Chemistry Reasoning Corpus · jablonkagroup
chempile-reasoning
Multi-step chemistry reasoning corpus — open-domain QA, NLI, and multiple-choice items with chains of reasoning for training and evaluating chemical reasoning models.
What chempile-reasoning is
ChemPile-Reasoning is a specialised text dataset developed by the Jablonka Group to evaluate and train machine learning models on complex scientific reasoning tasks. As a component of the ChemPile project, it compiles multi-step reasoning items across chemistry, materials modelling, and physics, combining community discussions with model-generated reasoning paths.
The dataset addresses the challenge of teaching computational models how to interpret experimental data and work through domain-specific questions. By pairing open-domain questions, multiple-choice items, and natural language inference tasks with analytical spectral reasoning, it provides structured materials for chemical structure elucidation and scientific problem solving.
How it was made
The corpus is divided into two primary sections: Stack Exchange forum data and spectral reasoning tasks. The Stack Exchange subset draws material from the chemistry, matter modelling, and physics communities, offering them in raw text formats as well as standardised completion and instruction templates with answer choices. The spectral subset covers reasoning over nuclear magnetic resonance (NMR) and infrared (IR) spectra, matching analytical questions with molecular structures denoted in SMILES (Simplified Molecular Input Line Entry System, a text-based chemical notation).
Dataset construction involves automated collection from Stack Exchange platforms alongside reasoning traces distilled from Claude-3.5-Sonnet and DeepSeek-R1 models. The data undergo domain-specific filtering for scientific relevance, quality control using expert validation and automated screening, text formatting into completion or instruction templates, and partitioning into train, validation, and test splits.
Key facts
- Total tokens
- ~21.3 billion
- Primary disciplines
- Chemistry, physics, matter modelling
- Spectral modalities
- Proton NMR, carbon NMR, infrared (IR) spectroscopy
- Distribution format
- Apache Parquet
- Language
- English
What you can use it for
- Spectral Structure Elucidation
- Training and evaluating systems to infer molecular structures and output corresponding SMILES strings from proton NMR, carbon NMR, and infrared spectra.
- Scientific Question Answering
- Developing language models that provide grounded explanations and solutions to peer-level queries in physical and chemical sciences.
- Educational Tutoring Systems
- Building educational agents capable of demonstrating step-by-step reasoning chains for chemistry and materials science problems.
- Multiple-Choice Benchmark Evaluation
- Assessing the multiple-choice reasoning accuracy of models across chemistry, physics, and spectral interpretation datasets.
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-reasoning at the moment you asked. Nothing is cached or stored — every row above came from that request.
How the columns are distributed
Over 3,207 rows of chemistry_stackexchange-completion_0/train
answer_choices
list- min
- 0
- median
- 0
- mean
- 0
- max
- 0
correct_output_index
floatNo distribution reported for this column.
100.0% empty
input
string_text- min
- 40
- median
- 1,302
- mean
- 1538.73
- max
- 9,847
output
string_label- 100%
text
string_text- min
- 40
- median
- 1,302
- mean
- 1538.73
- max
- 9,847
Strengths and limits
Good at
- Integrates both forum-based scientific problem discussions and specialised spectral analysis reasoning traces.
- Provides data across multiple standardised representations, including raw text, instruction templates, and completion formats.
- Supplies explicit SMILES strings and correctness indicators for evaluating spectral reasoning outputs.
Watch out for
- Content is restricted entirely to the English language.
- Data quality varies across sources, and certain machine-generated reasoning traces contain inconsistencies or errors.
- Evaluating the material requires advanced domain expertise in physical and chemical sciences.
Questions
Can I look inside chempile-reasoning 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-reasoning under?
Its authors declare cc-by-sa-4.0. Read the terms on the original page before reusing the data.
Which language models were used to generate the reasoning traces?
Reasoning traces were distilled from DeepSeek-R1 for multi-spectra reasoning across proton NMR, carbon NMR, and IR, and from Claude-3.5-Sonnet for single-spectrum proton NMR reasoning.
What formats are provided for the Stack Exchange subsets?
The chemistry, matter modelling, and physics Stack Exchange data are made available in unformatted raw data configurations, two completion configurations, and three instruction configurations.
How are chemical compounds represented in the dataset?
Molecules are represented as SMILES strings within the fields for target compounds and model-generated responses.
How to cite it
@article{mirza2025chempile0, title = {ChemPile: A 250GB Diverse and Curated Dataset for Chemical Foundation Models}, author = {Adrian Mirza and Nawaf Alampara and Martiño Ríos-García and others}, year = {2025}, journal = {arXiv preprint arXiv:2505.12534}}
Written from chempile-reasoning'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.