We present an ongoing initiative to provide open, very large, high-quality, and richly annotated textual datasets for almost 200 languages. At 30 trillion tokens, this is likely the largest generally available multilingual collection of LLM pre-training data. These datasets are derived from web crawls from different sources and accompanied with a complete, open-source pipeline for document selection from web archives, text extraction from HTML, language identification for noisy texts, exact and near-deduplication, annotation with, among others, register labels, text quality estimates, and personally identifiable information; and final selection and filtering. We report on data quality probes through contrastive and analytical statistics, through manual inspection of samples for some 20 languages, and through end-to-end evaluation of various language model architectures trained on this data. For multilingual LLM evaluation, we provide a comprehensive collection of benchmarks for nine European languages, with special emphasis on natively created tasks, mechanisms to mitigate prompt sensitivity, and refined normalization and aggregation of scores. Additionally, we train and evaluate a family of 57 monolingual encoder–decoder models, as well as about 30 “smallish” monolingual GPT-like reference models. Besides the monolingual data and models, we also present a very large collection of parallel texts automatically mined from this data, together with a novel parallel corpus synthesized via machine translation.

The development of high-quality multilingual models depends on well-curated multilingual datasets, which are typically obtained through filtering large web-crawled corpora. Such filtering approaches must be effective across a wide range of languages in order to improve model quality proportionally across the entire linguistic space. In recent years multiple Natural Language Processing (NLP) approaches for data processing and Large Language Model (LLM) training were developed. However, the most significant progress has been made for rich-resource languages such as English and Chinese, leaving multilingual methods largely underexplored. In this work, we present data selection strategies based on Internet registers (genres) and Web Document Score (WDS). We study 15 European languages utilizing HPLT 3.0 datasets. We train and evaluate four small generative language models per language, and compare how specific registers and WDS affect model training in separate and combined approaches. We use five common evaluation tasks covering models’ linguistic, knowledge-based, and reasoning skills. We analyze the impact of the studied strategies on each language, their average performance, and show how combining register-based selection with WDS improves evaluation performance. These results indicate that both registers and WDS can generalize across multiple languages, offering a new path towards multilingual data filtering.

We describe the progress of the High Performance Language Technologies (HPLT) project, a 3-year EU-funded project that started in September 2022. We focus on the up-to-date results on the release of free text datasets derived from web crawls, one of the central objectives of the project. The second release used a revised processing pipeline, and an enlarged set of input crawls. From 4.5 petabytes of web crawls we extracted 7.6T tokens of monolingual text in 193 languages, plus 380 million parallel sentences in 51 language pairs. We also release MultiHPLT, a cross-combination of the parallel data, which produces 1,275 pairs, as well as releasing the containing documents for all parallel sentences in order to enable research in document-level MT. We report changes in the pipeline, analysis and evaluation results for the second parallel data release based on machine translation systems. All datasets are released under a permissive CC0 licence.

Training state-of-the-art large language models requires vast amounts of clean and diverse textual data. However, building suitable multilingual datasets remains a challenge. In this work, we present HPLT v2, a collection of high-quality multilingual monolingual and parallel corpora, extending prior work of the HPLT project. The monolingual portion of the data contains 8T tokens covering 193 languages, while the parallel data contains 380M sentence pairs covering 51 languages. We document the entire data pipeline and release the code to reproduce it. We provide extensive analysis of the quality and characteristics of our data. Finally, we evaluate the performance of language models and machine translation systems trained on HPLT v2, demonstrating its value.

Web-scale corpora present valuable research opportunities but often lack detailed metadata, making them challenging to use in linguistics and social sciences. This study tackles this problem by exploring automatic methods to classify web corpora into specific categories, focusing on text registers such as Interactive Discussion and literary genres such as Politics and Social Sciences. We train two machine learning models to classify documents from the large web-crawled OSCAR dataset: a register classifier using the multilingual, manually annotated CORE corpus, and a genre classifier using a dataset based on Kindle US&UK. Fine-tuned from XLM-R Large, the register and genre classifiers achieved F1-scores of 0.74 and 0.70, respectively. Our analysis includes evaluating the distribution of the predicted text classes and examining the intersection of genre-register pairs using topic modelling. The results show expected combinations between certain registers and genres, such as the Lyrical register often aligning with the Literature & Fiction genre. However, most registers, such as Interactive Discussion, are divided across multiple genres, like Engineering & Transportation and Politics & Social Sciences, depending on the discussion topic. This enriched metadata provides valuable insights and supports new ways of studying digital cultural heritage.

In corpus linguistics, registers–language varieties suited to different contexts–have traditionally been defined by their situations of use, yet recent studies reveal significant situational variation within registers. Previous quantitative studies, however, have been limited to English, leaving this variation in other languages largely unexplored. To address this gap, we apply a quantitative situational analysis to a large multilingual web register corpus, using large language models (LLMs) to annotate texts in English, Finnish, French, Swedish, and Turkish for 23 situational parameters. Using clustering techniques, we identify six situational text types, such as “Advice”, “Opinion” and “Marketing”, each characterized by distinct situational features. We explore the relationship between these text types and traditional register categories, finding partial alignment, though no register maps perfectly onto a single cluster. These results support the quantitative approach to situational analysis and are consistent with earlier findings for English. Cross-linguistic comparisons show that language accounts for only a small part of situational variation within registers, suggesting registers are situationally similar across languages. This study demonstrates the utility of LLMs in multilingual register analysis and deepens our understanding of situational variation within registers.

This article introduces a resource-efficient method for developing question-answer (QA) datasets by extracting QA pairs from web-scale data using machine learning (ML). Our method benefits from recent advances in web register (genre) identification and consists of two ML steps with an additional post-processing step. First, using XLM-R and the multilingual CORE web register corpus series with categories such as QA Forum, we train a multilingual classifier to retrieve documents that are likely to contain QA pairs from web-scale data. Second, we develop a NER-style token classifier to identify the QA text spans within these documents. To this end, we experiment with training on a semi-synthetic dataset built on top of the English LFQA, a small set of manually cleaned web QA pairs in English and Finnish, and a Finnish web QA pair dataset cleaned using ChatGPT. The evaluation of our pipeline demonstrates its capability to efficiently retrieve a substantial volume of QA pairs. While the approach is adaptable to any language given the availability of language models and extensive web data, we showcase its efficiency in English and Finnish, developing the first open, non-synthetic and non-machine translated QA dataset for Finnish – Turku WebQA – comprising over 200,000 QA pairs.

Input saliency methods have recently become a popular tool for explaining predictions of deep learning models in NLP. Nevertheless, there has been little work investigating methods for aggregating prediction-level explanations to the class level, nor has a framework for evaluating such class explanations been established. We explore explanations based on XLM-R and the Integrated Gradients input attribution method, and propose 1) the Stable Attribution Class Explanation method (SACX) to extract keyword lists of classes in text classification tasks, and 2) a framework for the systematic evaluation of the keyword lists. We find that explanations of individual predictions are prone to noise, but that stable explanations can be effectively identified through repeated training and explanation. We evaluate on web register data and show that the class explanations are linguistically meaningful and distinguishing of the classes.