Hammam Abdelwahab

Mehdi Ali, Michael Fromm, Klaudia Thellmann et al.

The recent success of Large Language Models (LLMs) has been predominantly driven by curating the training dataset composition, scaling of model architectures and dataset sizes and advancements in pretraining objectives, leaving tokenizer influence as a blind spot. Shedding light on this underexplored area, we conduct a comprehensive study on the influence of tokenizer choice on LLM downstream performance by training 24 mono- and multilingual LLMs at a 2.6B parameter scale, ablating different tokenizer algorithms and parameterizations. Our studies highlight that the tokenizer choice can significantly impact the model's downstream performance and training costs. In particular, we find that the common tokenizer evaluation metrics fertility and parity are not always predictive of model downstream performance, rendering these metrics a questionable proxy for the model's downstream performance. Furthermore, we show that multilingual tokenizers trained on the five most frequent European languages require vocabulary size increases of factor three in comparison to English. While English-centric tokenizers have been applied to the training of multi-lingual LLMs in the past, we find that this approach results in a severe downstream performance degradation and additional training costs of up to 68%, due to an inefficient tokenization vocabulary.

Mehdi Ali, Michael Fromm, Klaudia Thellmann et al.

We present two multilingual LLMs, Teuken 7B-base and Teuken 7B-instruct, designed to embrace Europe's linguistic diversity by supporting all 24 official languages of the European Union. Trained on a dataset comprising around 60% non-English data and utilizing a custom multilingual tokenizer, our models address the limitations of existing LLMs that predominantly focus on English or a few high-resource languages. We detail the models' development principles, i.e., data composition, tokenizer optimization, and training methodologies. The models demonstrate strong performance across multilingual benchmarks, as evidenced by their performance on European versions of ARC, HellaSwag, and TruthfulQA.

Niclas Doll, Jasper Schulze Buschhoff, Shalaka Satheesh et al.

This paper narrows the performance gap between small, specialized models and significantly larger general-purpose models through domain adaptation via continual pre-training and merging. We address the scarcity of specialized non-English data by constructing a high-quality German medical corpus (FineMed-de) from FineWeb2. This corpus is used to continually pre-train and merge three well-known LLMs (ranging from $7B$ to $24B$ parameters), creating the DeFineMed model family. A comprehensive evaluation confirms that specialization dramatically enhances $7B$ model performance on German medical benchmarks. Furthermore, the pairwise win-rate analysis of the Qwen2.5-based models demonstrates an approximately $3.5$-fold increase in the win-rate against the much larger Mistral-Small-24B-Instruct through domain adaptation. This evidence positions specialized $7B$ models as a competitive, resource-efficient solution for complex medical instruction-following tasks. While model merging successfully restores instruction-following abilities, a subsequent failure mode analysis reveals inherent trade-offs, including the introduction of language mixing and increased verbosity, highlighting the need for more targeted fine-tuning in future work. This research provides a robust, compliant methodology for developing specialized LLMs, serving as the foundation for practical use in German-speaking healthcare contexts.

Nicolo' Brandizzi, Hammam Abdelwahab, Anirban Bhowmick et al.

This paper presents a comprehensive overview of the data preparation pipeline developed for the OpenGPT-X project, a large-scale initiative aimed at creating open and high-performance multilingual large language models (LLMs). The project goal is to deliver models that cover all major European languages, with a particular focus on real-world applications within the European Union. We explain all data processing steps, starting with the data selection and requirement definition to the preparation of the final filtered data. We distinguish between curated data and web data, as each of these categories is handled by distinct pipelines, with curated data undergoing minimal filtering and web data requiring extensive filtering and deduplication. This distinction guided the development of specialized algorithmic solutions for both pipelines. In addition to describing the processing methodologies, we provide an in-depth analysis of the datasets, increasing transparency and alignment with European data regulations. Finally, we share key insights and challenges faced during the project, offering recommendations for future endeavors in large-scale multilingual data preparation for LLMs.

Rebekka Görge, Sujan Sai Gannamaneni, Tabea Naeven et al.

Textual data used to train large language models (LLMs) exhibits multifaceted bias manifestations encompassing harmful language and skewed demographic distributions. Regulations such as the European AI Act require identifying and mitigating biases against protected groups in data, with the ultimate goal of preventing unfair model outputs. However, practical guidance and operationalization are lacking. We propose a comprehensive data bias detection and mitigation pipeline comprising four components that address two data bias types, namely representation bias and (explicit) stereotypes for a configurable sensitive attribute. First, we leverage LLM-generated word lists created based on quality criteria to detect relevant group labels. Second, representation bias is quantified using the Demographic Representation Score. Third, we detect and mitigate stereotypes using sociolinguistically informed filtering. Finally, we compensate representation bias through Grammar- and Context-Aware Counterfactual Data Augmentation. We conduct a two-fold evaluation using the examples of gender, religion and age. First, the effectiveness of each individual component on data debiasing is evaluated through human validation and baseline comparison. The findings demonstrate that we successfully reduce representation bias and (explicit) stereotypes in a text dataset. Second, the effect of data debiasing on model bias reduction is evaluated by bias benchmarking of several models (0.6B-8B parameters), fine-tuned on the debiased text dataset. This evaluation reveals that LLMs fine-tuned on debiased data do not consistently show improved performance on bias benchmarks, exposing critical gaps in current evaluation methodologies and highlighting the need for targeted data manipulation to address manifested model bias.