metadata

dataset_name: lit2vec-tldr-bart-dataset
pretty_name: Lit2Vec TL;DR Chemistry Dataset
tags:
  - summarization
  - chemistry
  - scientific-papers
  - knowledge-graphs
license: cc-by-4.0
task_categories:
  - summarization
language:
  - en
size_categories:
  - 10K<n<100K

Lit2Vec TL;DR Chemistry Dataset

Summary

The Lit2Vec TL;DR Chemistry Dataset is a curated collection of 19,992 chemistry research abstracts paired with short, TL;DR-style abstractive summaries.
It was created to support research in scientific text summarization, semantic indexing, and domain-specific knowledge graph construction.

Unlike generic summarization datasets, this corpus is:

Legally reusable → all abstracts are sourced from CC-BY licensed publications
Domain-specific → chemistry and closely related fields (materials science, chemical engineering, environmental science, etc.)
Schema-consistent → each summary follows a concise “methods–results–significance” style

Dataset Structure

Splits

Train: 17,992 records
Validation: 999 records
Test: 1,001 records

Example Record

{
  "corpus_id": 250384325,
  "doi": "10.3390/biom12070947",
  "title": "Diffusion of Vanadium Ions in Artificial Saliva...",
  "authors": [
    "Sónia I. G. Fangaia",
    "A. M. Cabral",
    "P. Nicolau",
    "Fernando Guerra",
    "M. Rodrigo",
    "A. C. Ribeiro",
    "A. Valente",
    "M. A. Esteso"
  ],
  "venue": "Biomolecules",
  "year": 2022,
  "fields_of_study": ["Chemistry", "Medicine"],
  "publication_date": "2022-07-01",
  "abstract": "In this study, diffusion coefficients of ammonium vanadate ...",
  "summary": "The study measured diffusion coefficients of ammonium vanadate in artificial saliva ...",
  "license_type": "cc-by",
  "license_publisher": "MDPI AG",
  "license_url": "https://www.mdpi.com/2218-273X/12/7/947/pdf?version=1657252511",
  "source_url": "https://www.semanticscholar.org/paper/1ca8e174a4bc3eebcf0eb328e8582f0008a01c06"
}

Features

corpus_id (int): Semantic Scholar corpus ID
doi (string): Digital Object Identifier
title (string): Paper title
authors (list[string]): Author names
venue (string): Journal or conference
year (int): Publication year
fields_of_study (list[string]): Disciplinary categories
publication_date (string, ISO date): Publication date
abstract (string): Full research abstract
summary (string): TL;DR-style summary (target label)
license_type (string): License (always “cc-by”)
license_publisher (string): Publisher name
license_url (string): OA license link or PDF link
source_url (string): Semantic Scholar source page

Usage

from datasets import load_dataset

DATASET_ID = "Bocklitz-Lab/lit2vec-tldr-bart-dataset"

# Prefer the modern `token=` param; fall back to use_auth_token if your version needs it.
try:
    ds = load_dataset(
        DATASET_ID,
        split=None,                # get all splits if defined
        token=True,                # uses your cached HF login
        revision="main",           # avoid refs/convert/parquet unless we ask for it
        cache_dir="./.hf_cache_fresh",
        download_mode="force_redownload",
    )
except TypeError:
    # Older versions
    ds = load_dataset(
        DATASET_ID,
        split=None,
        use_auth_token=True,
        revision="main",
        cache_dir="./.hf_cache_fresh",
        download_mode="force_redownload",
    )

print(ds)
print(ds["train"][0]["abstract"])
print(ds["train"][0]["summary"])

Applications

Abstractive summarization training (BART, DistilBART, T5, LLaMA fine-tuning)
Information retrieval in chemistry and materials science
Knowledge graph population from structured summaries
Domain-specific semantic search engines

Licensing

All abstracts are sourced from CC BY 4.0 licensed publications.
Summaries are machine-generated and also distributed under CC BY 4.0.
Attribution information (publisher, OA URL, DOI) is included in the metadata for each record.

Citation

If you use this dataset, please cite:

@dataset{lit2vec_tldr_2025,
  author       = {Mahmoud Amiri, Thomas bocklitz},
  title        = {Lit2Vec TL;DR Chemistry Dataset},
  year         = {2025},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/Bocklitz-Lab/lit2vec-tldr-bart-dataset}},
  note         = {Submitted to Nature Scientific Data}
}

Acknowledgements

Built using Semantic Scholar Open Research Corpus (S2ORC)