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.

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.

The improvements in neural machine translation make translation and post-editing pipelines ever more effective for a wider range of applications. In this paper, we evaluate the effectiveness of such a pipeline for the translation of scientific documents (limited here to article abstracts). Using a dedicated interface, we collect, then analyse the post-edits of approximately 350 abstracts (English→French) in the Natural Language Processing domain for two groups of post-editors: domain experts (academics encouraged to post-edit their own articles) on the one hand and trained translators on the other. Our results confirm that such pipelines can be effective, at least for high-resource language pairs. They also highlight the difference in the post-editing strategy of the two subgroups. Finally, they suggest that working on term translation is the most pressing issue to improve fully automatic translations, but that in a post-editing setup, other error types can be equally annoying for post-editors.

In this work, we introduce a comprehensive error typology specifically designed for evaluating two distinct tasks in machine-generated patent texts: claims-to-abstract generation, and the generation of the next claim given previous ones. We have also developed a benchmark, PatentEval, for systematically assessing language models in this context. Our study includes a comparative analysis, annotated by humans, of various models. These range from those specifically adapted during training for tasks within the patent domain to the latest general-purpose large language models (LLMs). Furthermore, we explored and evaluated some metrics to approximate human judgments in patent text evaluation, analyzing the extent to which these metrics align with expert assessments. These approaches provide valuable insights into the capabilities and limitations of current language models in the specialized field of patent text generation.