Molecular · jablonkagroup
chempile-mlift
A multimodal chemistry dataset of 61.5 million molecule images and text representations for property prediction.
What chempile-mlift is
ChemPile-MLIFT is a multimodal chemistry dataset created by the Jablonka Group as part of the broader ChemPile initiative. Designed for training and evaluating vision large language models (VLLMs), it provides structured chemical data to predict molecular properties from textual identifiers and visual structural depictions.
The dataset unifies records from numerous established chemical databases, including QM9, ChEMBL, and RDKit-derived property collections. By combining 2D molecular images with multiple string-based representations across 61.5 million examples, it enables computational researchers to benchmark multimodal systems on varied chemical prediction tasks.
How it was made
ChemPile-MLIFT aggregates chemical property records into distinct configurations covering drug discovery, safety, bioactivity, computational chemistry, and foundational databases. Every entry couples a rendered molecular image (in PNG format) with text notations: SMILES (Simplified Molecular Input Line Entry System), SELFIES (SELF-referencIng Embedded Strings), InChI (International Chemical Identifier), and systematic IUPAC nomenclature.
Each record includes task templates ('template' and 'template_original') that frame the chemical notations and target property values into prompts suitable for instruction-tuned models. Entries are pre-assigned into training, validation, and test splits to support systematic model evaluation across all supported property domains.
Key facts
- Total examples
- 61.5 million
- Image format
- PNG (one molecular image per example)
- Text representations
- SMILES, SELFIES, InChI, IUPAC
- Dataset splits
- Train, validation, test
- Prompt format
- Task and original prompt templates
What you can use it for
- ADMET property prediction
- Forecasting absorption, distribution, metabolism, excretion, and toxicity profiles, including blood-brain barrier permeability, clearance, and cytochrome P450 enzyme interactions.
- Toxicology and safety profiling
- Evaluating potential hazards such as mutagenicity (Ames test), hERG cardiac potassium channel inhibition, carcinogenicity, and acute lethality metrics.
- Target binding and bioactivity modeling
- Screening molecular efficacy against biological targets, including beta-secretase, muscarinic receptors, SARS-CoV-2 viral proteins, and Maximum Unbiased Validation benchmarks.
- Quantum and physical property calculation
- Predicting computational and physical metrics, such as quantum mechanical properties from QM8 and QM9, thermodynamic flash points, and RDKit molecular descriptors.
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 jablonkagroup/chempile-mlift at the moment you asked. Nothing is cached or stored — every row above came from that request.
Strengths and limits
Good at
- Provides 61.5 million examples that pair visual molecular diagrams directly with multiple text notations.
- Features four complementary molecular string representations (SMILES, SELFIES, InChI, and IUPAC) for every record.
- Spans diverse chemical subfields, covering pharmacology, toxicology, quantum mechanics, and reference knowledge bases.
- Offers standardised data fields and templated instruction formats across dozens of benchmark configurations.
Watch out for
- Requires multimodal architectures capable of processing both visual images and text sequences to utilise the full feature set.
- Varying target properties across distinct configurations necessitate task-specific model heads or configuration-specific tuning.
Questions
Can I look inside chempile-mlift 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 chempile-mlift under?
Its authors declare cc-by-nc-sa-4.0. Read the terms on the original page before reusing the data.
What molecular representations are included in each record?
Each record contains a 2D image of the molecule in PNG format alongside SMILES, SELFIES, InChI, and IUPAC textual representations.
How large is ChemPile-MLIFT?
The dataset contains 61.5 million individual examples across its various configurations, each accompanied by one image.
Which source databases are incorporated into the configurations?
Data is sourced from established chemistry repositories and benchmarks, including QM9, QM8, Tox21, ChEMBL v29, ZINC, ChEBI, and RDKit descriptor calculations.
How are task inputs and targets structured for language models?
Records include pre-formatted 'template' and 'template_original' fields that structure the molecular inputs and target chemical properties into standardized prompts.
Written from chempile-mlift'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.