---
license: apache-2.0
language: en
tags:
  - sentence-transformers
  - citation-recommendation
  - academic
  - feature-extraction
library_name: sentence-transformers
pipeline_tag: sentence-similarity
datasets:
  - custom
---

# MiniLM Citation v4

A sentence-transformer model fine-tuned for academic citation recommendation. Given a passage of academic writing, this model finds the most relevant papers to cite.

## Model Details

- **Base model**: [microsoft/MiniLM-L6-v2](https://huggingface.co/microsoft/MiniLM-L6-v2) (via all-MiniLM-L6-v2)
- **Dimensions**: 384 (Matryoshka: 128/256/384)
- **Training**: CachedMultipleNegativesRankingLoss with hard negatives
- **Training data**: 64K citation context → cited paper pairs from academic papers

## Performance

Evaluated on a benchmark of 3,420 citation contexts across 100 source papers:

| Method | MRR | R@1 | R@10 |
|--------|-----|-----|------|
| MiniLM-FT v4 (neural) | 0.400 | 27.5% | 72.0% |
| MiniLM-FT v4 (hybrid) | 0.428 | 30.2% | 74.8% |
| Cloud model (hybrid) | 0.550 | 38.5% | 88.5% |

For best results, use with [inCite](https://github.com/galenphall/incite) in hybrid mode (neural + BM25 fusion).

## Usage

```python
from sentence_transformers import SentenceTransformer

model = SentenceTransformer("galenphall/minilm-citation-v4")

# Embed a query (your writing)
query_embedding = model.encode("The relationship between CO2 emissions and temperature...")

# Embed papers as: "title. authors. year. journal. abstract"
paper_embedding = model.encode("Global Warming Effects. Smith and Jones. 2023. Nature. We study...")
```

Or use with inCite directly:

```bash
pip install incite
incite setup
incite recommend "your text here" -k 10
```

## Citation

```bibtex
@software{incite2025,
  author = {Hall, Galen},
  title = {inCite: Local-First Citation Recommendation},
  year = {2025},
  url = {https://github.com/galenphall/incite},
  license = {Apache-2.0}
}
```