Chemistry Pretraining · jablonkagroup
chempile-lift
A 35.5-billion-token dataset of natural language question-answer pairs for chemical property prediction using LLMs.
What chempile-lift is
ChemPile-LIFT is a large-scale chemistry dataset designed for training large language models (LLMs) on property prediction tasks. Created by jablonkagroup as part of the broader ChemPile project, it converts structured chemical knowledge and benchmarks into textual question-answer pairs where molecules are represented using SMILES strings.
The dataset addresses the challenge of enabling text-based foundation models to reason about chemical properties across drug discovery, toxicology, materials science, and chemical synthesis. It compiles data from established sources—including QM9, RDKit, the Open Reaction Database, and the Tox21 screening programme—into a unified text interface.
Containing 263.8 million examples across multiple configurations, ChemPile-LIFT enables researchers to train and evaluate language models on diverse molecular tasks, ranging from pharmacokinetic parameter prediction to quantum mechanical calculations and chemical entity recognition.
How it was made
ChemPile-LIFT gathers chemical records from existing chemistry datasets and databases such as QM9, RDKit, and Tox21. Structures are standardized with consistent formatting and SMILES representations across all configurations.
The structured records are converted into natural language question-answer prompts. Data processing involves expert curation, duplicate removal, quality checks, and partitioning into standardized train, validation, and test splits. The dataset is distributed in Parquet format, presenting a consistent text field alongside structured keys such as input, output, answer choices, and correct output indices.
Key facts
- Total examples
- 263.8 million
- Total tokens
- 35.5 billion
- Format
- Parquet
- Modality
- Text
- Language
- English
- Molecular representation
- SMILES
- Publisher
- jablonkagroup
- Creation date
- 16 May 2025
What you can use it for
- Drug Discovery and Pharmacokinetics
- Train language models to forecast ADMET endpoints, including blood-brain barrier permeability, intestinal absorption, clearance, half-life, and CYP450 enzyme interactions.
- Toxicological Screening
- Fine-tune models to predict chemical hazards, including mutagenicity (Ames test), carcinogenicity, acute toxicity (LD50), hERG channel inhibition, and clinical adverse drug effects.
- Materials Science Characterisation
- Predict thermodynamic and quantum properties, such as formation energies, flash points, band gaps, and characteristics of metal-organic frameworks and perovskites.
- Reaction and Synthesis Modelling
- Train models on reaction outcomes, reaction procedures, and yield predictions using collections derived from the Open Reaction Database, RHEA, and patent literature.
- Biomedical Named Entity Recognition
- Develop models to identify and extract chemical, disease, and biological entities from biomedical texts and safety data sheets.
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-lift at the moment you asked. Nothing is cached or stored — every row above came from that request.
How the columns are distributed
Over 11,420 rows of BACE/train
text
string_text- min
- 80
- median
- 176
- mean
- 190.70
- max
- 445
Strengths and limits
Good at
- Covers a massive breadth of chemical domains comprising 263.8 million examples and 35.5 billion tokens.
- Unifies diverse external data sources into a consistent text-based question-and-answer format suitable for language model training.
- Provides standardized train, validation, and test splits across individual domain configurations.
Watch out for
- Content is primarily in the English language and focuses strictly on chemistry and materials science terminology.
- Contains variable source data quality, with potential residual noise and inherent biases from underlying chemical databases.
- Certain subsets contain incomplete property annotations or missing values.
Questions
Can I look inside chempile-lift 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-lift under?
Its authors declare cc-by-nc-sa-4.0. Read the terms on the original page before reusing the data.
How are molecules represented in ChemPile-LIFT?
Molecules are represented as SMILES (Simplified Molecular Input Line Entry System) strings embedded directly inside the natural language prompts and answers.
What features are present in each dataset record?
Records provide a primary 'text' field containing the complete question and answer, alongside auxiliary fields including 'input', 'output', 'answer_choices', and 'correct_output_index'.
How does ChemPile-LIFT relate to the wider ChemPile project?
ChemPile-LIFT is a specific property-prediction component of the larger ChemPile initiative, which spans over 75 billion tokens covering code, scientific literature, images, and multiple molecular representations.
How to cite it
Adrian Mirza and Nawaf Alampara and Martiño Ríos-García and others, ChemPile: A 250GB Diverse and Curated Dataset for Chemical Foundation Models, arXiv preprint arXiv:2505.12534, 2025.
Written from chempile-lift'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.