aimgo/CorpusTrinum

---
license: cc-by-4.0
task_categories:
- sentence-similarity
- text-classification
- text-retrieval
language:
- la
size_categories:
- 10M<n<100M
---

<img src="https://cdn-uploads.huggingface.co/production/uploads/62cf05b026c94b143172379c/G1DxNSWgcnCPPszzLCeKG.png"
style="float:left;width:200px;height:200px;object-fit:cover;border-radius:50%;margin-right:16px;" />

A corpus containing both weak and strong Latin semantic triplets drawn from [Vicipaedia](https://la.wikipedia.org/) and the [Rosenthal Latin–English Parallel Dataset](https://huggingface.co/datasets/grosenthal/latin_english_parallel).

Weak triplets were calculated by calculating TF-IDF scores over all sentences in Vicipaedia. For each sentence, the most similar inexact match was selected as the positive sample and the most dissimilar match was selected as the negative sample.

Strong triplets were calculated by generating ModernBERT embeddings for each English translation in the Rosenthal Latin-English Parallel Dataset. A FAISS index to support positive retrieval, another FAISS index was constructed using the inverse of the distance metric for negative retrieval. Again, the positive sample was the most similar inexact match and the most dissimilar match was the negative sample.
This approach leverages high model performance on English language semantic similarity to inform the Latin equivalent, creating the first resource of its kind for Latin NLP.

If you use this in your work, please cite:
```
@misc{mccarthy2025trinumcorpus,
author = {McCarthy, A. M.},
title = {{CorpusTrinum}: A Latin Semantic Contrastive Triplet Dataset},
year = {2025},
howpublished = {\url{https://huggingface.co/datasets/aimgo/trinum-corpus}},
note = {Dataset}
}
```