The project is funded by the Swedish Research Council (VR) for 2025–2027 (No: 2024-01506) and investigates how large language models can represent sensory knowledge, with a particular focus on olfaction. Olfactory concepts provide a demanding test case because smell is comparatively difficult to describe linguistically, while current LLMs are trained primarily on text and therefore lack direct sensory experience. The project asks whether language models can nevertheless acquire representations of odors that correspond to human perceptual judgments. The computational work focuses on developing olfaction-enriched open-source LLMs using several complementary sources of information: natural-language descriptions of odor experiences, perceptual similarity and descriptor judgments, and chemical representations of odorants. The project also includes large-scale mining of olfactory information from text corpora, including extraction and filtering of odor-related narratives, followed by preprocessing and construction of training and evaluation datasets. Model development will rely primarily on parameter-efficient fine-tuning, including low-rank adapter methods and X-LoRA/mixture-of-experts approaches, allowing olfactory information from different data sources to be incorporated into existing general-purpose LLMs without full model retraining. Experiments will cover models of different sizes and will involve iterative fine-tuning, hyperparameter exploration, inference, and systematic evaluation. The resulting models will be evaluated on established olfactory tasks as well as a new benchmark designed to measure how closely model representations correspond to human olfactory perception. The project will ultimately provide open models, datasets, and evaluation resources for research at the intersection of computational linguistics, cognitive science, and olfactory perception.