The model was fine-tuned using Multiple Negatives Ranking Loss to learn meaningful representations that capture semantic relationships between biblical Greek texts.

Strengths

This model excels at:

Variant Detection: High similarity scores (>0.9) for verses that are textually identical or near-identical, even with minor orthographic differences
Semantic Clustering: Effectively groups related passages and parallel texts
Robust to Spelling Variations: Handles common manuscript variations in biblical Greek

Usage

Installation & Training

For installation instructions and training scripts, see the GitHub repository.

Inference Example

python

import unicodedata
from sentence_transformers import SentenceTransformer, util


def strip_accents_and_lowercase(s):
    return "".join(
        c for c in unicodedata.normalize("NFD", s) if unicodedata.category(c) != "Mn"
    ).lower()


xsent = [
    "οι δε φαρισεοι ακουσαντες οτι εφιμωσε τους σαδδουκεους συνηχθησαν επι το αυτο",
    "ειπεν δε αυτοις οταν προσευχησθε λεγετε πατερ αγιασθητω το ονομα σου ελθατω η βασιλια σου γενηθητω το θελημα σου ως εν ουρανω και επι γης",
    "νυν κρισις εστιν του κοσμου νυν ο αρχων τουτου τουτου νυν ο αρχων του κοσμου τουτου εκβληθησεται εξω",
    "διο προσλαμβανεσθαι αλληλους καθως και ο χς προσελαβετο υμας εις δοξαν του θυ",
]

ysent = [
    "οι δε φαρισαιοι ακουσαντες οτι εφιμωσε τους σαδδουκεους συνηχθησαν επι το αυτο",
    "ειπεν δε αυτοις οταν προσευχησθε λεγετε πατερ αγιασθητω το ονομα σου ελθατω η βασιλια σου γενηθητω το θελημα σου ως εν ουρανω και επι γης και ρυσαι ημας απο του πονηρου",
    "νυν δε προς σε ερχομαι και ταυτα λαλω εν τω κοσμω ινα εχωσιν την χαραν την εμην πεπληρωκενην εν αυτοις",
    "μετανοησαται ουν και επιστρεψαται προς το εξαλιφθηναι υμων τας αμαρτιας",
]

xsent_norm = [strip_accents_and_lowercase(s) for s in xsent]
ysent_norm = [strip_accents_and_lowercase(s) for s in ysent]

model = SentenceTransformer("Paulanerus/AncientGreekVariantSBERT")

x_embeddings = model.encode(xsent_norm, convert_to_tensor=True)
y_embeddings = model.encode(ysent_norm, convert_to_tensor=True)

print("Similarities:")
for i in range(len(xsent_norm)):
    similarity = util.cos_sim(x_embeddings[i], y_embeddings[i]).item()
    print(f"Pair {i + 1}: {similarity:.4f}")

Expected output:

Similarities:
Pair 1: 0.9882  # Near-identical verses (minor spelling difference)
Pair 2: 0.9000  # Same verse, one with additional text
Pair 3: 0.1772
Pair 4: 0.1724

Training Details

Base Model

Base: pranaydeeps/Ancient-Greek-BERT
Architecture: BERT-base (12 layers, 768 hidden dimensions)

Training Configuration

Parameter	Value
Batch Size	256
Epochs	8
Learning Rate	2e-5
Loss Function	MultipleNegativesRankingLoss
Warmup	10% of training steps
Hardware	NVIDIA A100 80GB PCIe

Preprocessing

All input text should be normalized by:

Removing diacritics/accents (NFD normalization)
Converting to lowercase

Evaluation Results

Evaluated on information retrieval task:

Metric	Score
Accuracy@1	0.43
Accuracy@3	1.00
Accuracy@5	1.00
Precision@3	0.74
Precision@10	0.88
MRR@10	0.715
NDCG@10	0.843
MAP@100	0.911

Limitations

Optimized specifically for Ancient Greek; may have reduced performance on other genres (Classical, Homeric, etc.)
Requires text preprocessing (accent stripping, lowercasing) for best results

Citation

bibtex

@misc{ancient-greek-variant-sbert,
  author = {Fröhlich, Paul},
  title = {Ancient Greek Variant SBERT: Fine-tuned Embeddings for Biblical text verses in Ancient Greek},
  year = {2026},
  howpublished = {\url{https://huggingface.co/Paulanerus/AncientGreekVariantSBERT}},
  note = {Model release}
}

Acknowledgments

This model builds upon the Ancient Greek BERT by Singh, Rutten, and Lefever (2021).

This work was funded by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) 513300936.