LeMaterial (LeMat-Bulk)

LeMat-Bulk is the flagship dataset of LeMaterial, an open source initiative led by Entalpic and Hugging Face to make computational materials data easier to work with. It unifies, cleans and standardises three of the most widely used bulk materials databases, Materials Project, Alexandria and OQMD, into one harmonised schema of around 6.7 million entries carrying seven materials properties. Because each source database defines and calculates properties differently, the main value here is consistency: energies, structures and metadata are brought into a single, comparable format so you are not stitching together incompatible dumps.

Access is through the Hugging Face dataset hub, where the data is served as Parquet and split into subsets by exchange correlation functional (compatible PBE, PBESol and SCAN, plus a non compatible split). A companion dataset, LeMat-BulkUnique, applies a structure fingerprint hashing algorithm to de-duplicate materials across the merged sources, which is useful when you want a clean, non redundant set. You can stream it with the datasets library or bulk download the Parquet files directly.

For RAG and AI work this is a strong base for materials informatics: training or fine-tuning property prediction and generative models, building structured retrieval over crystal structures and computed properties, or grounding a chemistry assistant in verifiable computed data. It is tabular and structured rather than prose, so for RAG you will typically retrieve records and render facts rather than chunk free text.

Watch-outs: the data is computed (DFT) rather than experimental, so values carry the usual approximations and functional dependence. The compilation is released under CC BY 4.0, but the underlying databases each ship their own terms, so check the original licences if provenance matters for your use. If you need broader materials or chemistry coverage, Materials Project and OQMD remain available directly as named alternatives.