Bocklitz-Lab/Lit2Vec-dataset

🧪 Lit2Vec Dataset

Lit2Vec is a large-scale, openly licensed dataset of 582,724 chemistry-specific research articles derived from the Semantic Scholar Open Research Corpus (S2ORC), curated to support semantic search, retrieval-augmented generation (RAG), and scientific NLP tasks in the chemistry domain.

Each article is distributed as a standalone JSON file containing:

• Clean full text
• Paragraph-level embeddings (1024-D vectors)
• TL;DR summaries
• Subfield classification scores
• Structured metadata
• Licensing information from Unpaywall, Crossref, and OpenAlex

---

📁 Repository Structure

Lit2Vec-dataset/
├── data/         # Full dataset split into multiple tar.gz archives (~4–5 GB each)
│   ├── data_part_000.tar.gz
│   ├── ...
│
├── sample/       # Small sample subset (~500MB) for quick testing
│   └── data_part_000.tar.gz
│
├── validation/   # Lightweight validation set (~15–18 MB shards)
│   ├── data_part_000.tar.gz
│   ├── ...
│
├── .gitattributes
└── .gitignore

Each .tar.gz archive contains thousands of individual JSON files, named by their S2ORC corpus_id, for example:

12345678.json
12345679.json
...

---

📄 Data Record Format

Each JSON file contains a single research article, structured according to the Lit2Vec v1.0 schema.

🔑 Top-Level Fields

Field	Type	Description
schema_version	string	Schema version ("1.0")
corpus_id	integer	S2ORC Corpus ID (used as filename)
metadata	object	Title, authors, year, venue, identifiers
abstract	string	Abstract text
fulltext	string	Full article text
paragraphs	array[string]	Paragraph-level text units
embeddings	array[array[float]]	1024-D float32 paragraph vectors (aligned with paragraphs)
abstract_embedding	array[float]	1024-D float32 vector for abstract
predicted_subfield	object	Map of subfield label → confidence score
tldr	string	Machine-generated two-sentence summary
unpaywall_license	object | null	License data from Unpaywall
crossref_license	object | null	License data from Crossref
openalex_license	object | null	License data from OpenAlex

📦 Minimal Example

{
  "schema_version": "1.0",
  "corpus_id": 37254803,
  "metadata": {
    "title": "Protective effect of EGCG...",
    "year": 2016,
    "externalids": { "DOI": "10.xxxx/xxxxx" },
    "url": "https://..."
  },
  "abstract": "Epigallocatechin gallate ...",
  "fulltext": "# Protective effect of ...",
  "paragraphs": ["37254803P0: passage: ...", "..."],
  "embeddings": [[0.0123, -0.0456, ...], ...],
  "abstract_embedding": [0.0345, -0.0789, ...],
  "predicted_subfield": {
    "Biochemistry": 0.997,
    "Medicinal Chemistry": 0.759
  },
  "tldr": "EGCG reduces lipid peroxidation ...",
  "unpaywall_license": {
    "best_oa_location": {
      "license": "cc-by"
    }
  },
  "crossref_license": {
    "license": "http://creativecommons.org/licenses/by/4.0/"
  },
  "openalex_license": null
}

---

🧠 Use Cases

Lit2Vec is optimized for:

• Semantic search — FAISS-style dense retrieval with paragraph embeddings
• Retrieval-Augmented Generation (RAG) — grounding LLMs on full-text chemistry papers
• Summarization — pre-generated TL;DRs for training and evaluation
• Multi-label classification — 18 chemistry subfields per article with confidence scores
• Citation & licensing analysis — metadata from Unpaywall, Crossref, OpenAlex
• Knowledge graph construction — integrate with PubChem, ChEMBL, patents, etc.

---

🚀 Quickstart

Python (manual extraction)

from datasets import load_dataset

dataset = load_dataset(
     "Bocklitz-Lab/Lit2Vec-dataset",
     data_files={"train": "data/*.tar.gz"},
     trust_remote_code=True,)

# Peek at one example
sample = dataset["train"][0]
print(sample["metadata"])  # Raw JSON string with title, authors, etc.
print(sample["predicted_subfield"])  # List of {label, score}

# Extract title from metadata
import json
metadata = json.loads(sample["metadata"])
print("Title:", metadata.get("title"))

# Display subfield predictions
print("Predicted subfields:")
for sub in sample["predicted_subfield"]:
    print(f"  {sub['label']}: {sub['score']:.3f}")

---

✅ Validation Subset

The /validation/ folder contains .tar.gz archives to:

• Test schema compliance
• Integrate with pipelines
• Validate embeddings and metadata formats

---

🪪 Licensing

Lit2Vec is fully open-access, curated under FAIR principles. All included articles passed license validation via:

• ✅ Unpaywall
• ✅ Crossref
• ✅ OpenAlex

Permissible licenses include:

• CC-BY
• CC-BY-SA
• CC-BY-NC
• CC-BY-NC-SA
• Public Domain

---

📖 Citation

Please cite the following if you use this dataset:

@article{amiri2025lit2vec,
  title={Lit2Vec: A Large-Scale, Derivative-Permissive Chemistry Corpus for Retrieval-Augmented Generation and Semantic Search},
  author={Amiri, Mahmoud and Bocklitz, Thomas},
  journal={Preprint},
  year={2025},
  note={\url{https://huggingface.co/datasets/Bocklitz-Lab/Lit2Vec-dataset}}
}

---

🤝 Contributing

We welcome community contributions for:

• Dataset converters (e.g., Hugging Face datasets format)
• Example notebooks and pipelines
• Benchmarking scripts

Open a pull request or contact us!

---

Lit2Vec is developed at the Leibniz Institute of Photonic Technology and the Institute of Physical Chemistry, Friedrich Schiller University Jena to accelerate open science in chemistry and AI.