Lit2Vec Subfield Classifier Dataset

Summary

The Lit2Vec Subfield Classifier Dataset is a curated and preprocessed collection of scientific research metadata designed for text classification and embedding-based machine learning tasks.
It includes over 39,900 chemistry abstract and tldr text annotated with domain subfields, dense text embeddings, and structured metadata, making it suitable for:

Scientific document classification
Subfield prediction and semantic tagging
Embedding-based retrieval and similarity search
Representation learning and transfer learning on scholarly text

All records are sourced from CC-BY licensed publications in the Semantic Scholar corpus.

Dataset Structure

Splits

Train: ~80% of records
Validation: ~10% of records
Test: ~10% of records

Each entry is a flat JSON object (1 per line) in .jsonl format.

Example Record

{
  "corpus_id": 105403827,
  "doi": "10.31031/PPS.2018.02.000549",
  "title": "Modeling of Chemical Reacting Transport Phenomena in a PEM Fuel Cell using Finite Volume Method",
  "authors": ["Mohammed Jourdani", "H. Mounir", "A. Marjani"],
  "author_ids": [null, "31345194", "30962775"],
  "venue": "Progress in Petrochemical Science",
  "year": 2018,
  "fields_of_study": ["Engineering", "Chemistry", "Environmental Science", "Materials Science"],
  "publication_date": "2018-09-19",
  "journal_name": "Progress in Petrochemical Science",
  "license_publisher": "Crimson Publishers",
  "license": "cc-by",
  "oa_url": "http://crimsonpublishers.com/pps/pdf/PPS.000549.pdf",
  "text": "A two-dimensional transient model using finite volume method enhances PEMFC design by simulating gas flow and solving movement and energy equations.",
  "field_classification": ["Energy Chemistry", "Chemical Engineering"],
  "text_type": "summary",
  "embedding": [-0.0148, -0.0213, 0.0076 , -0.0111, 0.0357, 0.0141],
  "label": [9, 11]
}

Features

Field Name	Type	Description
`corpus_id`	`int`	Semantic Scholar ID
`doi`	`string`	DOI identifier
`title`	`string`	Title of the publication
`authors`	`list[string]`	Author names
`author_ids`	`list[string]`	Semantic Scholar author IDs
`venue`	`string`	Journal or conference name
`year`	`int`	Year of publication
`fields_of_study`	`list[string]`	Top-level categories (e.g., Engineering, Chemistry)
`publication_date`	`string`	ISO 8601 date
`journal_name`	`string`	Full journal name
`license_publisher`	`string`	OA license publisher
`license`	`string`	License type (e.g., cc-by)
`oa_url`	`string`	Open access link to full text
`text`	`string`	Abstract or summary
`text_type`	`string`	Indicates type of `text` field (e.g., summary)
`field_classification`	`list[string]`	Expert-curated subfield labels
`embedding`	`list[float]`	Dense vector embedding of the text
`label`	`list[int]`	Numeric labels for classification

Label Mapping

The dataset includes a consistent mapping between subfield names and numeric labels:

Subfield	Label
Catalysis	0
Organic Chemistry	1
Polymer Chemistry	2
Inorganic Chemistry	3
Materials Science	4
Analytical Chemistry	5
Physical Chemistry	6
Biochemistry	7
Environmental Chemistry	8
Energy Chemistry	9
Medicinal Chemistry	10
Chemical Engineering	11
Supramolecular Chemistry	12
Radiochemistry & Nuclear Chemistry	13
Forensic & Legal Chemistry	14
Food Chemistry	15
Chemical Education	16
Others	17

For convenience, here is the label_to_index mapping in JSON format:

{
    "label_to_index": {
        "Catalysis": 0,
        "Organic Chemistry": 1,
        "Polymer Chemistry": 2,
        "Inorganic Chemistry": 3,
        "Materials Science": 4,
        "Analytical Chemistry": 5,
        "Physical Chemistry": 6,
        "Biochemistry": 7,
        "Environmental Chemistry": 8,
        "Energy Chemistry": 9,
        "Medicinal Chemistry": 10,
        "Chemical Engineering": 11,
        "Supramolecular Chemistry": 12,
        "Radiochemistry & Nuclear Chemistry": 13,
        "Forensic & Legal Chemistry": 14,
        "Food Chemistry": 15,
        "Chemical Education": 16,
        "Others": 17
    },
    "index_to_label": {
        "0": "Catalysis",
        "1": "Organic Chemistry",
        "2": "Polymer Chemistry",
        "3": "Inorganic Chemistry",
        "4": "Materials Science",
        "5": "Analytical Chemistry",
        "6": "Physical Chemistry",
        "7": "Biochemistry",
        "8": "Environmental Chemistry",
        "9": "Energy Chemistry",
        "10": "Medicinal Chemistry",
        "11": "Chemical Engineering",
        "12": "Supramolecular Chemistry",
        "13": "Radiochemistry & Nuclear Chemistry",
        "14": "Forensic & Legal Chemistry",
        "15": "Food Chemistry",
        "16": "Chemical Education",
        "17": "Others"
    }
}

Usage

from datasets import load_dataset

dataset = load_dataset("Bocklitz-Lab/lit2vec-subfield-classifier-dataset")

print(dataset["train"][0]["title"])
print(dataset["train"][0]["field_classification"])

Each entry includes:

A machine-readable embedding
A list of labels (IDs)
The original text, metadata, and license info

Applications

🧠 Text classification using BERT, RoBERTa, etc.
🔍 Semantic search with sentence-transformer or FAISS
🧬 Domain adaptation for scientific NLP tasks
🧭 Clustering and unsupervised topic modeling
📈 Benchmarking embedding models on scientific literature

Licensing

All entries are sourced from CC BY 4.0 licensed publications.
Each entry includes original attribution via license_publisher, oa_url, and doi.
You are free to reuse, modify, and distribute the dataset under the terms of Creative Commons Attribution 4.0 International License.

Citation

If you use this dataset in your research, please cite:

@dataset{lit2vec_classifier_2025,
  author       = {Mahmoud Amiri, Thomas Bocklitz},
  title        = {Lit2Vec Subfield Classifier Dataset},
  year         = {2025},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/Bocklitz-Lab/lit2vec-subfield-classifier-dataset}},
  note         = {Submitted to Nature Scientific Data}
}

Acknowledgements

Built on top of Semantic Scholar Open Research Corpus (S2ORC)
Flattened and cleaned using custom preprocessing by the Bocklitz Lab
Embeddings generated from proprietary or publicly available models (details forthcoming in accompanying paper)

Datasets:

Bocklitz-Lab
/

lit2vec-subfield-classifier-dataset

You need to agree to share your contact information to access this dataset