Theologian Embedder (theologian-embedder-chirho)

A fine-tuned MiniLM-L12-v2 sentence transformer that creates a theological embedding space, clustering orthodox statements together and separating them from heterodox ones.

Part of the Theological Guardrails Pipeline.

How It Works

Trained via contrastive learning (triplet loss):

• Anchor: Orthodox theological statement
• Positive: Similar orthodox statement
• Negative: Heterodox / heretical statement

The resulting embedding space groups orthodox statements together while pushing heterodox statements far away.

Evaluation Results

Metric	Value
Orthodox-Orthodox avg similarity	0.8881
Heterodox-Heterodox avg similarity	0.6126
Orthodox-Heterodox avg similarity	-0.7174
Separation gap	1.6055
Pearson correlation	0.970

Usage

from sentence_transformers import SentenceTransformer
import numpy as np

model = SentenceTransformer("loveJesus/theologian-embedder-chirho")

orthodox = model.encode("Christ is fully God and fully man, two natures in one person.")
heretical = model.encode("Jesus was merely a created being.")

similarity = np.dot(orthodox, heretical) / (np.linalg.norm(orthodox) * np.linalg.norm(heretical))
print(f"Similarity: {similarity:.3f}")  # Should be low/negative

Use Cases

• Zero-shot heresy detection: Compare new statements against orthodox centroid
• Semantic search: Find theologically similar statements
• Clustering: Group theological positions by similarity
• Anomaly detection: Flag statements that are far from known orthodox/heterodox clusters

Architecture

• Base model: sentence-transformers/all-MiniLM-L12-v2
• Training: Triplet loss with online hard mining
• Embedding dimension: 384
• Training data: ~22,500 theological statements from the theologian-dataset-chirho

Orthodoxy Basis

First six ecumenical councils (Nicaea I through Constantinople III).

Limitations

• Trained on English theological texts only
• Best at detecting heresies covered in the training data
• Embedding similarity is a signal, not a definitive classification
• Should be used alongside the classifier and explainer for best results

Related Models

Model	Task
theologian-classifier-chirho	Multi-label heresy classification (RoBERTa-large)
theologian-explainer-chirho	Explanation generation (Flan-T5-base)
theologian-dataset-chirho	Training dataset

License

MIT

Citation

@misc{lovejesus2026theologianembedder,
  title={Theologian Embedder: Contrastive Theological Embedding Space},
  author={loveJesus},
  year={2026},
  publisher={HuggingFace},
  url={https://huggingface.co/loveJesus/theologian-embedder-chirho}
}