José Pombal

Pedro Henrique Martins, Patrick Fernandes, João Alves et al. · meta-ai

The quality of open-weight LLMs has seen significant improvement, yet they remain predominantly focused on English. In this paper, we introduce the EuroLLM project, aimed at developing a suite of open-weight multilingual LLMs capable of understanding and generating text in all official European Union languages, as well as several additional relevant languages. We outline the progress made to date, detailing our data collection and filtering process, the development of scaling laws, the creation of our multilingual tokenizer, and the data mix and modeling configurations. Additionally, we release our initial models: EuroLLM-1.7B and EuroLLM-1.7B-Instruct and report their performance on multilingual general benchmarks and machine translation.

Ricardo Rei, Nuno M. Guerreiro, José Pombal et al.

We present the joint contribution of Unbabel and Instituto Superior Técnico to the WMT 2023 Shared Task on Quality Estimation (QE). Our team participated on all tasks: sentence- and word-level quality prediction (task 1) and fine-grained error span detection (task 2). For all tasks, we build on the COMETKIWI-22 model (Rei et al., 2022b). Our multilingual approaches are ranked first for all tasks, reaching state-of-the-art performance for quality estimation at word-, span- and sentence-level granularity. Compared to the previous state-of-the-art COMETKIWI-22, we show large improvements in correlation with human judgements (up to 10 Spearman points). Moreover, we surpass the second-best multilingual submission to the shared-task with up to 3.8 absolute points.

José Pombal, André F. Cruz, João Bravo et al.

In recent years, machine learning algorithms have become ubiquitous in a multitude of high-stakes decision-making applications. The unparalleled ability of machine learning algorithms to learn patterns from data also enables them to incorporate biases embedded within. A biased model can then make decisions that disproportionately harm certain groups in society -- limiting their access to financial services, for example. The awareness of this problem has given rise to the field of Fair ML, which focuses on studying, measuring, and mitigating unfairness in algorithmic prediction, with respect to a set of protected groups (e.g., race or gender). However, the underlying causes for algorithmic unfairness still remain elusive, with researchers divided between blaming either the ML algorithms or the data they are trained on. In this work, we maintain that algorithmic unfairness stems from interactions between models and biases in the data, rather than from isolated contributions of either of them. To this end, we propose a taxonomy to characterize data bias and we study a set of hypotheses regarding the fairness-accuracy trade-offs that fairness-blind ML algorithms exhibit under different data bias settings. On our real-world account-opening fraud use case, we find that each setting entails specific trade-offs, affecting fairness in expected value and variance -- the latter often going unnoticed. Moreover, we show how algorithms compare differently in terms of accuracy and fairness, depending on the biases affecting the data. Finally, we note that under specific data bias conditions, simple pre-processing interventions can successfully balance group-wise error rates, while the same techniques fail in more complex settings.

Sérgio Jesus, José Pombal, Duarte Alves et al.

Evaluating new techniques on realistic datasets plays a crucial role in the development of ML research and its broader adoption by practitioners. In recent years, there has been a significant increase of publicly available unstructured data resources for computer vision and NLP tasks. However, tabular data -- which is prevalent in many high-stakes domains -- has been lagging behind. To bridge this gap, we present Bank Account Fraud (BAF), the first publicly available privacy-preserving, large-scale, realistic suite of tabular datasets. The suite was generated by applying state-of-the-art tabular data generation techniques on an anonymized,real-world bank account opening fraud detection dataset. This setting carries a set of challenges that are commonplace in real-world applications, including temporal dynamics and significant class imbalance. Additionally, to allow practitioners to stress test both performance and fairness of ML methods, each dataset variant of BAF contains specific types of data bias. With this resource, we aim to provide the research community with a more realistic, complete, and robust test bed to evaluate novel and existing methods.

José Pombal, Pedro Saleiro, Mário A. T. Figueiredo et al.

The unparalleled ability of machine learning algorithms to learn patterns from data also enables them to incorporate biases embedded within. A biased model can then make decisions that disproportionately harm certain groups in society. Much work has been devoted to measuring unfairness in static ML environments, but not in dynamic, performative prediction ones, in which most real-world use cases operate. In the latter, the predictive model itself plays a pivotal role in shaping the distribution of the data. However, little attention has been heeded to relating unfairness to these interactions. Thus, to further the understanding of unfairness in these settings, we propose a taxonomy to characterize bias in the data, and study cases where it is shaped by model behaviour. Using a real-world account opening fraud detection case study as an example, we study the dangers to both performance and fairness of two typical biases in performative prediction: distribution shifts, and the problem of selective labels.

Duarte M. Alves, Nuno M. Guerreiro, João Alves et al.

Large language models (LLMs) are a promising avenue for machine translation (MT). However, current LLM-based MT systems are brittle: their effectiveness highly depends on the choice of few-shot examples and they often require extra post-processing due to overgeneration. Alternatives such as finetuning on translation instructions are computationally expensive and may weaken in-context learning capabilities, due to overspecialization. In this paper, we provide a closer look at this problem. We start by showing that adapter-based finetuning with LoRA matches the performance of traditional finetuning while reducing the number of training parameters by a factor of 50. This method also outperforms few-shot prompting and eliminates the need for post-processing or in-context examples. However, we show that finetuning generally degrades few-shot performance, hindering adaptation capabilities. Finally, to obtain the best of both worlds, we propose a simple approach that incorporates few-shot examples during finetuning. Experiments on 10 language pairs show that our proposed approach recovers the original few-shot capabilities while keeping the added benefits of finetuning.

José Pombal, Pedro Saleiro, Mário A. T. Figueiredo et al.

Data valuation is a ML field that studies the value of training instances towards a given predictive task. Although data bias is one of the main sources of downstream model unfairness, previous work in data valuation does not consider how training instances may influence both performance and fairness of ML models. Thus, we propose Fairness-Aware Data vauatiOn (FADO), a data valuation framework that can be used to incorporate fairness concerns into a series of ML-related tasks (e.g., data pre-processing, exploratory data analysis, active learning). We propose an entropy-based data valuation metric suited to address our two-pronged goal of maximizing both performance and fairness, which is more computationally efficient than existing metrics. We then show how FADO can be applied as the basis for unfairness mitigation pre-processing techniques. Our methods achieve promising results -- up to a 40 p.p. improvement in fairness at a less than 1 p.p. loss in performance compared to a baseline -- and promote fairness in a data-centric way, where a deeper understanding of data quality takes center stage.

Miguel Moura Ramos, Duarte M. Alves, Hippolyte Gisserot-Boukhlef et al.

This report presents EuroLLM-22B, a large language model trained from scratch to support the needs of European citizens by covering all 24 official European Union languages and 11 additional languages. EuroLLM addresses the issue of European languages being underrepresented and underserved in existing open large language models. We provide a comprehensive overview of EuroLLM-22B's development, including tokenizer design, architectural specifications, data filtering, and training procedures. Across a broad set of multilingual benchmarks, EuroLLM-22B demonstrates strong performance in reasoning, instruction following, and translation, achieving results competitive with models of comparable size. To support future research, we release our base and instruction-tuned models, our multilingual web pretraining data and updated EuroBlocks instruction datasets, as well as our pre-training and evaluation codebases.

Duarte M. Alves, José Pombal, Nuno M. Guerreiro et al.

While general-purpose large language models (LLMs) demonstrate proficiency on multiple tasks within the domain of translation, approaches based on open LLMs are competitive only when specializing on a single task. In this paper, we propose a recipe for tailoring LLMs to multiple tasks present in translation workflows. We perform continued pretraining on a multilingual mixture of monolingual and parallel data, creating TowerBase, followed by finetuning on instructions relevant for translation processes, creating TowerInstruct. Our final model surpasses open alternatives on several tasks relevant to translation workflows and is competitive with general-purpose closed LLMs. To facilitate future research, we release the Tower models, our specialization dataset, an evaluation framework for LLMs focusing on the translation ecosystem, and a collection of model generations, including ours, on our benchmark.

Beatriz Canaverde, Duarte M. Alves, José Pombal et al.

In a conversation, a helpful assistant must reliably follow user directives, even as they refine, modify, or contradict earlier requests. Yet most instruction-following benchmarks focus on single-turn or short multi-turn scenarios, leaving open how well models handle long-horizon instruction-following tasks. To bridge this gap, we present SEQUOR, an automatic benchmark for evaluating constraint adherence in long multi-turn conversations. SEQUOR consists of simulated persona-driven interactions built with constraints extracted from real-world conversations. Our results show that even when following a single constraint, instruction-following accuracy consistently decreases as the conversation grows longer, with drops exceeding 11%. This decline becomes larger when models have to follow multiple constraints simultaneously, reducing their accuracy by over 40%. In scenarios where constraints are added or replaced at arbitrary points of the conversation, model accuracy decreases by more than 9%. Taken together, our results reveal that current models still struggle to follow user instructions in multi-turn conversations, and provide a way for better measuring instruction-following capabilities in assistants.

José Pombal, Dongkeun Yoon, Patrick Fernandes et al. · cmu

The use of language models for automatically evaluating long-form text (LLM-as-a-judge) is becoming increasingly common, yet most LLM judges are optimized exclusively for English, with strategies for enhancing their multilingual evaluation capabilities remaining largely unexplored in the current literature. This has created a disparity in the quality of automatic evaluation methods for non-English languages, ultimately hindering the development of models with better multilingual capabilities. To bridge this gap, we introduce M-Prometheus, a suite of open-weight LLM judges ranging from 3B to 14B parameters that can provide both direct assessment and pairwise comparison feedback on multilingual outputs. M-Prometheus models outperform state-of-the-art open LLM judges on multilingual reward benchmarks spanning more than 20 languages, as well as on literary machine translation (MT) evaluation covering 4 language pairs. Furthermore, M-Prometheus models can be leveraged at decoding time to significantly improve generated outputs across all 3 tested languages, showcasing their utility for the development of better multilingual models. Lastly, through extensive ablations, we identify the key factors for obtaining an effective multilingual judge, including backbone model selection and training on synthetic multilingual feedback data instead of translated data. We release our models, training dataset, and code.

Ricardo Rei, Nuno M. Guerreiro, José Pombal et al.

Fine-tuning pretrained LLMs has been shown to be an effective strategy for reaching state-of-the-art performance on specific tasks like machine translation. However, this process of adaptation often implies sacrificing general-purpose capabilities, such as conversational reasoning and instruction-following, hampering the utility of the system in real-world applications that require a mixture of skills. In this paper, we introduce Tower+, a suite of models designed to deliver strong performance across both translation and multilingual general-purpose text capabilities. We achieve a Pareto frontier between translation specialization and multilingual general-purpose capabilities by introducing a novel training recipe that builds on Tower (Alves et al., 2024), comprising continued pretraining, supervised fine-tuning, preference optimization, and reinforcement learning with verifiable rewards. At each stage of training, we carefully generate and curate data to strengthen performance on translation as well as general-purpose tasks involving code generation, mathematics problem solving, and general instruction-following. We develop models at multiple scales: 2B, 9B, and 72B. Our smaller models often outperform larger general-purpose open-weight and proprietary LLMs (e.g., Llama 3.3 70B, GPT-4o). Our largest model delivers best-in-class translation performance for high-resource languages and top results in multilingual Arena Hard evaluations and in IF-MT, a benchmark we introduce for evaluating both translation and instruction-following. Our findings highlight that it is possible to rival frontier models in general capabilities, while optimizing for specific business domains, such as translation and localization.

Pedro Henrique Martins, João Alves, Patrick Fernandes et al. · meta-ai

This report presents EuroLLM-9B, a large language model trained from scratch to support the needs of European citizens by covering all 24 official European Union languages and 11 additional languages. EuroLLM addresses the issue of European languages being underrepresented and underserved in existing open large language models. We provide a comprehensive overview of EuroLLM-9B's development, including tokenizer design, architectural specifications, data filtering, and training procedures. We describe the pre-training data collection and filtering pipeline, including the creation of EuroFilter, an AI-based multilingual filter, as well as the design of EuroBlocks-Synthetic, a novel synthetic dataset for post-training that enhances language coverage for European languages. Evaluation results demonstrate EuroLLM-9B's competitive performance on multilingual benchmarks and machine translation tasks, establishing it as the leading open European-made LLM of its size. To support open research and adoption, we release all major components of this work, including the base and instruction-tuned models, the EuroFilter classifier, and the synthetic post-training dataset.

José Pombal, Nuno M. Guerreiro, Ricardo Rei et al.

As language models improve and become capable of performing more complex tasks across modalities, evaluating them automatically becomes increasingly challenging. Developing strong and robust task-specific automatic metrics gets harder, and human-annotated test sets -- which are expensive to create -- saturate more quickly. A compelling alternative is to design reliable strategies to automate the creation of test data and evaluation, but previous attempts either rely on pre-existing data, or focus solely on individual tasks. We present Zero-shot Benchmarking (ZSB), a framework for creating high-quality benchmarks for any task by leveraging language models for both synthetic test data creation and evaluation. ZSB is simple and flexible: it requires only the creation of a prompt for data generation and one for evaluation; it is scalable to tasks and languages where collecting real-world data is costly or impractical; it is model-agnostic, allowing the creation of increasingly challenging benchmarks as models improve. To assess the effectiveness of our framework, we create benchmarks for five text-only tasks and a multi-modal one: general capabilities in four languages (English, Chinese, French, and Korean), translation, and general vision-language capabilities in English. We then rank a broad range of open and closed systems on our benchmarks. ZSB rankings consistently correlate strongly with human rankings, outperforming widely-adopted standard benchmarks. Through ablations, we find that strong benchmarks can be created with open models, and that judge model size and dataset variety are crucial drivers of performance. We release all our benchmarks, and code to reproduce our experiments and to produce new benchmarks.

José Pombal, Nuno M. Guerreiro, Ricardo Rei et al.

As automatic metrics become increasingly stronger and widely adopted, the risk of unintentionally "gaming the metric" during model development rises. This issue is caused by metric interference (MINT), i.e., the use of the same or related metrics for both model tuning and evaluation. MINT can misguide practitioners into being overoptimistic about the performance of their systems: as system outputs become a function of the interfering metric, their estimated quality loses correlation with human judgments. In this work, we analyze two common cases of MINT in machine translation-related tasks: filtering of training data, and decoding with quality signals. Importantly, we find that MINT strongly distorts instance-level metric scores, even when metrics are not directly optimized for-questioning the common strategy of leveraging a different, yet related metric for evaluation that is not used for tuning. To address this problem, we propose MINTADJUST, a method for more reliable evaluation under MINT. On the WMT24 MT shared task test set, MINTADJUST ranks translations and systems more accurately than state-of-the-art metrics across a majority of language pairs, especially for high-quality systems. Furthermore, MINTADJUST outperforms AUTORANK, the ensembling method used by the organizers.

José Pombal, Sweta Agrawal, Patrick Fernandes et al.

Automatic translation systems offer a powerful solution to bridge language barriers in scenarios where participants do not share a common language. However, these systems can introduce errors leading to misunderstandings and conversation breakdown. A key issue is that current systems fail to incorporate the rich contextual information necessary to resolve ambiguities and omitted details, resulting in literal, inappropriate, or misaligned translations. In this work, we present a framework to improve large language model-based translation systems by incorporating contextual information in bilingual conversational settings during training and inference. We validate our proposed framework on two task-oriented domains: customer chat and user-assistant interaction. Across both settings, the system produced by our framework-TowerChat-consistently results in better translations than state-of-the-art systems like GPT-4o and TowerInstruct, as measured by multiple automatic translation quality metrics on several language pairs. We also show that the resulting model leverages context in an intended and interpretable way, improving consistency between the conveyed message and the generated translations.

António Farinhas, Nuno M. Guerreiro, José Pombal et al.

Large language models are increasingly used for mental health support, yet their conversational coherence alone does not ensure clinical appropriateness. Existing general-purpose safeguards often fail to distinguish between therapeutic disclosures and genuine clinical crises, leading to safety failures. To address this gap, we introduce a clinically grounded risk taxonomy, developed in collaboration with PhD-level psychologists, that identifies actionable harm (e.g., self-harm and harm to others) while preserving space for safe, non-crisis therapeutic content. We release MindGuard-testset, a dataset of real-world multi-turn conversations annotated at the turn level by clinical experts. Using synthetic dialogues generated via a controlled two-agent setup, we train MindGuard, a family of lightweight safety classifiers (with 4B and 8B parameters). Our classifiers reduce false positives at high-recall operating points and, when paired with clinician language models, help achieve lower attack success and harmful engagement rates in adversarial multi-turn interactions compared to general-purpose safeguards. We release all models and human evaluation data.

José Pombal, Maya D'Eon, Nuno M. Guerreiro et al.

Demand for mental health support through AI chatbots is surging, though current systems present several limitations, like sycophancy or overvalidation, and reinforcement of maladaptive beliefs. A core obstacle to the creation of better systems is the scarcity of benchmarks that capture the complexity of real therapeutic interactions. Most existing benchmarks either only test clinical knowledge through multiple-choice questions or assess single responses in isolation. To bridge this gap, we present MindEval, a framework designed in collaboration with Ph.D-level Licensed Clinical Psychologists for automatically evaluating language models in realistic, multi-turn mental health therapy conversations. Through patient simulation and automatic evaluation with LLMs, our framework balances resistance to gaming with reproducibility via its fully automated, model-agnostic design. We begin by quantitatively validating the realism of our simulated patients against human-generated text and by demonstrating strong correlations between automatic and human expert judgments. Then, we evaluate 12 state-of-the-art LLMs and show that all models struggle, scoring below 4 out of 6, on average, with particular weaknesses in problematic AI-specific patterns of communication. Notably, reasoning capabilities and model scale do not guarantee better performance, and systems deteriorate with longer interactions or when supporting patients with severe symptoms. We release all code, prompts, and human evaluation data.

Marcos Treviso, Nuno M. Guerreiro, Sweta Agrawal et al.

While machine translation (MT) systems are achieving increasingly strong performance on benchmarks, they often produce translations with errors and anomalies. Understanding these errors can potentially help improve the translation quality and user experience. This paper introduces xTower, an open large language model (LLM) built on top of TowerBase designed to provide free-text explanations for translation errors in order to guide the generation of a corrected translation. The quality of the generated explanations by xTower are assessed via both intrinsic and extrinsic evaluation. We ask expert translators to evaluate the quality of the explanations across two dimensions: relatedness towards the error span being explained and helpfulness in error understanding and improving translation quality. Extrinsically, we test xTower across various experimental setups in generating translation corrections, demonstrating significant improvements in translation quality. Our findings highlight xTower's potential towards not only producing plausible and helpful explanations of automatic translations, but also leveraging them to suggest corrected translations.