liuganghuggingface/moldata-chemx

ChemX chemistry information-extraction datasets

ChemX benchmarks chemical information extraction from literature across nanomaterials and small molecules. Records retain extraction context and source article identifiers; the original article PDFs are not included.

Part of the Molecular Datasets collection.

Get started

from datasets import load_dataset
dataset = load_dataset("liuganghuggingface/moldata-chemx", "complexes")

Where the data comes from

• Original data: https://ai-chem.github.io/ChemX/
• Original study: ChemX chemistry information-extraction datasets primary reference

Tables

Config	Splits	Rows	Targets
complexes	train	907	target (regression; units not specified)
co_crystals	train	70	photostability_change photostability change (non_numeric; units not specified)
oxazolidinones	train	2,923	target_value (constant; units not specified)
benzimidazoles	train	1,721	target_value (classification; units not specified)
eyedrops	train	163	perm (cm/s) (non_numeric; units not specified), logP (regression; units not specified)

data/ contains rows with readable molecular structures; identical observations are collapsed. observations/ contains all source rows.

Source label columns are text, including values such as <10. SMILES_original retains the input representation; SMILES_canonical provides a canonical representation. structure_key groups molecular components, and structure_status records parsing outcomes.

Files and modalities

Modality	Where	Content
text and source labels	files/Complexes.parquet (0 MB)	Original ChemX Complexes table with DOI, article metadata, extraction locations and labels. Source article PDFs are not bundled by the authors.
text and source labels	files/Co-crystals.parquet (0 MB)	Original ChemX Co-crystals table with DOI, article metadata, extraction locations and labels. Source article PDFs are not bundled by the authors.
text and source labels	files/Nanomag.parquet (0 MB)	Original ChemX Nanomag table with DOI, article metadata, extraction locations and labels. Source article PDFs are not bundled by the authors.
text and source labels	files/Synergy.parquet (0 MB)	Original ChemX Synergy table with DOI, article metadata, extraction locations and labels. Source article PDFs are not bundled by the authors.
text and source labels	files/Nanozymes.parquet (0 MB)	Original ChemX Nanozymes table with DOI, article metadata, extraction locations and labels. Source article PDFs are not bundled by the authors.
text and source labels	files/Oxazolidinones.parquet (0 MB)	Original ChemX Oxazolidinones table with DOI, article metadata, extraction locations and labels. Source article PDFs are not bundled by the authors.
text and source labels	files/Benzimidazoles.parquet (0 MB)	Original ChemX Benzimidazoles table with DOI, article metadata, extraction locations and labels. Source article PDFs are not bundled by the authors.
text and source labels	files/EyeDrops.parquet (0 MB)	Original ChemX EyeDrops table with DOI, article metadata, extraction locations and labels. Source article PDFs are not bundled by the authors.
text and source labels	files/SelTox.parquet (0 MB)	Original ChemX SelTox table with DOI, article metadata, extraction locations and labels. Source article PDFs are not bundled by the authors.
text and source labels	files/Cytotox.parquet (0 MB)	Original ChemX Cytotox table with DOI, article metadata, extraction locations and labels. Source article PDFs are not bundled by the authors.

Download an individual file:

from huggingface_hub import hf_hub_download
path = hf_hub_download("liuganghuggingface/moldata-chemx", "files/Complexes.parquet", repo_type="dataset")

Data quality

• many sources: Compiled from several labs or papers, whose protocols differ.
• all structures readable: All 5,784 SMILES were read with RDKit.
• 21% undefined stereo: 21% of molecules have a stereocenter whose configuration is not given, so mirror-image forms cannot be told apart.
• 1 of 7 stereoisomer sets differs: Found 7 groups of molecules that differ only in stereochemistry. In 1, the values are 10-fold apart (log-scale labels), more than one standard deviation apart, or in different classes. That can be a real effect or a mistake in the source; a model that ignores stereochemistry treats these molecules as the same.
• 0% in rarer class: In a typical task, 0% of labeled molecules are in the rarer class. Accuracy is misleading here; use precision and recall.
• 72.2% measured per task: In a typical task, 72.2% of molecules have a value. A missing value means not measured, not inactive.
• 1 single-value target: These targets have the same value for every molecule, so there is nothing to learn from them.
• 365 extreme values: 365 of 2,603 values lie more than 6 robust standard deviations from the median of their target (on a log scale when values span 1,000-fold or more). They are often typing or unit mistakes, and sometimes real; look at them before training.
• mixes MIC, pMIC: Column “target_type” holds MIC, pMIC. Values of different types are not comparable, so train on one type at a time.
• 176 molecules per task: A typical task has 176 labeled molecules. Results on fewer than 1,000 molecules vary a lot between splits.
• train set only: The source released only its train set (5,784 rows). Hold out part of it for testing.
• dated 1989–2022: Column “year” holds years from 1989 to 2022, so you can train on older data and test on newer data.

Dataset notes

• Five nanomaterial tables contain native material records without SMILES.
• Records contain article DOI, PDF identifiers, and extraction locations. Article PDFs are available from their original publishers.
• Literature extraction accuracy and property prediction are different tasks.

License

MIT

Citation

Cite the original study and data source:

• ChemX chemistry information-extraction datasets primary reference
• Original release/download documentation