Dataset Card for SafeChem

Dataset Summary

SafeChem is a regulatory-grounded benchmark dataset of 32,614 chemical substances for multi-label GHS (Globally Harmonized System) hazard prediction and LLM safety reliability evaluation. Unlike prior molecular benchmarks constructed by querying pharmaceutical databases, SafeChem is seeded from a curated hazardous materials registry, ensuring coverage of real-world industrial and safety-critical chemicals including solvents, reactive gases, agrochemicals, and heavy metal compounds underrepresented in drug-focused benchmarks.

Each substance is annotated with molecular structure representations (canonical SMILES, InChI, InChI key), eight physicochemical descriptors, free-text physical descriptions, and 30 binary GHS hazard classification labels spanning toxicological, physical, and environmental hazard categories derived from European regulatory sources (CLP Regulation and ECHA C&L Inventory as aggregated by PubChem).

The dataset supports two benchmark tasks:

• Task 1 — Multi-label hazard prediction: Scaffold-split evaluation of GHS hazard classification from molecular structure across 17 retained hazard classes with sufficient test-set support.
• Task 2 — LLM safety hallucination benchmark: Evaluation of omission hallucination rate, commission hallucination, and silent abstention behavior across LLMs on a curated 500-substance high-hazard subset.

Dataset Details

Dataset Sources

• Primary source: Hazardous Materials Management Information System (HMMIS) regulatory registry (anonymized for double-blind review)
• Molecular enrichment: PubChem Compound and GHS Classification databases (https://pubchem.ncbi.nlm.nih.gov)
• GHS annotations: ECHA CLP Regulation submissions and C&L Inventory, Japan NITE, and other national regulatory bodies as aggregated by PubChem
• Paper: Anonymous submission to NeurIPS 2026 Datasets and Benchmarks Track (link to be added in camera-ready version)

Dataset Structure

Files

File	Description	Size
full_dataset.csv	Complete 32,614-substance master file with all 57 columns	32,614 rows
split_train.csv	Training partition (70%, scaffold split)	~22,829 rows
split_val.csv	Validation partition (15%, scaffold split)	~4,892 rows
split_test.csv	Test partition (15%, scaffold split)	~4,893 rows

Column Groups

Molecular identifiers

• cas_number, cid, inchi, inchi_key, canonical_smiles, isomeric_smiles, iupac_name, molecular_formula

Molecular descriptors (pre-computed via RDKit and PubChem)

• molecular_weight, xlogp, tpsa, complexity, h_bond_donor_count, h_bond_acceptor_count, rotatable_bond_count, heavy_atom_count

Physical properties

• boiling_point, flash_point, physical_description

GHS hazard labels (30 binary columns, 1 = hazard applies, 0 = not classified)

• Health: acute_toxicity_oral, acute_toxicity_dermal, acute_toxicity_inhalation, carcinogenicity, germ_cell_mutagenicity, reproductive_toxicity, skin_corrosion_irritation, skin_sensitization, eye_damage_irritation, respiratory_sensitization, aspiration_hazard, stot_single_exposure, stot_repeated_exposure
• Physical: flammable_liquids, flammable_gases, flammable_solids, flammable_aerosols, explosives, oxidizing_liquids, oxidizing_solids, oxidizing_gases, organic_peroxides, self_reactive_substances, pyrophoric_substances, self_heating_substances, water_reactive_substances, gases_under_pressure, metal_corrosion
• Environmental: acute_aquatic_toxicity, chronic_aquatic_toxicity

Dataset Splits

Splits were generated using Bemis-Murcko scaffold splitting to prevent structural data leakage between train and test, combined with iterative stratification to preserve per-label positive rates across partitions. This is the standard evaluation protocol for molecular ML benchmarks. A 70/15/15 train/validation/test ratio was applied.

Task 1 benchmark uses 17 of the 30 labels — those with at least 10 positive instances in the test partition after scaffold splitting. The remaining 13 labels are retained in the full dataset and used in Task 2.

Dataset Creation

Curation Rationale

Existing molecular ML benchmarks (MoleculeNet, Tox21) are populated primarily from pharmaceutical and bioassay databases, biasing their chemical space toward drug-like molecules. Real-world chemical safety workflows involve industrial solvents, reactive gases, agrochemicals, and heavy metal compounds that are systematically underrepresented in these benchmarks. SafeChem was created to fill this gap by grounding the substance population in a regulatory hazardous materials registry rather than database availability.

Data Collection and Processing

1. Seed corpus: Substance identifiers (CAS numbers) were extracted from a regulatory hazardous materials management information system (HMMIS).
2. PubChem enrichment: CAS numbers were resolved to PubChem CIDs via the PubChem identifier exchange service. Records were retained only if they had a valid canonical SMILES and at least one GHS hazard classification entry.
3. Descriptor retrieval: Molecular descriptors were retrieved from PubChem's computed properties endpoint and cross-validated against RDKit recomputation; discrepancies exceeding 5% used the RDKit value.
4. Deduplication: Records were deduplicated by PubChem CID. Label conflicts across duplicate records (20 CIDs, 23 label-level conflicts) were resolved by union — a hazard class was assigned positive if any source record carried a positive classification.
5. Filtering: Substances with no positive GHS labels were excluded. The final corpus contains 32,614 unique substances.

Source Data Producers

GHS hazard annotations are regulatory submissions from government bodies and industry to ECHA, NITE, and other national authorities, as aggregated by PubChem. Molecular structure data is sourced from PubChem's Compound database.

Annotations

GHS hazard labels are binary indicators derived from regulatory classification submissions rather than manual annotation by the dataset authors. Zero values indicate the absence of a positive GHS classification in PubChem's aggregated records; this predominantly reflects a regulatory determination that the hazard does not apply, though absence of data cannot be fully excluded for all substances.

Personal and Sensitive Information

The dataset contains only chemical substance data (molecular structures, physicochemical properties, regulatory classifications). No personal or sensitive information about individuals is included.

Bias, Risks, and Limitations

• Jurisdictional scope: GHS annotations are aggregated by PubChem from multiple regulatory bodies. The majority derive from ECHA submissions (CLP Regulation and C&L Inventory), with a minority from Japan's NITE and other national authorities. The precise jurisdictional distribution cannot be fully characterised as PubChem does not expose per-record source provenance. Classifications may not align with US OSHA HazCom 2012 or other national GHS implementations for all substances.
• Label noise: A subset of annotations derives from industry self-notifications in the ECHA C&L Inventory, which are self-reported and may exhibit inter-submitter inconsistency not fully resolved by union merging.
• Scaffold clustering: Four labels exhibit train/test prevalence deviations due to scaffold clustering (stot_repeated_exposure, flammable_solids, metal_corrosion, stot_single_exposure); per-label metrics for these should be interpreted with caution.
• Single compounds only: All substances are single compounds; mixture toxicology is not captured.
• Temporal snapshot: Labels reflect the state of regulatory knowledge at the time of PubChem data collection and may not reflect subsequent regulatory updates.

Glossary

• GHS: Globally Harmonized System of Classification and Labelling of Chemicals — the international standard for chemical hazard communication.
• CLP Regulation: EU Regulation (EC) No 1272/2008 on classification, labelling and packaging of substances and mixtures, which implements GHS in the European Union.
• ECHA C&L Inventory: European Chemicals Agency Classification and Labelling Inventory — a database of GHS classifications notified by industry.
• Scaffold split: A dataset partitioning strategy that separates molecules by their Bemis-Murcko core ring system, preventing structurally similar molecules from appearing in both train and test.
• OHR (Omission Hallucination Rate): The fraction of true positive hazard labels that an LLM fails to predict — the safety-critical false-negative rate.
• EPR (Empty Prediction Rate): The fraction of LLM responses that return no hazard labels, indistinguishable from a confirmed negative classification in automated pipelines.
• SMILES: Simplified Molecular Input Line Entry System — a string notation for representing molecular structures.
• AUPRC: Area Under the Precision-Recall Curve — the primary evaluation metric for Task 1, preferred over AUROC under class imbalance.