MatPROV

MatPROV is a dataset of materials synthesis procedures extracted from scientific papers using large language models (LLMs) and represented in PROV-DM–compliant structures. Further details on MatPROV are described in our paper "MatPROV: A Provenance Graph Dataset of Material Synthesis Extracted from Scientific Literature.”

---

Files

MatPROV/
├── MatPROV.jsonl  # Main dataset (2,367 synthesis procedures)
├── ground-truth/  # Expert-annotated ground truth
│ └─ <DOI>.json
├── few-shot/.     # Prompt examples used for synthesis procedure extraction
│ └─ <DOI>.txt
└── doi_status.csv # Status of each paper DOI across the pipeline

Note: In file names under ground-truth/ and few-shot/, forward slashes (/) in DOIs are replaced with underscores (_).

---

Data format

The main dataset file is MatPROV.jsonl, where each line corresponds to one paper’s structured record. Each record contains:

• doi: DOI of the source paper
• label: Identifier for the extracted synthesis procedure, encoding the material's chemical composition and key synthesis characteristics (e.g., CuGaTe2_ball-milling)
• prov_jsonld: A PROV-JSONLD structure describing the synthesis procedure

Example

{
  "doi": "10.1002/advs.201600035",
  "label": "Fe1+xNb0.75Ti0.25Sb_composition variation",
  "prov_jsonld": {
    "@context": [
      {"xsd": "http://www.w3.org/2001/XMLSchema#", "prov": "http://www.w3.org/ns/prov#"},
      "https://openprovenance.org/prov-jsonld/context.jsonld",
      "URL of MatPROV's context schema omitted for double-blind review"
    ],
    "@graph": [
      {
        "@type": "Entity",
        "@id": "e1",
        "label": [{"@value": "Fe", "@language": "EN"}],
        "type": [{"@value": "material"}],
        "matprov:purity": [{"@value": "99.97%", "@type": "xsd:string"}]
      }
      ...
    ]
  }
}

Visualization

You can visualize the PROV-JSONLD data in MatPROV using the online tool at: https://matprov-project.github.io/prov-jsonld-viz/ To do this, copy the value of the "prov_jsonld" field from any record in MatPROV.jsonl and paste it into the “PROV-JSONLD Editor” panel of the tool. A directed graph of the synthesis procedure will then be generated, as shown in the figure below.

Dataset construction summary

• Source papers collected: 1648
• Relevant Text Extraction
  • 32 papers contained no synthesis-related text
  • → 1616 papers remained
• Synthesis Procedure Extraction
  • 48 papers contained no synthesis procedure
  • → 1568 papers remained (final dataset)

The DOIs of these 1568 papers and their extracted data are included in MatPROV.jsonl. For details on the filtering status of each DOI, see doi_status.csv.

Ground Truth annotations

• A subset of papers was manually annotated by a single domain expert.
• Files are stored in ground-truth/ and named as <DOI>.json.

Few-shot examples

• Prompt examples used for LLM extraction are provided in few-shot/.
• Files are stored in few-shot/ and named as <DOI>.txt.

Links

• Paper: https://arxiv.org/abs/2509.01042
• Code: https://github.com/MatPROV-project/matprov-experiments

Citation

If you use MatPROV, please cite:

@inproceedings{tsuruta2025matprov,
  title={Mat{PROV}: A Provenance Graph Dataset of Material Synthesis Extracted from Scientific Literature},
  author={Hirofumi Tsuruta and Masaya Kumagai},
  booktitle={NeurIPS 2025 Workshop on AI for Accelerated Materials Design},
  year={2025}
}