Matbench Discovery is a leaderboard that benchmarks machine-learning force fields on crystal discovery, geometry optimization, phonons, molecular dynamics and diatomics.
DatasetFree
Step 1: Open the leaderboard and pick a column preset.
Step 2: Choose the full test set or unique prototypes.
Step 3: Filter by training data, openness or targets.
Step 4: Adjust the score weights if you value a task differently.
Step 5: Export the table, or use the submit a model link to add your own.
It weights discovery F1 at 50%, geometry RMSD at 10% and phonon thermal conductivity error at 40%. MD and diatomics are left out.
Yes. The page has weight controls, and the other weights adjust proportionally so the total stays at 100%.
Yes. The page has a Submit a model link.
Yes. You can filter by training data, openness and targets, and require models that predict forces.
Claude Protein Binder DesignA data release of 1,440 de novo miniprotein binders designed by Claude models acting as autonomous protein-design agents, with wet-lab results from two contract research organizations. It covers 16 targets.QokedasTraining-data company that turns real-world scientific signals and observations that are not normally written down into datasets for improving AI models in science.Pretrain BehaviorsPretrain-Behaviors is a behavior-focused text dataset from IFM covering reasoning, planning, data science, games and format rewriting. It is part of the K2 Horizon collection.