BGE-M3 Latin Intertextuality Embedding Model (v2)

This model is a fine-tuned version of BAAI/bge-m3 for generating embeddings of Latin texts to detect intertextual relationships.

The model is part of the Loci Similes benchmark setup (Schelb et al., 2026), evaluated on expert-verified Latin intertextual links. It is designed to work with the LociSimiles Python package API: https://julianschelb.github.io/locisimiles/api/

Model Description

• Task: Sentence embedding for detecting intertextual links between classical Latin authors
• Model type: Sentence Transformer (Embedding Model)
• Base model: BAAI/bge-m3
• Loss: online contrastive loss
• Training data: one of five cross-validation splits of the Loci Similes benchmark
• Language: Latin
• License: Apache 2.0

Relation to v1

This replaces julian-schelb/bge-m3-emb-lat-intertext-v1, which was trained for the first version of the paper on an earlier revision of the dataset. The task and the interface are unchanged, so v2 is a drop-in replacement; v1 stays available for reproducing the earlier results.

Usage

This model was trained with prompt prefixes and should be used with the matching prompt names, otherwise retrieval quality drops noticeably:

• prompt_name="query" for query texts (Jerome), corresponding to a "Query: " prefix
• prompt_name="match" for candidate texts (classical authors), corresponding to a "Candidate: " prefix

from sentence_transformers import SentenceTransformer
from sentence_transformers.util import cos_sim

model = SentenceTransformer("julian-schelb/bge-m3-emb-lat-intertext-v2")

# Jerome text and candidates (1 positive match, 2 unrelated)
queries = [
    "omnia fert aetas, animum quoque; saepe ego longos cantando puerum memini me condere soles."
]
candidates = [
    "saepe ego longos cantando puerum memini me condere soles.",  # positive match
    "Gallia est omnis divisa in partes tres",                     # unrelated (Caesar)
    "in nova fert animus mutatas dicere formas",                  # unrelated (Ovid)
]

query_embeddings = model.encode(queries, prompt_name="query")
candidate_embeddings = model.encode(candidates, prompt_name="match")

scores = cos_sim(query_embeddings, candidate_embeddings)
print(scores)
print(f"Best candidate: {scores[0].argmax().item()}")

In a retrieval pipeline, embed the corpus once with prompt_name="match" and each query with prompt_name="query", then rank by cosine similarity. Retrieval is typically followed by a classification model — see the *-3class-lat-intertext-v1 classifiers in the same collection.

Citation

@misc{schelb2026locisimilesbenchmarkextracting,
      title={Loci Similes: A Benchmark for Extracting Intertextualities in Latin Literature},
      author={Julian Schelb and Michael Wittweiler and Marie Revellio and Barbara Feichtinger and Andreas Spitz},
      year={2026},
      eprint={2601.07533},
      archivePrefix={arXiv},
      primaryClass={cs.IR},
      url={https://arxiv.org/abs/2601.07533},
}