Chemistry Instruction Corpus · jablonkagroup
chempile-instruction
A multi-turn dialogue dataset for instruction tuning chemistry-focused large language models across multiple subdomains.
What chempile-instruction is
ChemPile-Instruction is an instruction-tuning corpus developed by the Jablonka Group to support the fine-tuning of large language models in the chemical sciences. It consists of multi-turn dialogues and reasoning traces generated from chemical literature, educational materials, and deduction problems.
General language models frequently struggle with the technical dialogue and deductive reasoning demanded by chemistry. ChemPile-Instruction provides structured conversational data annotated by reasoning skill—including factual knowledge, mathematical calculation, and logical deduction—across major subdisciplines such as organic, inorganic, physical, and analytical chemistry.
How it was made
The dataset is organised into three distinct configurations: educational dialogues derived from textbooks and courses (chempile-education), research discussions drawn from academic publications (chempile-paper-100m), and chemical logic problems (chempile-reasoning). The dialogues were generated using gpt-4o-mini-2024-07-18 across four distinct prompt strategies designed to elicit different conversational styles: engaging explanations, advanced vocabulary, encyclopaedic wiki-style responses, and unconstrained natural model outputs.
Each entry is structured into standard LiteLLM message sequences containing user and assistant turns. Entries are annotated using ChemBench taxonomic categories to specify the required problem-solving skills (requires-knowledge, requires-calculation, requires-reasoning) and chemistry subdisciplines, alongside metadata detailing the source dataset, split, configuration, and prompt strategy.
Key facts
- Modality
- Text
- Language
- English
- Data format
- Parquet
- Generation model
- gpt-4o-mini-2024-07-18
- Configurations
- chempile-education, chempile-paper-100m, chempile-reasoning
What you can use it for
- Instruction tuning for chemistry models
- Fine-tuning base language models on multi-turn scientific interactions to enhance conversational capability and task adherence in chemical topics.
- Chemistry educational systems
- Developing interactive tutoring tools and virtual assistants capable of explaining foundational concepts at varying levels of technical complexity.
- Research paper discussion assistants
- Training specialised models to discuss, summarise, and contextualise advanced research publications in academic chemistry.
- Chemical reasoning enhancement
- Training computational models on multi-step deduction, chemical problem-solving, and domain-specific calculations.
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-instruction at the moment you asked. Nothing is cached or stored — every row above came from that request.
How the columns are distributed
Over 60,171 rows of chempile-education/train
first_tag
list- min
- 0
- median
- 0
- mean
- 0.44
- max
- 3
messages
list- min
- 1
- median
- 20
- mean
- 19.47
- max
- 138
second_tag
list- min
- 0
- median
- 1
- mean
- 1.16
- max
- 9
Strengths and limits
Good at
- Covers seven major subfields including Analytical Chemistry, Inorganic Chemistry, Organic Chemistry, Physical Chemistry, and Materials Science.
- Provides multi-turn conversational structures suited for contextual learning and dialogue adaptation.
- Uses four distinct prompting styles to train models on both accessible explanations and advanced scientific vocabulary.
- Includes taxonomy tags that classify conversations by cognitive demands such as calculation, factual retrieval, and deduction.
Watch out for
- Relies on synthetic generation via gpt-4o-mini-2024-07-18, meaning factual errors or hallucinations may be present.
- Focuses strictly on educational and academic research chemistry, excluding industrial chemical processes.
- Restricted entirely to the English language.
- Subject to the inherent biases present in the underlying training sources and generation model.
Questions
Can I look inside chempile-instruction 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-instruction under?
Its authors declare cc-by-4.0. Read the terms on the original page before reusing the data.
What chemistry subdisciplines are covered?
The dataset categorises conversations across Analytical Chemistry, General Chemistry, Inorganic Chemistry, Materials Science, Organic Chemistry, Physical Chemistry, and Technical Chemistry.
What skill categories are tagged in each dialogue?
Conversations are classified by required problem-solving skills into requires-knowledge for factual information, requires-calculation for mathematical processing, and requires-reasoning for logical deduction.
How are the dialogue styles differentiated?
Data generation uses four prompt configurations: engaging for accessible multi-step explanations, hard for dense and esoteric technical vocabulary, wiki for encyclopaedic factual delivery, and none for unconstrained generation.
What subsets are available in the dataset?
It provides three subsets: chempile-education from textbooks and online courses, chempile-paper-100m from academic research papers, and chempile-reasoning containing chemical logic problems.
How to cite it
Adrian Mirza, Nawaf Alampara, 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-instruction'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.