README.md

---
language: en
license: mit
library_name: model2vec
tags:
  - model2vec
  - static-embeddings
  - text-classification
  - topic-classification
  - scientific-classification
  - s-index
  - dataset-classification
datasets:
  - jimnoneill/dataset-to-field-training-10k
base_model: minishlab/potion-base-32m
metrics:
  - accuracy
model-index:
  - name: dataset-to-field
    results:
      - task:
          type: text-classification
          name: Research Field Classification
        dataset:
          name: Dataset Research Field Training Data
          type: jimnoneill/dataset-to-field-training-10k
        metrics:
          - type: accuracy
            value: 0.926
            name: Domain Accuracy
          - type: accuracy
            value: 0.858
            name: Field Accuracy
          - type: accuracy
            value: 0.736
            name: Subfield Accuracy
          - type: accuracy
            value: 0.626
            name: Topic Accuracy
---

# Dataset Research Field Classifier

A fine-tuned static embedding model for classifying scientific datasets into research fields. It maps datasets to the 4,516 topics in the [OpenAlex taxonomy](https://docs.openalex.org/api-entities/topics), 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](https://github.com/data-S-index/hub) for more information about our S-index and the Challenge.

## Model Description

This model is fine-tuned from [minishlab/potion-base-32m](https://huggingface.co/minishlab/potion-base-32m) on ground truth topic classifications aligned with the [OpenAlex topics taxonomy](https://docs.openalex.org/api-entities/topics). It uses Model2Vec's static embedding approach for fast, efficient inference without requiring a GPU.

### Topic Hierarchy

The classifier uses a 4-level hierarchical classification system based on the OpenAlex topics taxonomy:
- **4 Domains**: Physical Sciences, Life Sciences, Social Sciences, Health Sciences
- **~26 Fields**: Chemistry, Physics, Medicine, Computer Science, etc.
- **~250 Subfields**: More specific research areas
- **4,516 Topics**: Granular research topics

## Performance

Evaluated on a held-out test set (1,525 samples):

| Level | Accuracy |
|-------|----------|
| Domain | **92.6%** |
| Field | **85.8%** |
| Subfield | **73.6%** |
| Topic (exact) | **62.6%** |

### Comparison with Base Model

| 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%** |
| Improvement | +15.4 | +25.3 | +45.7 | +46.4 |

## Usage

```python
from model2vec import StaticModel
import numpy as np

# Load model
model = StaticModel.from_pretrained("jimnoneill/dataset-to-field")

# Prepare your text
text = "Machine learning approaches for protein structure prediction using deep neural networks"

# Get embedding
embedding = model.encode([text])

# For full classification pipeline, see:
# https://github.com/data-S-index/dataset-to-field
```

## Training Details

- **Base Model**: minishlab/potion-base-32m
- **Training Samples**: 8,636
- **Test Samples**: 1,525
- **Topics Covered**: 1,135 / 4,516 (25%)
- **Training Time**: ~2 minutes on RTX 4090
- **Framework**: Model2Vec + PyTorch Lightning

## Training Data

The model was trained on ground truth topic classifications derived from the [OpenAlex topics taxonomy](https://docs.openalex.org/api-entities/topics). The training data includes scientific records with titles, subjects, and descriptions mapped to specific topics.

Dataset: [jimnoneill/dataset-to-field-training-10k](https://huggingface.co/datasets/jimnoneill/dataset-to-field-training-10k)

## Intended Use

- Classifying scientific publications, datasets, and research outputs into research fields
- Mapping DataCite records to research topics
- Bibliometric analysis and research trend identification
- Integration with the S-Index scientific indexing pipeline

## Limitations

- Trained on ~25% of the 4,516 topics; rare topics may have lower accuracy
- Domain distribution is skewed toward Physical Sciences (53%)
- Best suited for English-language scientific content

## Citation

If you use this model, please cite:

```bibtex
@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

## Acknowledgments

- [OpenAlex](https://openalex.org/) for the topic taxonomy
- [minishlab/potion-base-32m](https://huggingface.co/minishlab/potion-base-32m) base embedding model
- [Model2Vec](https://github.com/MinishLab/model2vec) for model distillation and training