10.2139/ssrn.4312742. M Saiful Bari, Yazeed Alnumay, Norah A Alzahrani, Nouf M Alotaibi, Hisham A Alyahya, Sultan AlRashed, Faisal A Mirza, Shaykhah Z Alsubaie, Hassan A Alahmed, Ghadah Alabduljabbar, et al. Allam: Large language models for arabic and english. arXiv preprint arXiv:2407.15390 ,

    arXiv

  3. [6]

    Overview of the IWSLT 2017 evaluation cam- paign

    Mauro Cettolo, Marcello Federico, Luisa Bentivogli, Jan Niehues, Sebastian Stüker, Katsuhito Su- doh, Koichiro Yoshino, and Christian Federmann. Overview of the IWSLT 2017 evaluation cam- paign. In Proceedings of the 14th International Conference on Spoken Language Translation , pp. 2–14, Tokyo, Japan, December 14-15

    work page 2017

  4. [10]

    Kuwain 1.5B: An Arabic SLM via Language Injection

    Khalil Hennara, Sara Chrouf, Mohamed Motaism Hamed, Zeina Aldallal, Omar Hadid, and Safwan AlModhayan. Kuwain 1.5 b: An arabic slm via language injection. arXiv preprint arXiv:2504.15120,

    work page Pith review arXiv

  5. [11]

    Acegpt, localizing large language models in arabic

    Huang Huang, Fei Yu, Jianqing Zhu, Xuening Sun, Hao Cheng, Dingjie Song, Zhihong Chen, Abdulmohsen Alharthi, Bang An, Juncai He, et al. Acegpt, localizing large language models in arabic. arXiv preprint arXiv:2309.12053 ,

    arXiv

  6. [12]

    Gpt-4o system card

    14 Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Os- trow, Akila Welihinda, Alan Hayes, Alec Radford, et al. Gpt-4o system card. arXiv preprint arXiv:2410.21276,

    arXiv

  7. [14]

    Llamax: Scaling linguistic horizons of llm by enhancing translation capabilities beyond 100 languages.arXiv preprint arXiv:2407.05975,

    Yinquan Lu, Wenhao Zhu, Lei Li, Yu Qiao, and Fei Yuan. Llamax: Scaling linguistic horizons of llm by enhancing translation capabilities beyond 100 languages.arXiv preprint arXiv:2407.05975,

    arXiv

  8. [15]

    Arat5: Text- to-text transformers for arabic language generation.arXiv preprint arXiv:2109.12068 ,

    El Moatez Billah Nagoudi, AbdelRahim Elmadany, and Muhammad Abdul-Mageed. Arat5: Text- to-text transformers for arabic language generation.arXiv preprint arXiv:2109.12068 ,

    arXiv

Show all 25 references
  1. [16]

    Ricardo Rei, Craig Stewart, Ana C Farinha, and Alon Lavie

    URL https://arxiv.org/abs/1910.10683. Ricardo Rei, Craig Stewart, Ana C Farinha, and Alon Lavie. Comet: A neural framework for mt evaluation. arXiv preprint arXiv:2009.09025 ,

    1910 arXiv

  2. [17]

    Gemma 2: Improving open language models at a practical size.arXiv preprint arXiv:2408.00118 ,

    Gemma Team, Morgane Riviere, Shreya Pathak, Pier Giuseppe Sessa, Cassidy Hardin, Surya Bhu- patiraju, Léonard Hussenot, Thomas Mesnard, Bobak Shahriari, Alexandre Ramé, et al. Gemma 2: Improving open language models at a practical size.arXiv preprint arXiv:2408.00118 ,

    arXiv

  3. [18]

    Silma Team

    URLhttps://arxiv.org/abs/ 2207.04672. Silma Team. Silma

    arXiv

  4. [20]

    Prompting palm for translation: Assessing strategies and performance

    David Vilar, Markus Freitag, Colin Cherry, Jiaming Luo, Viresh Ratnakar, and George Fos- ter. Prompting palm for translation: Assessing strategies and performance. arXiv preprint arXiv:2211.09102,

    arXiv

  5. [21]

    Few-shot learning with multilingual gen- erative language models

    Victoria Lin Xi, Todor Mihaylov, Mikel Artetxe, Tianlu Wang, Shuohui Chen, Daniel Simig, Myle Ott, Naman Goyal, Shruti Bhosale, Du Jingfei, et al. Few-shot learning with multilingual gen- erative language models. InProceedings of the 2022 Conference on Empirical Methods in Nat...

    2022

  6. [22]

    A paradigm shift in ma- chine translation: Boosting translation performance of large language models

    Haoran Xu, Young Jin Kim, Amr Sharaf, and Hany Hassan Awadalla. A paradigm shift in ma- chine translation: Boosting translation performance of large language models. arXiv preprint arXiv:2309.11674,

    arXiv

  7. [23]

    X- alma: Plug & play modules and adaptive rejection for quality translation at scale.arXiv preprint arXiv:2410.03115,

    Haoran Xu, Kenton Murray, Philipp Koehn, Hieu Hoang, Akiko Eriguchi, and Huda Khayrallah. X- alma: Plug & play modules and adaptive rejection for quality translation at scale.arXiv preprint arXiv:2410.03115,

    arXiv

  8. [25]

    Multilingual machine translation with large language models: Empirical results and analysis

    Wenhao Zhu, Hongyi Liu, Qingxiu Dong, Jingjing Xu, Shujian Huang, Lingpeng Kong, Jiajun Chen, and Lei Li. Multilingual machine translation with large language models: Empirical results and analysis. arXiv preprint arXiv:2304.04675 ,

    arXiv

  9. [26]

    ةنيدملﺍ"،

    16 A Evaluation details B Traning Details We trained Mutarjim using a two-stage approach (pre-training and fine-tuning) on 8 NVIDIA H100 GPUs. Table 7 summarizes the key hyperparameters for both phases. Hyperparameter Pre-training Fine-tuning Max Learning Rate 1 × 10−4 8 × 10−...

  10. [2016]

    URLhttps://aclanthology.org/2016

    Baltic Journal of Modern Computing. URLhttps://aclanthology.org/2016. eamt-2.8. Ahmet Üstün, Viraat Aryabumi, Zheng-Xin Yong, Wei-Yin Ko, Daniel D’souza, Gbemileke Onilude, Neel Bhandari, Shivalika Singh, Hui-Lee Ooi, Amr Kayid, et al. Aya model: An instruction finetuned open-...

    2016 arXiv

  11. [2017]

    URL https://aclanthology.org/2017.iwslt-1.1

    International Workshop on Spoken Language Trans- lation. URL https://aclanthology.org/2017.iwslt-1.1. Cohere For AI. c4ai-command-r-07-arabic-2025,

    2017

  12. [2020]

    Chain- of-dictionary prompting elicits translation in large language models

    Hongyuan Lu, Haoran Yang, Haoyang Huang, Dongdong Zhang, Wai Lam, and Furu Wei. Chain- of-dictionary prompting elicits translation in large language models. arxiv e-prints, page.arXiv preprint arXiv:2305.06575,

    arXiv

  13. [2021]

    W Yang, C Li, J Zhang, and C Zong

    URL https://arxiv.org/abs/2010.11934. W Yang, C Li, J Zhang, and C Zong. Bigtranslate: Augmenting large language models with mul- tilingual translation capability over 100 languages. arxiv 2023.arXiv preprint arXiv:2305.18098 ,

    2010 arXiv

  14. [2022]

    Tower: An open multilingual large language model for translation-related tasks.arXiv preprint arXiv:2402.17733 ,

    Duarte M Alves, José Pombal, Nuno M Guerreiro, Pedro H Martins, João Alves, Amin Farajian, Ben Peters, Ricardo Rei, Patrick Fernandes, Sweta Agrawal, et al. Tower: An open multilingual large language model for translation-related tasks.arXiv preprint arXiv:2402.17733 ,

    arXiv

  15. [2023]

    In- context examples selection for machine translation.arXiv preprint arXiv:2212.02437 ,

    Sweta Agrawal, Chunting Zhou, Mike Lewis, Luke Zettlemoyer, and Marjan Ghazvininejad. In- context examples selection for machine translation.arXiv preprint arXiv:2212.02437 ,

    arXiv

  16. [2024]

    Aya23: Openweightreleases to further multilingual progress.arXiv preprint arXiv:2405.15032 ,

    13 Viraat Aryabumi, John Dang, Dwarak Talupuru, Saurabh Dash, David Cairuz, Hangyu Lin, Bharat Venkitesh, MadelineSmith, JonAnderCampos, YiChernTan, etal. Aya23: Openweightreleases to further multilingual progress.arXiv preprint arXiv:2405.15032 ,

    arXiv

  17. [2025]

    Wmt24++: Expanding the language coverage of wmt24 to 55 languages & dialects.arXiv preprint arXiv:2502.12404 ,

    Daniel Deutsch, Eleftheria Briakou, Isaac Caswell, Mara Finkelstein, Rebecca Galor, Juraj Juraska, Geza Kovacs, Alison Lui, Ricardo Rei, Jason Riesa, et al. Wmt24++: Expanding the language coverage of wmt24 to 55 languages & dialects.arXiv preprint arXiv:2502.12404 ,

    arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.