Dataset Research Field Classifier

Fine-tuned CPU-based classification of scientific datasets into research fields using OpenAlex topics taxonomy.

Overview

This classifier assigns a research field to scientific datasets based on their metadata using a fine-tuned embedding model. It maps datasets to the 4,516 topics in the OpenAlex taxonomy, along with their hierarchical subfield, field, and domain classifications. This was developed as part of our NIH S-index Challenge Phase 2 proposal. We refer to the S-index Hub for more information about our S-index and the Challenge.

Key features:

• 🚀 ~48,000 records/second on CPU with parallel processing (no GPU required)
• 📊 4,516 OpenAlex topics with full hierarchy
• 🎯 92.6% domain accuracy, 62.6% topic accuracy (fine-tuned model)
• 💻 Single dependency install — works on any machine

Quick Start

# Clone the repository (requires Git LFS for model files)
git lfs install
git clone https://github.com/data-S-index/dataset-to-field.git
cd dataset-to-field
git lfs pull  # Download model files (~140 MB)

# Install
pip install -e .

# Classify a single record
from openalex_classifier import TopicClassifier

classifier = TopicClassifier()
classifier.initialize()

result = classifier.classify({
    "title": "Climate change impact on marine ecosystems",
    "subjects": ["Climate", "Marine Biology", "Ecology"]
})

print(result)
# {
#   'topic': {'id': 1234, 'name': 'Marine Ecology and Climate Change', 'score': 0.72},
#   'subfield': {'id': 23, 'name': 'Ecology'},
#   'field': {'id': 5, 'name': 'Environmental Science'},
#   'domain': {'id': 2, 'name': 'Life Sciences'}
# }

Installation

This repository uses Git LFS to store model files. Make sure Git LFS is installed before cloning.

# 1. Install Git LFS (if not already installed)
git lfs install

# 2. Clone the repository
git clone https://github.com/data-S-index/dataset-to-field.git
cd dataset-to-field

# 3. Pull model files (~140 MB)
git lfs pull

# 4. Install the package
pip install -e .

Requirements

• Python 3.10+
• Git LFS (for model files)
• ~200MB disk space for models
• No GPU required

Usage

Python API

from openalex_classifier import TopicClassifier

# Initialize (loads model and topic embeddings)
classifier = TopicClassifier()
classifier.initialize()

# Single record
result = classifier.classify(record)

# Batch processing (faster)
results = classifier.classify_batch(records)

Command Line

# Classify NDJSON file
python -m openalex_classifier.cli input.ndjson output.ndjson

# With progress bar
python -m openalex_classifier.cli input.ndjson output.ndjson --progress

Input Format

Records should have at minimum a title field. Additional fields improve classification:

{
  "id": "10.5281/zenodo.123456",
  "title": "Dataset of marine temperature measurements",
  "subjects": ["Oceanography", "Climate Science"],
  "description": "Temperature readings from Pacific Ocean buoys..."
}

Output Format

{
  "dataset_id": "10.5281/zenodo.123456",
  "topic": {
    "id": 1234,
    "name": "Ocean Temperature and Climate Variability",
    "score": 0.68
  },
  "subfield": {"id": 23, "name": "Oceanography"},
  "field": {"id": 5, "name": "Earth and Planetary Sciences"},
  "domain": {"id": 2, "name": "Physical Sciences"}
}

Performance

Metric	Value
Throughput	48,304 records/sec (32 workers)
Total Records Classified	48.7 million
Processing Time	~15 minutes
Compressed Output	394 MB

Tested on 32-core AMD Threadripper with real DataCite metadata. Performance scales linearly with CPU cores.

Pre-classified Output

The complete classified DataCite dataset (46.2M records) is available for download:

• classified_output.zip (394 MB)

Validation Results

The fine-tuned model was validated against a held-out test set of 1,525 records with ground truth OpenAlex classifications:

Level	Accuracy	Description
Domain	92.6%	4 domains: Physical Sciences, Life Sciences, Social Sciences, Health Sciences
Field	85.8%	~26 fields: Chemistry, Medicine, Computer Science, etc.
Subfield	73.6%	~250 subfields: more specific research areas
Topic	62.6%	4,516 topics: granular research topics

Model Comparison

Model	Domain	Field	Subfield	Topic
Base (potion-32m)	77.2%	60.5%	27.9%	16.2%
Fine-tuned	92.6%	85.8%	73.6%	62.6%

The fine-tuned model achieves +15% improvement on domain accuracy and +46% improvement on exact topic matching compared to the base embedding model.

Training Data

The model was fine-tuned on 10,500 scientific records with ground truth topic classifications aligned with the OpenAlex taxonomy.

• Dataset: jimnoneill/dataset-to-field-training-10k
• Model: jimnoneill/dataset-to-field

Method

1. Text Extraction: Concatenate title + subjects/keywords + description from metadata
2. Semantic Embedding: Fine-tuned Model2Vec static embeddings
3. Topic Matching: Cosine similarity against 4,516 pre-embedded topics
4. Hierarchical Output: Return topic → subfield → field → domain

Topic Hierarchy

Domain (4)
└── Field (26)
    └── Subfield (254)
        └── Topic (4,516)

Example:

• Domain: Physical Sciences
• Field: Computer Science
• Subfield: Artificial Intelligence
• Topic: Natural Language Processing

Citation

If you use this classifier, please cite:

@software{dataset-to-field,
  author = {O'Neill, James, Patel, Bhavesh},
  title = {Dataset Research Field Classifier},
  year = {2026},
  url = {https://github.com/data-S-index/dataset-to-field}
}

License

MIT License - see LICENSE for details.

Acknowledgments

• OpenAlex for the topic taxonomy
• minishlab/potion-base-32m base embedding model
• Model2Vec for model distillation and training