Chemistry Education Corpus · jablonkagroup
chempile-education
A chemistry education text dataset covering textbooks, university lectures, and competition problems for language models.
What chempile-education is
ChemPile-Education is a natural language dataset compiled by the Jablonka Group as part of the broader ChemPile initiative. It aggregates foundational chemistry knowledge structured as learners encounter it across textbooks, lecture halls, and academic competitions. The corpus addresses the requirement for specialised educational text to support chemistry reasoning and comprehension in artificial intelligence systems.
Its materials span introductory undergraduate topics through to graduate-level theory, divided into four primary sub-collections: open-source textbooks, university lecture transcripts, competition examination problems with complete solutions, and educational video transcripts.
How it was made
Data collection integrates multiple ingestion pathways tailored to original material formats. Open-source textbook pages from LibreTexts Chemistry are harvested using custom HTML scrapers, covering general, organic, inorganic, physical, and biological chemistry. Transcripts from MIT OpenCourseWare are gathered through keyword-targeted course retrieval spanning chemistry, chemical engineering, biology, and physics. Educational video transcripts under Creative Commons licences are gathered via automated search queries and cleaned using GPT-4.1.
Competition items from the American Chemical Society US Chemistry Olympiad undergo PDF parsing and question-and-answer extraction assisted by Gemini 2.0 Flash Thinking Experimental 01-21. These competition entries are subjected to programmatic quality filtering, requiring solutions of at least 250 characters, followed by manual review from chemistry professionals. All subsets are formatted with standard metadata fields, partitioned into training, validation, and test splits, and stored in Parquet format.
Key facts
- Total tokens
- ~129.26 million
- Subsets
- LibreText Chemistry, MIT OCW Transcripts, US Olympiad Problems, YouTube Transcripts
- Language
- English
- Modality
- Text
- Data format
- Parquet
- Publication reference
- arXiv:2505.12534
What you can use it for
- Chemical Language Model Pre-training and Tuning
- Supplies domain-specific instructional text to pre-train or fine-tune neural language models on chemistry concepts, terminology, and problem-solving steps.
- Educational Tutoring and Question-Answering Systems
- Provides solved competition exercises and pedagogical explanations to underpin chemistry automated tutoring agents and assessment tools.
- Domain Adaptation
- Adapts general-domain foundation models to chemical reasoning by exposing them to structured lecture notes, textbook chapters, and spoken explanations.
- Information Retrieval and Knowledge Bases
- Serves as an indexed knowledge corpus for chemical concept retrieval, question matching, and retrieval-augmented generation pipelines.
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-education at the moment you asked. Nothing is cached or stored — every row above came from that request.
How the columns are distributed
Over 53,051 rows of LibreText_Chemistry-default/train
text
string_text- min
- 501
- median
- 5,219
- mean
- 7662.92
- max
- 228k
url
string_text- min
- 0
- median
- 151
- mean
- 154.47
- max
- 550
Strengths and limits
Good at
- Spans multiple educational tiers, moving from introductory university coursework to challenging Olympiad competition problems.
- Pairs automated model-based text extraction with human expert verification on competition problem solutions.
- Combines diverse textual styles, including formal prose from textbooks, conversational explanations from transcripts, and structured multiple-choice problems.
Watch out for
- Exclusively contains English-language content.
- Focuses on educational materials and fundamental concepts rather than recent primary research literature.
- Some automated speech-to-text transcripts may contain transcription errors.
Questions
Can I look inside chempile-education 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-education under?
Its authors declare cc-by-nc-sa-4.0. Read the terms on the original page before reusing the data.
Which chemistry sub-disciplines are included in the dataset?
The dataset spans general chemistry, organic chemistry, inorganic chemistry, physical chemistry, and biochemistry, as well as cross-disciplinary material touching biology, chemical engineering, and physics.
What information is provided for the US Olympiad problems?
Each Olympiad entry contains problem metadata, the original question statement, multiple-choice options, the designated correct answer, a detailed explanatory solution, and a quality-control verification flag.
What role did language models play in constructing this corpus?
GPT-4.1 was utilised to clean and structure educational video transcripts, while Gemini 2.0 Flash Thinking Experimental 01-21 was employed to extract structured question-and-answer pairs from Olympiad documents.
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-education'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.