ÜberWeb: Insights from Multilingual Curation for a 20-Trillion-Token Dataset

• URL: http://arxiv.org/abs/2602.15210v2
• Date: Tue, 24 Feb 2026 00:18:11 GMT
• Title: ÜberWeb: Insights from Multilingual Curation for a 20-Trillion-Token Dataset
• Authors: DatologyAI, :, Aldo Gael Carranza, Kaleigh Mentzer, Ricardo Pio Monti, Alex Fang, Alvin Deng, Amro Abbas, Anshuman Suri, Brett Larsen, Cody Blakeney, Darren Teh, David Schwab, Diego Kiner, Fan Pan, Haakon Mongstad, Haoli Yin, Jack Urbanek, Jason Lee, Jason Telanoff, Josh Wills, Luke Merrick, Maximilian Böther, Parth Doshi, Paul Burstein, Pratyush Maini, Rishabh Adaiga, Sid Joshi, Spandan Das, Tony Jiang, Vineeth Dorna, Zhengping Wang, Bogdan Gaza, Ari Morcos, Matthew Leavitt,
• Abstract summary: We study multilingual data curation across thirteen languages.<n>In controlled bilingual experiments, improving data quality for any single language benefits others.<n>We operationalize this approach within an effort that produced a 20T-token pretraining corpus.
• Score: 16.940142252787144
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: Multilinguality is a core capability for modern foundation models, yet training high-quality multilingual models remains challenging due to uneven data availability across languages. A further challenge is the performance interference that can arise from joint multilingual training, commonly referred to as the "curse of multilinguality". We study multilingual data curation across thirteen languages and find that many reported regressions are not inherent to multilingual scaling but instead stem from correctable deficiencies in data quality and composition rather than fundamental capacity limits. In controlled bilingual experiments, improving data quality for any single language benefits others: curating English improves non-English performance in 12 of 13 languages, while curating non-English yields reciprocal improvements in English. Bespoke per-language curation produces substantially larger within-language improvements. Extending these findings to large-scale general-purpose training mixtures, we show that curated multilingual allocations comprising under 8% of total tokens remain remarkably effective. We operationalize this approach within an effort that produced a 20T-token pretraining corpus derived entirely from public sources. Models with 3B and 8B parameters trained on a 1T-token random subset achieve competitive multilingual accuracy with 4-10x fewer training FLOPs than strong public baselines, establishing a new Pareto frontier in multilingual performance versus compute. Moreover, these benefits extend to frontier model scale: the 20T-token corpus served as part of the pretraining dataset for Trinity Large (400B/A13B), which exhibits strong multilingual performance relative to its training FLOPs. These results show that targeted, per-language data curation mitigates multilingual interference and enables compute-efficient multilingual scaling.

Related papers

• Revisiting Multilingual Data Mixtures in Language Model Pretraining [20.282622416939997]
  We study the impact of different multilingual data mixtures in pretraining large language models.<n>We find that combining English and multilingual data does not necessarily degrade the in-language performance of either group.<n>We do not observe a significant "curse of multilinguality" as the number of training languages increases.
  arXiv  Detail & Related papers  (2025-10-29T20:46:03Z)
• Leveraging Multilingual Training for Authorship Representation: Enhancing Generalization across Languages and Domains [41.44674318564781]
  Authorship representation (AR) learning has demonstrated strong performance in authorship attribution tasks.<n>We introduce a novel method for multilingual AR learning that incorporates two key innovations.<n>Our model is trained on over 4.5 million authors across 36 languages and 13 domains.
  arXiv  Detail & Related papers  (2025-09-20T04:43:24Z)
• Centurio: On Drivers of Multilingual Ability of Large Vision-Language Model [66.17354128553244]
  Most Large Vision-Language Models (LVLMs) to date are trained predominantly on English data.<n>We investigate how different training mixes tip the scale for different groups of languages.<n>We train Centurio, a 100-language LVLM, offering state-of-the-art performance in an evaluation covering 14 tasks and 56 languages.
  arXiv  Detail & Related papers  (2025-01-09T10:26:14Z)
• Improving In-context Learning of Multilingual Generative Language Models with Cross-lingual Alignment [42.624862172666624]
  We propose a simple yet effective cross-lingual alignment framework exploiting pairs of translation sentences. It aligns the internal sentence representations across different languages via multilingual contrastive learning. Experimental results show that even with less than 0.1 textperthousand of pre-training tokens, our alignment framework significantly boosts the cross-lingual abilities of generative language models.
  arXiv  Detail & Related papers  (2023-11-14T11:24:08Z)
• Breaking Language Barriers in Multilingual Mathematical Reasoning: Insights and Observations [59.056367787688146]
  This paper pioneers exploring and training powerful Multilingual Math Reasoning (xMR) LLMs. We construct the first multilingual math reasoning instruction dataset, MGSM8KInstruct, encompassing ten distinct languages. By utilizing translation, we construct the first multilingual math reasoning instruction dataset, MGSM8KInstruct, encompassing ten distinct languages.
  arXiv  Detail & Related papers  (2023-10-31T08:09:20Z)
• High-resource Language-specific Training for Multilingual Neural Machine Translation [109.31892935605192]
  We propose the multilingual translation model with the high-resource language-specific training (HLT-MT) to alleviate the negative interference. Specifically, we first train the multilingual model only with the high-resource pairs and select the language-specific modules at the top of the decoder. HLT-MT is further trained on all available corpora to transfer knowledge from high-resource languages to low-resource languages.
  arXiv  Detail & Related papers  (2022-07-11T14:33:13Z)
• Multilingual Transfer Learning for QA Using Translation as Data Augmentation [13.434957024596898]
  We explore strategies that improve cross-lingual transfer by bringing the multilingual embeddings closer in the semantic space. We propose two novel strategies, language adversarial training and language arbitration framework, which significantly improve the (zero-resource) cross-lingual transfer performance. Empirically, we show that the proposed models outperform the previous zero-shot baseline on the recently introduced multilingual MLQA and TyDiQA datasets.
  arXiv  Detail & Related papers  (2020-12-10T20:29:34Z)
• Cross-lingual Machine Reading Comprehension with Language Branch Knowledge Distillation [105.41167108465085]
  Cross-lingual Machine Reading (CLMRC) remains a challenging problem due to the lack of large-scale datasets in low-source languages. We propose a novel augmentation approach named Language Branch Machine Reading (LBMRC) LBMRC trains multiple machine reading comprehension (MRC) models proficient in individual language. We devise a multilingual distillation approach to amalgamate knowledge from multiple language branch models to a single model for all target languages.
  arXiv  Detail & Related papers  (2020-10-27T13:12:17Z)
• Leveraging Monolingual Data with Self-Supervision for Multilingual Neural Machine Translation [54.52971020087777]
  Using monolingual data significantly boosts the translation quality of low-resource languages in multilingual models. Self-supervision improves zero-shot translation quality in multilingual models. We get up to 33 BLEU on ro-en translation without any parallel data or back-translation.
  arXiv  Detail & Related papers  (2020-05-11T00:20:33Z)
• Improving Massively Multilingual Neural Machine Translation and Zero-Shot Translation [81.7786241489002]
  Massively multilingual models for neural machine translation (NMT) are theoretically attractive, but often underperform bilingual models and deliver poor zero-shot translations. We argue that multilingual NMT requires stronger modeling capacity to support language pairs with varying typological characteristics. We propose random online backtranslation to enforce the translation of unseen training language pairs.
  arXiv  Detail & Related papers  (2020-04-24T17:21:32Z)
• Knowledge Distillation for Multilingual Unsupervised Neural Machine Translation [61.88012735215636]
  Unsupervised neural machine translation (UNMT) has recently achieved remarkable results for several language pairs. UNMT can only translate between a single language pair and cannot produce translation results for multiple language pairs at the same time. In this paper, we empirically introduce a simple method to translate between thirteen languages using a single encoder and a single decoder.
  arXiv  Detail & Related papers  (2020-04-21T17:26:16Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.