We are LamaLab, a research group working on machine learning and foundation models for chemistry and materials science at the Helmholtz Institute for Polymers in Energy Applications (HIPOLE Jena) and the Friedrich-Schiller University Jena.
We apply data-driven methods to discover materials that work in the real world — building models that act as navigation systems for chemical space, in close collaboration with experimental partners, to bridge the gap between computational predictions and practical applications. We use large language models to unlock the tacit knowledge buried in the scientific literature, design better representations and inductive biases, and train models that are right for the right reasons — not merely well correlated with the ground truth.
We are a collaborative team with a flat hierarchy: everyone shapes the ideas, the code, and the direction of the science. And we publish what we build — because computational work without open code is mere advertisement. Please use our code, build on it, and contribute back.
Our publications, team, mentoring approach, and open positions live at lamalab.org.
corral: do AI "scientists" actually reason, or do they reach the right answer for the wrong reasons? A framework to probe scientific reasoning in LLM agents.
ChemBench: a benchmark measuring the chemical knowledge and reasoning of large language models against the expertise of human chemists.
ChemPile: a 250 GB diverse, curated, open dataset for training chemical foundation models.
MaCBench: probing the limits of multimodal language models across chemistry and materials science.
PolyMetriX: a Python ecosystem for digital polymer chemistry.
SECS: end-to-end elucidation of molecular structures directly from raw spectra.
llm-tutorial: a hands-on tutorial on LLMs and agents for the chemical sciences.
GPMs-book: an open book on general-purpose models for the chemical sciences (gpmbook.lamalab.org).
matextract-book: a practical guide to extracting data from the scientific literature with LLMs (matextract.pub).
Besides our funding agencies, we thank Modal and Sourcery for compute credits.