LearningMatter/

ALM-Bench

Training pairs for the Atomistic Language Model and the ALM Bench benchmark. Each row pairs a crystal structure with a natural-language instruction/target. Two clearly-separated families:

	ALM Bench: instruction-following / editing + OOD (the benchmark contribution)




	

bucket rows description

atomtxt 270 k directional property editing ("make the formation energy lower"; density/volume)

app 20 k application-conditioned… See the full description on the dataset page: https://huggingface.co/datasets/LearningMatter/ALM-Bench.