We present an enhanced benchmark for evaluating linguistic acceptability in Danish. We first analyze the most common errors found in written Danish. Based on this analysis, we introduce a set of fourteen corruption functions that generate incorrect sentences by systematically introducing errors into existing correct Danish sentences. To ensure the accuracy of these corruptions, we assess their validity using both manual and automatic methods. The results are then used as a benchmark for evaluating Large Language Models on a linguistic acceptability judgement task. Our findings demonstrate that this extension is both broader and more comprehensive than the current state of the art. By incorporating a greater variety of corruption types, our benchmark provides a more rigorous assessment of linguistic acceptability, increasing task difficulty, as evidenced by the lower performance of LLMs on our benchmark compared to existing ones. Our results also suggest that our benchmark has a higher discriminatory power which allows to better distinguish well-performing models from low-performing ones.

With the rapid advances of large language models, it becomes increasingly important to systematically evaluate their multilingual and multicultural capabilities. Previous cultural evaluation benchmarks focus mainly on basic cultural knowledge that can be encoded in linguistic form. Here, we propose SommBench, a multilingual benchmark to assess sommelier expertise, a domain deeply grounded in the senses of smell and taste. While language models learn about sensory properties exclusively through textual descriptions, SommBench tests whether this textual grounding is sufficient to emulate expert-level sensory judgment. SommBench comprises three main tasks: Wine Theory Question Answering (WTQA), Wine Feature Completion (WFC), and Food-Wine Pairing (FWP). SommBench is available in multiple languages: English, Slovak, Swedish, Finnish, German, Danish, Italian, and Spanish. This helps separate a language model’s wine expertise from its language skills. The benchmark datasets were developed in close collaboration with a professional sommelier and native speakers of the respective languages, resulting in 1,024 questions for wine theory question answering, 1,000 examples for wine feature completion, and 1,000 examples of food-wine pairing. We provide results for the most popular language models, including closed-weights models such as Gemini 2.5, and open-weights models, such as GPT-OSS and Qwen 3. Our results show that the most capable models perform well on wine theory question answering (up to 97% correct with a closed-weights model), yet feature completion (peaking at 65%) and food-wine pairing show (MCC ranging between 0 and 0.39) turn out to be more challenging. These results position SommBench as an interesting and challenging benchmark for evaluating the sommelier expertise of language models. The benchmark is publicly available at https://github.com/sommify/sommbench.

Large-scale datasets are foundational for research and development in natural language processing. However, current approaches face three key challenges: (1) reliance on ambiguously licensed sources restricting use, sharing, and derivative works; (2) static dataset releases that prevent community contributions and diminish longevity; and (3) quality assurance processes restricted to publishing teams rather than leveraging community expertise. To address these limitations, we introduce two contributions: the Dynaword approach and Danish Dynaword. The Dynaword approach is a framework for creating large-scale, open datasets that can be continuously updated through community collaboration. Danish Dynaword is a concrete implementation that validates this approach and demonstrates its potential. Danish Dynaword contains over five times as many tokens as comparable releases, is exclusively openly licensed, and has received multiple contributions across industry, the public sector and research institutions. The repository includes light-weight tests to ensure data formatting, quality, and documentation, establishing a sustainable framework for ongoing community contributions and dataset evolution.

Language and culture are deeply intertwined, yet it has been unclear how and where multilingual large language models encode culture. Here, we build on an established methodology for identifying language-specific neurons to localize and isolate culture-specific neurons, carefully disentangling their overlap and interaction with language-specific neurons. To facilitate our experiments, we introduce MUREL, a curated dataset of 85.2 million tokens spanning six different cultures. Our localization and intervention experiments show that LLMs encode different cultures in distinct neuron populations, predominantly in upper layers, and that these culture neurons can be modulated largely independently of language-specific neurons or those specific to other cultures. These findings suggest that cultural knowledge and propensities in multilingual language models can be selectively isolated and edited, with implications for fairness, inclusivity, and alignment. Code and data are available at https://github.com/namazifard/Culture_Neurons

Preprocessing large and possibly multimodal datasets remains a key bottleneck in many machine learning workflows, particularly when random access to samples is needed for global shuffling and sorting. Existing approaches, including widely used formats like JSONL and frameworks such as Huggingface Datasets and MosaicML Streaming, typically incur substantial computational, memory, and storage overhead in such settings. Here, we introduce MLDataForge, a Python-based open-source framework designed for scalable dataset pre-processing and access. Our key contributions are: (1) optimized readers for Mosaic Data Shards (MDS) that substantially improve throughput, reduce peak storage usage, and support sample-level compression; (2) JINX (JSON Indexed ’N’ eXtended), a novel, index-augmented JSONL-compatible format supporting structured footers and binary sidecar files; and (3) a lazy-loading mechanism that defers data loading, decompression, and decoding JINX files until sample fields are accessed. We empirically evaluate MLDataForge and our contributions on a representative 200 GB supervised fine-tuning dataset for vision language models. Our best configuration – zstd-compressed JINX with binary sidecar and lazy loading – yields at least a decimal order-of-magnitude throughput increase compared to the best baselines for iteration, global shuffling, and sorting. These advances enable substantial gains in data preprocessing performance, facilitating more scalable and resource-efficient model training pipelines.

Large language models (LLMs) require immense resources for training and inference. Quantization, a technique that reduces the precision of model parameters, offers a promising solution for improving LLM efficiency and sustainability. While post-training quantization methods typically achieve 4-8 bits per parameter, recent research suggests that training LLMs with 1.58 bits per weight parameter from scratch can maintain model accuracy while greatly reducing memory requirements and energy consumption at inference time. Here, we investigate a training strategy for quantization-aware pre-training, where the models are first trained with 16-bit precision and then transition into 1.58-bit quantization-aware training. Our results on 11 downstream tasks, show that this 16-to-1.58-bit training strategy is preferable over full 1.58-bit training and leaves models closer to those which have undergone 16-bit training. We further investigate the effects of retaining the optimizer state at the transition point and gradually phasing in quantization strength - finding that both techniques alleviate the magnitude of loss spikes, but also that these effects can be compensated through further training.

We develop a multilingual version of the Wug Test, an artificial word completion experiment that is typically used to test the morphological knowledge of children, and apply it to the GPT family of large language models (LLMs). LLMs’ performance on this test was evaluated by native speakers of six different languages, who judged whether the inflected and derived forms generated by the models conform to the morphological rules of their language. Our results show that LLMs can generalize their morphological knowledge to new, unfamiliar words, but that their success in generating the “correct” generalization (as judged by native human speakers) is predicted by a language’s morphological complexity (specifically, integrative complexity). We further find that the amount of training data has surprisingly little on LLMs’ morphological generalization abilities within the scope of the analyzed languages. These findings highlight that “morphology matters”, and have important implications for improving low-resource language modeling.

Language models can serve as a valuable tool for software developers to increase productivity. Large generative models can be used for code generation and code completion, while smaller encoder-only models are capable of performing code search tasks using natural language queries. These capabilities are heavily influenced by the quality and diversity of the available training data. Source code datasets used for training usually focus on the most popular languages and testing is mostly conducted on the same distributions, often overlooking low-resource programming languages. Motivated by the NLP generalization taxonomy proposed by Hupkes et.,al., we propose a new benchmark dataset called GenCodeSearchNet (GeCS) which builds upon existing natural language code search datasets to systemically evaluate the programming language understanding generalization capabilities of language models. As part of the full dataset, we introduce a new, manually curated subset StatCodeSearch that focuses on R, a popular but so far underrepresented programming language that is often used by researchers outside the field of computer science. For evaluation and comparison, we collect several baseline results using fine-tuned BERT-style models and GPT-style large language models in a zero-shot setting.

Graph neural networks have triggered a resurgence of graph-based text classification methods, defining today’s state of the art. We show that a wide multi-layer perceptron (MLP) using a Bag-of-Words (BoW) outperforms the recent graph-based models TextGCN and HeteGCN in an inductive text classification setting and is comparable with HyperGAT. Moreover, we fine-tune a sequence-based BERT and a lightweight DistilBERT model, which both outperform all state-of-the-art models. These results question the importance of synthetic graphs used in modern text classifiers. In terms of efficiency, DistilBERT is still twice as large as our BoW-based wide MLP, while graph-based models like TextGCN require setting up an 𝒪(N²) graph, where N is the vocabulary plus corpus size. Finally, since Transformers need to compute 𝒪(L²) attention weights with sequence length L, the MLP models show higher training and inference speeds on datasets with long sequences.