amayuelas/arxiv-cond-matter

arXiv cond-matter — Materials Synthesis Recipes

A dataset of 56,943 materials-synthesis recipes extracted from condensed-matter (cond-mat) papers on arXiv. Each record links a paper to a structured synthesis recipe describing the target material, the starting materials, and a step-by-step procedure, automatically extracted from the paper's full text by a large language model.

Dataset summary

• Records: 56,943 (one synthesis recipe per paper)
• Source: arXiv condensed-matter (cond-mat) paper metadata + full text
• Recipes with a DOI: 40,544
• Extraction model: google/gemma-4-31B-it
• Language: English
• Format: JSON Lines (arxiv_cond_matter.jsonl)

Material categories

Category	Count
Nanomaterials	14,115
Ceramics	10,473
Superconductors	7,686
Semiconductors	7,670
Magnetic	5,974
Metals	3,591
Chemicals	1,697
Polymers	1,660
Composites	1,270
N/A	1,030
Other	961
Biomaterials	800
Textiles	15
Amorphous Metals	1

Data fields

Each line is a JSON object with the following fields:

Field	Type	Description
title	string	Title of the source arXiv paper.
authors	list[string]	Author names as "First Last".
doi	string	DOI of the published version (may be empty).
paper_link	string	URL to the arXiv abstract page.
sample_name	string	Identifier "<final_material_formula>_<doi>".
starting_materials	list[string]	Starting reagents as "name (formula)".
starting_materials_structured	list[object]	Starting reagents as {name, formula} objects.
final_material	string	Target material as "name (formula)".
final_material_structured	object	Target material as {name, formula}.
keywords	list[string]	Synthesis / material keywords.
category	string	Material category (see table above).
recipe_text	string	Structured, human-readable synthesis recipe (target material, reagents, environment, equipment, procedure, notes).
recipe_llm_model	string	Model used to extract the recipe.
published	bool	Whether the paper has a published (peer-reviewed) version.

Example

{
  "title": "Langmuir blodgett assembly of densely aligned single walled carbon nanotubes from bulk materials",
  "authors": ["Xiaolin Li", "Li Zhang", "Xinran Wang", "..."],
  "doi": "10.1021/ja071114e",
  "paper_link": "https://arxiv.org/abs/0704.0113",
  "sample_name": "C_10.1021/ja071114e",
  "starting_materials": ["Single-walled carbon nanotubes (C)", "1,2-dichloroethane (C2H4Cl2)"],
  "final_material": "Aligned Single-Walled Carbon Nanotube Monolayers (C)",
  "category": "Nanomaterials",
  "keywords": ["Nanomaterials", "Langmuir-Blodgett", "Single-Walled Carbon Nanotubes"],
  "recipe_text": "## Target Material:\n    Name: Aligned Single-Walled Carbon Nanotube Monolayers\n    ...",
  "recipe_llm_model": "google/gemma-4-31B-it",
  "published": true
}

Usage

from datasets import load_dataset

ds = load_dataset("amayuelas/arxiv-cond-matter", split="train")
print(ds[0]["recipe_text"])

How it was built

1. Condensed-matter (cond-mat) paper metadata was collected from arXiv.
2. Paper PDFs were parsed to full text.
3. A structured synthesis recipe (target material, starting materials, environment, equipment, procedure) was extracted from each paper using google/gemma-4-31B-it.
4. Recipes were matched back to their paper metadata to produce one record per paper.

Limitations

• Recipes are LLM-extracted and may contain extraction errors, hallucinated details, or missing steps. They have not been exhaustively verified against the source papers.
• Chemical formulas and quantities may be incomplete (N/A) when not stated in the source text.
• Coverage is limited to papers available on arXiv cond-mat that contained an identifiable synthesis procedure.

License

Released under CC-BY-4.0. The underlying arXiv metadata and papers retain their original licenses; please cite the original papers (via the provided doi / paper_link) when using individual recipes.