Sampo Pyysalo

Joshua Tan, Nicholas Vincent, Katherine Elkins et al.

Open source projects have made incredible progress in producing widely usable machine learning models and systems, but open source alone will face challenges in fully democratizing access to AI. Unlike previous generations of open source software, open source and open weight AI models require substantial resources to activate and maintain -- e.g., data and compute for pre-training, post-training, and deployment -- which only a few actors can currently provide. This position paper argues that open source AI must be complemented by public AI: infrastructure and institutions that ensure models are accessible, sustainable, and governed in the public interest. To achieve the full promise of AI models as prosocial public goods, we need to build public infrastructure to power and deliver open source software and models.

BigScience Workshop, Teven Le Scao, Angela Fan et al. · allen-ai, berkeley

Large language models (LLMs) have been shown to be able to perform new tasks based on a few demonstrations or natural language instructions. While these capabilities have led to widespread adoption, most LLMs are developed by resource-rich organizations and are frequently kept from the public. As a step towards democratizing this powerful technology, we present BLOOM, a 176B-parameter open-access language model designed and built thanks to a collaboration of hundreds of researchers. BLOOM is a decoder-only Transformer language model that was trained on the ROOTS corpus, a dataset comprising hundreds of sources in 46 natural and 13 programming languages (59 in total). We find that BLOOM achieves competitive performance on a wide variety of benchmarks, with stronger results after undergoing multitask prompted finetuning. To facilitate future research and applications using LLMs, we publicly release our models and code under the Responsible AI License.

Renhao Pei, Yihong Liu, Sampo Pyysalo et al.

Large language models (LLMs) offer a promising approach to machine translation (MT) for extremely low-resource languages by incorporating linguistic resources through in-context learning. However, LLMs often struggle to apply grammatical information effectively during translation. Inspired by recent progress in chain-of-thought reasoning, we investigate whether low-resource MT can benefit from structured intermediate steps of linguistic analysis and grammatical reasoning. We propose a pipeline for automatically generating step-by-step linguistic reasoning traces from Universal Dependencies treebanks, dictionaries, and grammar-rule banks. We evaluate these traces in three settings: in-context learning (ICL), supervised fine-tuning (SFT), and reinforcement fine-tuning (RFT), on Xibe and Chintang as test cases. Our results show that linguistic reasoning traces are most effective as inference-time guidance: in ICL, reliable sentence-specific traces substantially improve translation performance across most models, languages, and metrics. In contrast, using the linguistic reasoning traces as training data yields smaller and less consistent gains, as models learn the trace format but often generate erroneous content. These findings suggest that LLMs can leverage grammatical information for low-resource MT when given reliable linguistic analyses, while learning to generate such analyses remains a major bottleneck.

Stephan Oepen, Nikolay Arefev, Mikko Aulamo et al.

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 24 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 a handful of 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.

Risto Luukkonen, Ville Komulainen, Jouni Luoma et al.

Large language models (LLMs) excel in many tasks in NLP and beyond, but most open models have very limited coverage of smaller languages and LLM work tends to focus on languages where nearly unlimited data is available for pretraining. In this work, we study the challenges of creating LLMs for Finnish, a language spoken by less than 0.1% of the world population. We compile an extensive dataset of Finnish combining web crawls, news, social media and eBooks. We pursue two approaches to pretrain models: 1) we train seven monolingual models from scratch (186M to 13B parameters) dubbed FinGPT, 2) we continue the pretraining of the multilingual BLOOM model on a mix of its original training data and Finnish, resulting in a 176 billion parameter model we call BLUUMI. For model evaluation, we introduce FIN-bench, a version of BIG-bench with Finnish tasks. We also assess other model qualities such as toxicity and bias. Our models and tools are openly available at https://turkunlp.org/gpt3-finnish.

Wazir Ali, Sampo Pyysalo

Large Language Models (LLMs) have gained significant attention due to their high performance on a wide range of natural language tasks since the release of ChatGPT. The LLMs learn to understand and generate language by training billions of model parameters on vast volumes of text data. Despite being a relatively new field, LLM research is rapidly advancing in various directions. In this paper, we present an overview of LLM families, including LLaMA, PaLM, GPT, and MoE, and the methods developed to create and enhance LLMs for official European Union (EU) languages. We provide a comprehensive summary of common monolingual and multilingual datasets used for pretraining large language models.

Joona Kytöniemi, Jousia Piha, Akseli Reunamo et al.

We introduce FIN-bench-v2, a unified benchmark suite for evaluating large language models in Finnish. FIN-bench-v2 consolidates Finnish versions of widely used benchmarks together with an updated and expanded version of the original FIN-bench into a single, consistently formatted collection, covering multiple-choice and generative tasks across reading comprehension, commonsense reasoning, sentiment analysis, world knowledge, and alignment. All datasets are converted to HuggingFace Datasets, which include both cloze and multiple-choice prompt formulations with five variants per task, and we incorporate human annotation or review for machine-translated resources such as GoldenSwag and XED. To select robust tasks, we pretrain a set of 2.15B-parameter decoder-only models and use their learning curves to compute monotonicity, signal-to-noise, non-random performance, and model ordering consistency, retaining only tasks that satisfy all criteria. We further evaluate a set of larger instruction-tuned models to characterize performance across tasks and prompt formulations. All datasets, prompts, and evaluation configurations are publicly available via our fork of the Language Model Evaluation Harness at https://github.com/LumiOpen/lm-evaluation-harness. Supplementary resources are released in a separate repository at https://github.com/TurkuNLP/FIN-bench-v2.

Ona de Gibert, Graeme Nail, Nikolay Arefyev et al.

We present the HPLT (High Performance Language Technologies) language resources, a new massive multilingual dataset including both monolingual and bilingual corpora extracted from CommonCrawl and previously unused web crawls from the Internet Archive. We describe our methods for data acquisition, management and processing of large corpora, which rely on open-source software tools and high-performance computing. Our monolingual collection focuses on low- to medium-resourced languages and covers 75 languages and a total of ~5.6 trillion word tokens de-duplicated on the document level. Our English-centric parallel corpus is derived from its monolingual counterpart and covers 18 language pairs and more than 96 million aligned sentence pairs with roughly 1.4 billion English tokens. The HPLT language resources are one of the largest open text corpora ever released, providing a great resource for language modeling and machine translation training. We publicly release the corpora, the software, and the tools used in this work.

Risto Luukkonen, Jonathan Burdge, Elaine Zosa et al.

The pretraining of state-of-the-art large language models now requires trillions of words of text, which is orders of magnitude more than available for the vast majority of languages. While including text in more than one language is an obvious way to acquire more pretraining data, multilinguality is often seen as a curse, and most model training efforts continue to focus near-exclusively on individual large languages. We believe that multilinguality can be a blessing: when the lack of training data is a constraint for effectively training larger models for a target language, augmenting the dataset with other languages can offer a way to improve over the capabilities of monolingual models for that language. In this study, we introduce Poro 34B, a 34 billion parameter model trained for 1 trillion tokens of Finnish, English, and programming languages, and demonstrate that a multilingual training approach can produce a model that substantially advances over the capabilities of existing models for Finnish and excels in translation, while also achieving competitive performance in its class for English and programming languages. We release the model parameters, scripts, and data under open licenses at https://huggingface.co/LumiOpen/Poro-34B.

Akseli Reunamo, Laura-Maria Peltonen, Hans Moen et al.

This paper reports on pretraining ModernBERT encoder models in six different sizes, ranging from 51M to 475M parameters, with a focus on limited multilingualism, emphasizing languages relevant to Finland. Our models are competitive with, or superior to, existing multilingual models. They outperform monolingual models on tasks that require a context longer than 512 tokens. We present empirical results on using different data in the final stage of training. The code and models are publicly released.

Taishi Nakamura, Mayank Mishra, Simone Tedeschi et al. · ibm-research, stanford

Pretrained language models are an integral part of AI applications, but their high computational cost for training limits accessibility. Initiatives such as Bloom and StarCoder aim to democratize access to pretrained models for collaborative community development. Despite these efforts, such models encounter challenges such as limited multilingual capabilities, risks of catastrophic forgetting during continual pretraining, and the high costs of training models from scratch, alongside the need to align with AI safety standards and regulatory frameworks. This paper presents Aurora-M, a 15B parameter multilingual open-source model trained on English, Finnish, Hindi, Japanese, Vietnamese, and code. Continually pretrained from StarCoderPlus on 435B additional tokens, Aurora-M surpasses 2T tokens in total training token count. It is the first open-source multilingual model fine-tuned on human-reviewed safety instructions, thus aligning its development not only with conventional red-teaming considerations, but also with the specific concerns articulated in the Biden-Harris Executive Order on the Safe, Secure, and Trustworthy Development and Use of Artificial Intelligence. We evaluate Aurora-M across a wide range of tasks and languages, showcasing its robustness against catastrophic forgetting and its superior performance in multilingual settings, particularly in safety evaluations. We open-source Aurora-M and its variants to encourage responsible open-source development of large language models at https://huggingface.co/aurora-m.

Elaine Zosa, Ville Komulainen, Sampo Pyysalo

As LLMs gain more popularity as chatbots and general assistants, methods have been developed to enable LLMs to follow instructions and align with human preferences. These methods have found success in the field, but their effectiveness has not been demonstrated outside of high-resource languages. In this work, we discuss our experiences in post-training an LLM for instruction-following for English and Finnish. We use a multilingual LLM to translate instruction and preference datasets from English to Finnish. We perform instruction tuning and preference optimization in English and Finnish and evaluate the instruction-following capabilities of the model in both languages. Our results show that with a few hundred Finnish instruction samples we can obtain competitive performance in Finnish instruction-following. We also found that although preference optimization in English offers some cross-lingual benefits, we obtain our best results by using preference data from both languages. We release our model, datasets, and recipes under open licenses at https://huggingface.co/LumiOpen/Poro-34B-chat-OpenAssistant

Niklas Muennighoff, Alexander M. Rush, Boaz Barak et al.

The current trend of scaling language models involves increasing both parameter count and training dataset size. Extrapolating this trend suggests that training dataset size may soon be limited by the amount of text data available on the internet. Motivated by this limit, we investigate scaling language models in data-constrained regimes. Specifically, we run a large set of experiments varying the extent of data repetition and compute budget, ranging up to 900 billion training tokens and 9 billion parameter models. We find that with constrained data for a fixed compute budget, training with up to 4 epochs of repeated data yields negligible changes to loss compared to having unique data. However, with more repetition, the value of adding compute eventually decays to zero. We propose and empirically validate a scaling law for compute optimality that accounts for the decreasing value of repeated tokens and excess parameters. Finally, we experiment with approaches mitigating data scarcity, including augmenting the training dataset with code data or removing commonly used filters. Models and datasets from our 400 training runs are freely available at https://github.com/huggingface/datablations.

Li-Hsin Chang, Sampo Pyysalo, Jenna Kanerva et al.

Language models based on deep neural networks have facilitated great advances in natural language processing and understanding tasks in recent years. While models covering a large number of languages have been introduced, their multilinguality has come at a cost in terms of monolingual performance, and the best-performing models at most tasks not involving cross-lingual transfer remain monolingual. In this paper, we consider the question of whether it is possible to pre-train a bilingual model for two remotely related languages without compromising performance at either language. We collect pre-training data, create a Finnish-English bilingual BERT model and evaluate its performance on datasets used to evaluate the corresponding monolingual models. Our bilingual model performs on par with Google's original English BERT on GLUE and nearly matches the performance of monolingual Finnish BERT on a range of Finnish NLP tasks, clearly outperforming multilingual BERT. We find that when the model vocabulary size is increased, the BERT-Base architecture has sufficient capacity to learn two remotely related languages to a level where it achieves comparable performance with monolingual models, demonstrating the feasibility of training fully bilingual deep language models. The model and all tools involved in its creation are freely available at https://github.com/TurkuNLP/biBERT

Stefan Daniel Dumitrescu, Andrei-Marius Avram, Sampo Pyysalo

Large-scale pretrained language models have become ubiquitous in Natural Language Processing. However, most of these models are available either in high-resource languages, in particular English, or as multilingual models that compromise performance on individual languages for coverage. This paper introduces Romanian BERT, the first purely Romanian transformer-based language model, pretrained on a large text corpus. We discuss corpus composition and cleaning, the model training process, as well as an extensive evaluation of the model on various Romanian datasets. We open source not only the model itself, but also a repository that contains information on how to obtain the corpus, fine-tune and use this model in production (with practical examples), and how to fully replicate the evaluation process.

Sampo Pyysalo, Jenna Kanerva, Antti Virtanen et al.

Deep neural language models such as BERT have enabled substantial recent advances in many natural language processing tasks. Due to the effort and computational cost involved in their pre-training, language-specific models are typically introduced only for a small number of high-resource languages such as English. While multilingual models covering large numbers of languages are available, recent work suggests monolingual training can produce better models, and our understanding of the tradeoffs between mono- and multilingual training is incomplete. In this paper, we introduce a simple, fully automated pipeline for creating language-specific BERT models from Wikipedia data and introduce 42 new such models, most for languages up to now lacking dedicated deep neural language models. We assess the merits of these models using the state-of-the-art UDify parser on Universal Dependencies data, contrasting performance with results using the multilingual BERT model. We find that UDify using WikiBERT models outperforms the parser using mBERT on average, with the language-specific models showing substantially improved performance for some languages, yet limited improvement or a decrease in performance for others. We also present preliminary results as first steps toward an understanding of the conditions under which language-specific models are most beneficial. All of the methods and models introduced in this work are available under open licenses from https://github.com/turkunlp/wikibert.

Laurie Burchell, Ona de Gibert, Nikolay Arefyev et al.

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.

Amanda Myntti, Erik Henriksson, Veronika Laippala et al.

Pretraining data curation is a cornerstone in Large Language Model (LLM) development, leading to growing research on quality filtering of large web corpora. From statistical quality flags to LLM-based labelling systems, datasets are divided into categories, frequently reducing to a binary: those passing the filters are deemed as valuable examples, others are discarded as useless or detrimental. However, a more detailed understanding of the contribution of different kinds of texts to model performance is still largely lacking. In this article, we present the first study utilising registers or genres - a widely used standard in corpus linguistics to model linguistic variation - to curate pretraining datasets and investigate the effect of register on the performance of LLMs. We train small generative models with register classified data and evaluate them using standard benchmarks, and show that the register of pretraining data substantially affects model performance. We uncover surprising relationships between the pretraining material and the resulting models: using the News register results in subpar performance, and on the contrary, including the Opinion class, covering texts such as reviews and opinion blogs, is highly beneficial. While a model trained on the entire unfiltered dataset outperforms those trained on datasets limited to a single register, combining well-performing registers like How-to-Instructions, Informational Description, and Opinion leads to major improvements. Furthermore, analysis of individual benchmark results reveals key differences in the strengths and drawbacks of specific register classes as pretraining data. These findings show that register is an important explainer of model variation and can facilitate more deliberate future data selection practices.

Elisa Bassignana, Filip Ginter, Sampo Pyysalo et al.

Relation Extraction (RE) remains a challenging task, especially when considering realistic out-of-domain evaluations. One of the main reasons for this is the limited training size of current RE datasets: obtaining high-quality (manually annotated) data is extremely expensive and cannot realistically be repeated for each new domain. An intermediate training step on data from related tasks has shown to be beneficial across many NLP tasks.However, this setup still requires supplementary annotated data, which is often not available. In this paper, we investigate intermediate pre-training specifically for RE. We exploit the affinity between syntactic structure and semantic RE, and identify the syntactic relations which are closely related to RE by being on the shortest dependency path between two entities. We then take advantage of the high accuracy of current syntactic parsers in order to automatically obtain large amounts of low-cost pre-training data. By pre-training our RE model on the relevant syntactic relations, we are able to outperform the baseline in five out of six cross-domain setups, without any additional annotated data.

Elisa Bassignana, Filip Ginter, Sampo Pyysalo et al.

Most research in Relation Extraction (RE) involves the English language, mainly due to the lack of multi-lingual resources. We propose Multi-CrossRE, the broadest multi-lingual dataset for RE, including 26 languages in addition to English, and covering six text domains. Multi-CrossRE is a machine translated version of CrossRE (Bassignana and Plank, 2022), with a sub-portion including more than 200 sentences in seven diverse languages checked by native speakers. We run a baseline model over the 26 new datasets and--as sanity check--over the 26 back-translations to English. Results on the back-translated data are consistent with the ones on the original English CrossRE, indicating high quality of the translation and the resulting dataset.

Samuel Rönnqvist, Amanda Myntti, Aki-Juhani Kyröläinen et al.

In recent years, several methods have been proposed for explaining individual predictions of deep learning models, yet there has been little study of how to aggregate these predictions to explain how such models view classes as a whole in text classification tasks. In this work, we propose a method for explaining classes using deep learning models and the Integrated Gradients feature attribution technique by aggregating explanations of individual examples in text classification to general descriptions of the classes. We demonstrate the approach on Web register (genre) classification using the XML-R model and the Corpus of Online Registers of English (CORE), finding that the method identifies plausible and discriminative keywords characterizing all but the smallest class.

Li-Hsin Chang, Sampo Pyysalo, Jenna Kanerva et al.

In this paper, we present a quantitative evaluation of differences between alternative translations in a large recently released Finnish paraphrase corpus focusing in particular on non-trivial variation in translation. We combine a series of automatic steps detecting systematic variation with manual analysis to reveal regularities and identify categories of translation differences. We find the paraphrase corpus to contain highly non-trivial translation variants difficult to recognize through automatic approaches.

Li-Hsin Chang, Iiro Rastas, Sampo Pyysalo et al.

Essays as a form of assessment test student knowledge on a deeper level than short answer and multiple-choice questions. However, the manual evaluation of essays is time- and labor-consuming. Automatic clustering of essays, or their fragments, prior to manual evaluation presents a possible solution to reducing the effort required in the evaluation process. Such clustering presents numerous challenges due to the variability and ambiguity of natural language. In this paper, we introduce two datasets of undergraduate student essays in Finnish, manually annotated for salient arguments on the sentence level. Using these datasets, we evaluate several deep-learning embedding methods for their suitability to sentence clustering in support of essay grading. We find that the choice of the most suitable method depends on the nature of the exam question and the answers, with deep-learning methods being capable of, but not guaranteeing better performance over simpler methods based on lexical overlap.

Liina Repo, Valtteri Skantsi, Samuel Rönnqvist et al.

We explore cross-lingual transfer of register classification for web documents. Registers, that is, text varieties such as blogs or news are one of the primary predictors of linguistic variation and thus affect the automatic processing of language. We introduce two new register annotated corpora, FreCORE and SweCORE, for French and Swedish. We demonstrate that deep pre-trained language models perform strongly in these languages and outperform previous state-of-the-art in English and Finnish. Specifically, we show 1) that zero-shot cross-lingual transfer from the large English CORE corpus can match or surpass previously published monolingual models, and 2) that lightweight monolingual classification requiring very little training data can reach or surpass our zero-shot performance. We further analyse classification results finding that certain registers continue to pose challenges in particular for cross-lingual transfer.

Jouni Luoma, Sampo Pyysalo

Named entity recognition (NER) is frequently addressed as a sequence classification task where each input consists of one sentence of text. It is nevertheless clear that useful information for the task can often be found outside of the scope of a single-sentence context. Recently proposed self-attention models such as BERT can both efficiently capture long-distance relationships in input as well as represent inputs consisting of several sentences, creating new opportunitites for approaches that incorporate cross-sentence information in natural language processing tasks. In this paper, we present a systematic study exploring the use of cross-sentence information for NER using BERT models in five languages. We find that adding context in the form of additional sentences to BERT input systematically increases NER performance on all of the tested languages and models. Including multiple sentences in each input also allows us to study the predictions of the same sentences in different contexts. We propose a straightforward method, Contextual Majority Voting (CMV), to combine different predictions for sentences and demonstrate this to further increase NER performance with BERT. Our approach does not require any changes to the underlying BERT architecture, rather relying on restructuring examples for training and prediction. Evaluation on established datasets, including the CoNLL'02 and CoNLL'03 NER benchmarks, demonstrates that our proposed approach can improve on the state-of-the-art NER results on English, Dutch, and Finnish, achieves the best reported BERT-based results on German, and is on par with performance reported with other BERT-based approaches in Spanish. We release all methods implemented in this work under open licenses.

Joakim Nivre, Marie-Catherine de Marneffe, Filip Ginter et al.

Universal Dependencies is an open community effort to create cross-linguistically consistent treebank annotation for many languages within a dependency-based lexicalist framework. The annotation consists in a linguistically motivated word segmentation; a morphological layer comprising lemmas, universal part-of-speech tags, and standardized morphological features; and a syntactic layer focusing on syntactic relations between predicates, arguments and modifiers. In this paper, we describe version 2 of the guidelines (UD v2), discuss the major changes from UD v1 to UD v2, and give an overview of the currently available treebanks for 90 languages.

Antti Virtanen, Jenna Kanerva, Rami Ilo et al.

Deep learning-based language models pretrained on large unannotated text corpora have been demonstrated to allow efficient transfer learning for natural language processing, with recent approaches such as the transformer-based BERT model advancing the state of the art across a variety of tasks. While most work on these models has focused on high-resource languages, in particular English, a number of recent efforts have introduced multilingual models that can be fine-tuned to address tasks in a large number of different languages. However, we still lack a thorough understanding of the capabilities of these models, in particular for lower-resourced languages. In this paper, we focus on Finnish and thoroughly evaluate the multilingual BERT model on a range of tasks, comparing it with a new Finnish BERT model trained from scratch. The new language-specific model is shown to systematically and clearly outperform the multilingual. While the multilingual model largely fails to reach the performance of previously proposed methods, the custom Finnish BERT model establishes new state-of-the-art results on all corpora for all reference tasks: part-of-speech tagging, named entity recognition, and dependency parsing. We release the model and all related resources created for this study with open licenses at https://turkunlp.org/finbert .

Marek Rei, Gamal K. O. Crichton, Sampo Pyysalo

Sequence labeling architectures use word embeddings for capturing similarity, but suffer when handling previously unseen or rare words. We investigate character-level extensions to such models and propose a novel architecture for combining alternative word representations. By using an attention mechanism, the model is able to dynamically decide how much information to use from a word- or character-level component. We evaluated different architectures on a range of sequence labeling datasets, and character-level extensions were found to improve performance on every benchmark. In addition, the proposed attention-based architecture delivered the best results even with a smaller number of trainable parameters.