Hugging Face's logo

Datasets:
structure-epflai
/
coconut

Modalities:
Text
Formats:
parquet
Languages:
English
Size:
100K - 1M
Tags:
chemistry
natural-products
molecular-structures
Libraries:
Datasets
pandas
Polars
License:
Dataset card Data Studio Files
xet
 Community
coconut / README.md
rohangautam's picture
rohangautam
Add COCONUT October 2026 source-preserving CSV and Parquet release
53d4288 verified
preview code
|
Raw History Blame Contribute Delete
2.43 kB
metadata 
license: cc0-1.0
language:
  - en
pretty_name: COCONUT (2026-10)
tags:
  - chemistry
  - natural-products
  - molecular-structures
configs:
  - config_name: full
    default: true
    data_files:
      - split: train
        path: data/*.parquet

COCONUT: 2026-10 source-preserving mirror

This dataset contains 738,794 structure records and 44 columns from the full COCONUT CSV export. It contains molecular structures and metadata, not measured spectra. The source download page releases the data under CC0.

Source: 2026-10 full CSV archive. See the upstream download page.

Data and preprocessing

All source rows, columns, their order, and parsed CSV values are retained. All columns use the string type, including numeric and boolean descriptors. Empty cells remain empty strings. Pipe-delimited metadata remains source text. No molecule standardization, stereochemistry removal, deduplication, or chemistry filter is applied. canonical_smiles is the upstream representation, unchanged. Cast descriptor columns explicitly when computing chemical-space statistics.

The train split holds the full collection. It is a storage convention, not a scientific training split. No validation or test partition has been created.

from datasets import load_dataset

dataset = load_dataset("structure-epflai/coconut", "full", split="train")

Validation and provenance

Every parsed CSV value was compared with the generated Parquet value. Unique source identifiers: 738,794; duplicate identifier rows: 0. RDKit parsing was checked on the first 256 structures; 0 failed. This is a smoke check, not a full chemical audit.

manifest.json records source/output SHA-256 checksums, row counts, empty-cell counts, software versions, and the source commit and conversion-script hash. The original ZIP is retained under source/; the exact conversion script is included as collate_coconut.py. Upstream descriptors and annotations have not been independently verified. The CSV export is not the complete database SQL dump, which also includes active/inactive records and relational data.

Citation

Venkata Chandrasekhar et al. COCONUT 2.0: a comprehensive overhaul and curation of the collection of open natural products database. Nucleic Acids Research. https://doi.org/10.1093/nar/gkae1063