RoD-TAL

RoD-TAL is a multimodal benchmark for Romanian driving-license exam tasks, introduced in:

RoD-TAL: A Benchmark for Answering Questions in Romanian Driving License Exams
Findings of EACL 2026, arXiv: 2507.19666

The dataset supports four task families:

• IR (Information Retrieval): retrieve relevant legal articles
• QA (Question Answering): answer text-only driving-exam questions
• VIR (Visual Information Retrieval): retrieve legal references and indicators from visual questions
• VQA (Visual Question Answering): answer image-grounded driving-exam questions

What is included

This dataset repository contains:

• Question splits (split_1 to split_4): exam-style queries with metadata and answer options
• Legal corpus (corpus_laws): article-level Romanian traffic-law passages
• Traffic-sign corpus (corpus_traffic_signs): sign metadata and sign images
• Qrels for legal retrieval (qrels_laws): query-document relevance labels
• Qrels for sign retrieval (qrels_traffic_signs): query-sign relevance labels
• Images (images/questions, images/traffic_signs): visual assets referenced by multimodal splits

Split overview

• split_1: text QA/IR split with train and test subsets
• split_2: additional text QA/IR split
• split_3: multimodal split (includes traffic_signs and image field)
• split_4: multimodal split (includes traffic_signs and image field)

For qrels_laws, the current release includes non-empty splits: qrels_queries_split_1, qrels_queries_split_1_train, qrels_queries_split_1_test, and qrels_queries_split_3.

Usage

from datasets import load_dataset

# Text-only splits (train/test available for split_1)
queries_train = load_dataset("unstpb-nlp/RoD-TAL", "split_1", split="split_1_train")
queries_test = load_dataset("unstpb-nlp/RoD-TAL", "split_1", split="split_1_test")

# Additional text split
queries_split_2 = load_dataset("unstpb-nlp/RoD-TAL", "split_2", split="split_2_all")

# Legal corpus
corpus_laws = load_dataset("unstpb-nlp/RoD-TAL", "corpus_laws", split="corpus")

# Retrieval labels for legal IR
qrels_laws_train = load_dataset("unstpb-nlp/RoD-TAL", "qrels_laws", split="qrels_queries_split_1_train")
qrels_laws_test = load_dataset("unstpb-nlp/RoD-TAL", "qrels_laws", split="qrels_queries_split_1_test")

# Visual/multimodal splits
queries_visual_split_3 = load_dataset("unstpb-nlp/RoD-TAL", "split_3", split="split_3_all")
queries_visual_split_4 = load_dataset("unstpb-nlp/RoD-TAL", "split_4", split="split_4_all")

# Traffic sign retrieval labels
qrels_signs = load_dataset("unstpb-nlp/RoD-TAL", "qrels_traffic_signs", split="qrels_queries_split_3")

Related resources

• GitHub repository (code + notebooks): https://github.com/vladman-25/RoD-TAL
• Fine-tuned embedding model: https://huggingface.co/unstpb-nlp/multilingual-e5-small-RoD-TAL

The experiment notebooks in the GitHub repository cover corpus scraping, IR, QA, VIR, VQA, and post-hoc analyses.

Citation

If you use RoD-TAL in your research, please cite our paper:

@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}
}