This is a Latin sentence-transformers model finetuned from bowphs/LaBerta. It maps Latin sentences and paragraphs to a dense vector space and can be used for semantic correspondances detection or information retrieval (e.g. for detection of quotations and allusions).

Model Details

Model Description

Model Type: Sentence Transformer
Base Model: bowphs/LaBerta
Maximum Sequence Length: 512 tokens
Output Dimensionality: 768 dimensions
Similarity Function: cosine
Pooling Mode: cls
Supported Modality: Text

Model Sources

Documentation: https://sbert.net
Repository: https://github.com/huggingface/sentence-transformers
Hugging Face: https://huggingface.co/models?library=sentence-transformers

Full Model Architecture

SentenceTransformer(
  (0): Transformer(transformer_task=feature-extraction, architecture=RobertaModel)
  (1): Pooling(embedding_dimension=768, pooling_mode=cls)
)

Usage

Direct Usage (Sentence Transformers)

pip install -U sentence-transformers

python

from sentence_transformers import SentenceTransformer

model = SentenceTransformer("Tdelaselle/PatriLaSE")
sentences = [
    "quis ergo sanat omnes languores tuos nisi qui propitius fit omnibus iniquitatibus tuis?",
    "qui propitiatur omnibus iniquitatibus tuis qui sanat omnes infirmitates tuas",
]
embeddings = model.encode(sentences)
print(embeddings.shape)

Training Details

Training Dataset

Size: 183227 training samples
Corpus Path: Patristic Latin sentences
Loss: MaskedDenoisingAutoEncoderLoss

Training Hyperparameters

epochs: 3
batch_size: 16
learning_rate: 2e-05
weight_decay: 0.0
warmup_ratio: 0.0

Framework Versions

pytorch: 2.11.0+cu130
sentence_transformers: 5.4.1
transformers: 4.57.6

TdelaSelle/PatriLaSE

Patristic Latin Sentence Embeddings