We annotate PubMed Central paragraphs for document type, domain, and educational quality using a two-stage pipeline: Llama-3.1-70B labels 400K paragraphs, then a fine-tuned XLM-RoBERTa propagates annotations to the full corpus. This paragraph-level approach captures content diversity within scientific articles that document-level labels miss. The resulting Biomed-Enriched corpus contains 2M clinical case paragraphs, providing a publicly available alternative to restricted clinical datasets. For decoders, continual pretraining experiments enable targeted improvements, with clinical upsampling boosting performance by 4 points on MMLU ProfMed and educational filtering improving MedQA and MedMCQA by ~1 point. Combinations of these techniques led to faster convergence, reaching the same performance with a third of training tokens. For encoders, our best recipe matches BioClinical-ModernBERT on 11 tasks (77.3% vs 77.1% F1) while using 2.5x fewer tokens and only public data.

Standardized benchmarks have become the dominant metric for measuring progress in large language models, yet their validity is increasingly compromised by data contamination and the unclear relationship between benchmark scores and genuine language understanding. We introduce Gaperon, a suite of fully open bilingual (French-English) language models designed as an experimental testbed to investigate evaluation dynamics under realistic training conditions. Our study makes three core contributions. First, we demonstrate mismatches between benchmark performance and generation quality: models that excel on benchmarks may underperform in qualitative text generation, and vice versa. Second, through our deliberately contaminated Gaperon-Garlic variant, we show that competitive benchmark scores can be recovered via late-stage contamination with only moderate degradation of generation quality, and surprisingly, such contamination also improves performance on held-out benchmarks. Third, we provide empirical evidence that widely used neural quality filters, particularly those trained to favor instructional or educational content, amplify benchmark contamination in pretraining corpora, with the DCLM classifier systematically ranking benchmark samples in the top-5 percentiles of samples. We release all models, data mixtures, checkpoints, and evaluation code to support reproducibility and further investigation.

Language models have long been shown to embed geographical information in their hidden representations. This line of work has recently been revisited by extending this result to Large Language Models (LLMs). In this paper, we propose to fill the gap between well-established and recent literature by observing how geographical knowledge evolves when scaling language models. We show that geographical knowledge is observable even for tiny models, and that it scales consistently as we increase the model size. Notably, we observe that larger language models cannot mitigate the geographical bias that is inherent to the training data.

The representation degeneration problem is a phenomenon that is widely observed among self-supervised learning methods based on Transformers. In NLP, it takes the form of anisotropy, a singular property of hidden representations which makes them unexpectedly close to each other in terms of angular distance (cosine-similarity). Some recent works tend to show that anisotropy is a consequence of optimizing the cross-entropy loss on long-tailed distributions of tokens. We show in this paper that anisotropy can also be observed empirically in language models with specific objectives that should not suffer directly from the same consequences. We also show that the anisotropy problem extends to Transformers trained on other modalities. Our observations tend to demonstrate that anisotropy might actually be inherent to Transformers-based models.

Static subword tokenization algorithms have been an essential component of recent works on language modeling. However, their static nature results in important flaws that degrade the models’ downstream performance and robustness. In this work, we propose MANTa, a Module for Adaptive Neural TokenizAtion. MANTa is a differentiable tokenizer trained end-to-end with the language model. The resulting system offers a trade-off between the expressiveness of byte-level models and the speed of models trained using subword tokenization. In addition, our tokenizer is highly explainable since it produces an explicit segmentation of sequences into blocks. We evaluate our pre-trained model on several English datasets from different domains as well as on synthetic noise. We find that MANTa improves robustness to character perturbations and out-of-domain data. We then show that MANTa performs comparably to other models on the general-domain GLUE benchmark. Finally, we show that it is considerably faster than strictly byte-level models.