Materials Embeddings · LeMaterial
LeMat-GenBench-embeddings
Precomputed embeddings for evaluating generative materials models on crystal-structure prediction tasks.
What LeMat-GenBench-embeddings is
LeMat-GenBench-embeddings is a scientific data collection published by LeMaterial. It provides precalculated feature representations built for the LeMat-GenBench benchmark, targeting researchers across materials science, chemistry, physics, and engineering.
The dataset addresses the computational burden of calculating feature sets during model assessment. It offers ready-to-use embeddings that serve as standardised reference features for testing generative materials models on crystal-structure prediction and associated downstream tasks.
How it was made
The dataset consists of between one million and ten million text-modality records packaged in Parquet format. The stored embeddings represent materials data and are structured for direct ingestion into analysis pipelines via libraries such as Polars, Dask, and Datasets.
Key facts
- Data format
- Parquet
- Modality
- Text
- Fields
- Materials science, chemistry, physics, engineering
What you can use it for
- Crystal-structure prediction evaluation
- Benchmarking generative materials models on their ability to predict stable crystal structures using consistent precomputed features.
- Generative model benchmarking
- Comparing the performance of different materials generation architectures against the standardised LeMat-GenBench evaluation suite.
- Downstream materials property prediction
- Using the precomputed representations as fixed feature inputs for downstream physical and chemical property modelling tasks.
Explore the data
The schema, real rows, full-text search and a SQL filter — read in your browser, nothing downloaded.
Explore
Read the real rows without downloading anything
Reading rows…
LiveRead from LeMaterial/LeMat-GenBench-embeddings at the moment you asked. Nothing is cached or stored — every row above came from that request.
How the columns are distributed
Over 1,200,000 rows of default/train
immutable_id
string_text- min
- 6
- median
- 12
- mean
- 11.98
- max
- 13
91.7% empty
mace_embeddings
list- min
- 640
- median
- 640
- mean
- 640
- max
- 640
orb_embeddings
list- min
- 256
- median
- 256
- mean
- 256
- max
- 256
uma_embeddings
list- min
- 128
- median
- 128
- mean
- 128
- max
- 128
Strengths and limits
Good at
- Provides drop-in precalculated representations that remove the need to generate features from scratch for benchmark tasks.
- Stored in Parquet format, offering compatibility with modern tabular tools such as Polars, Dask, and Datasets.
Watch out for
- The specific encoder model and dimensional details used to produce the embeddings are not specified.
- The features are designed around the LeMat-GenBench benchmark, which may limit relevance outside crystal-structure evaluation workflows.
Questions
Can I look inside LeMat-GenBench-embeddings without downloading it?
Yes. This page reads its schema, real rows and per-column statistics in the browser, with full-text search and a SQL filter. Nothing is downloaded to your machine.
What licence is LeMat-GenBench-embeddings under?
Its authors declare apache-2.0. Read the terms on the original page before reusing the data.
What is the primary function of LeMat-GenBench-embeddings?
It provides precomputed embedding representations designed as drop-in features for assessing generative materials models on crystal-structure prediction and downstream evaluation tasks.
How many records are in the dataset?
The dataset contains between one million and ten million records.
What file formats and software libraries are supported?
The dataset is distributed in Parquet format and supports integration with Polars, Dask, Datasets, and mlcroissant.
Written from LeMat-GenBench-embeddings's own dataset card. Statements citing a figure that does not appear there are removed before publication. Licence and running details come from our registry.