EmbedMol: An Open Billion-scale Molecular Embedding Dataset for Molecular Discovery

(For NeurIPS 2025 Datasets and Benchmark Track Review ONLY)

Description of the Data and File Structure

This repository contains the complete EmbedMol-1B dataset, partitioned into 40 parts. Each part is provided as a tarball (e.g., embedmol-1b-part-39.tar.gz). Inside each tarball, the embeddings are further split into four batches, with each batch stored as a compressed NumPy array of approximately 900 MB. Each tarball is about 3.7 GB.

Metadata for each embedding instance is provided in embedmol-1b-metadata.tar.gz, organized as CSV files (e.g., metadata_part_39_batch_all.csv) for each of the 40 parts. Each metadata entry includes the ID, original SMILES string, source dataset, and a validity flag, as detailed in the accompanying paper.

Due to the large size (hundreds of GB per file) of the generated indices for evaluation reproducibility and the immense size (tens of TB) of EmbedMol-11B, we are unable to upload these files directly to Hugging Face, per their hosting policy. However, we will work with and provide access to reviewers upon request. We thank the reviewers in advance for their understanding.

Future Hosting Plan

To ensure future accessibility, and in compliance with Hugging Face policies, we are actively working with the Hugging Face team to host the entire dataset through their platform. Given the scale of our dataset, coordination with Hugging Face is required. Public access and persistent hosting will be enabled as soon as feasible.

Original Datasets

Data was derived from the following sources:

• GDB-13: https://gdb.unibe.ch/downloads/
• ZINC-22: https://cartblanche.docking.org/