multilingual-e5-small-RoD-TAL

Fine-tuned sentence embedding model for Romanian legal retrieval in the driving-license domain.

• Base model: intfloat/multilingual-e5-small
• Fine-tuned model: unstpb-nlp/multilingual-e5-small-RoD-TAL
• Dataset: unstpb-nlp/RoD-TAL
• Primary task: dense retrieval of Romanian legal references for exam-style questions

Model details

This checkpoint is domain-adapted for the RoD-TAL benchmark (Romanian driving-law QA/retrieval). It is intended for retrieval and RAG pipelines where legal passages must be matched to user questions.

Setup highlights:

• Contrastive training (MultipleNegativesRankingLoss / InfoNCE)
• Positive pairs: question ↔ correct legal references
• Hard negatives mined from top candidates of a base retriever
• Document encoding with 512-token truncation

Usage

from sentence_transformers import SentenceTransformer

model = SentenceTransformer("unstpb-nlp/multilingual-e5-small-RoD-TAL")

queries = [
    "query: În ce situație trebuie să acord prioritate pietonilor?"
]
passages = [
    "passage: Art. X - Conducătorul este obligat să acorde prioritate pietonilor angajați regulamentar în traversare.",
    "passage: Art. Y - ..."
]

q_emb = model.encode(queries, normalize_embeddings=True)
p_emb = model.encode(passages, normalize_embeddings=True)

scores = (q_emb @ p_emb.T)[0]
print(scores)

Query formatting recommendation

This is an E5-family model, so using explicit query: / passage: prefixes is recommended in inference.

For RoD-TAL-style retrieval, best reported performance used question + answer options as query text (concatenated), rather than question-only.

Intended uses

• Legal article retrieval in Romanian traffic-law assistants
• Retrieval stage for Romanian legal RAG systems
• Domain-specific benchmarking on RoD-TAL IR/VIR tasks

Limitations

• Truncation at 512 tokens can reduce recall for long legal documents
• Performance is domain-focused (Romanian driving-law); transfer to other legal domains may drop
• Reported gains depend on retrieval protocol and corpus preparation

Training and evaluation summary

Reported IR test metrics for the fine-tuned retriever on RoD-TAL:

• Recall@10: 88.14
• Precision@10: 23.28
• nDCG@10: 81.41

Citation

If you use this model, please cite:

@misc{man2025rodtalbenchmarkansweringquestions,
  title={RoD-TAL: A Benchmark for Answering Questions in Romanian Driving License Exams},
  author={Andrei Vlad Man and Răzvan-Alexandru Smădu and Cristian-George Craciun and Dumitru-Clementin Cercel and Florin Pop and Mihaela-Claudia Cercel},
  year={2025},
  eprint={2507.19666},
  archivePrefix={arXiv},
  primaryClass={cs.CL},
  url={https://arxiv.org/abs/2507.19666}
}