gamma-lab-umd/temporal-lalm

Temporal LALM: Relative Temporal Audio MCQA

Multiple-choice questions probing relative temporal reasoning over audio: identifying which sound event starts earliest, ends latest, or has the longest duration within a clip. Built on the TACOS audio collection.

Tasks

task	question	#MCQs
earliest_start	Which sound event starts earliest?	528
latest_end	Which sound event ends latest?	499
longest_duration	Which sound event has the longest duration?	630
total		1657

718 unique audio clips; each clip may appear under multiple tasks. The correct-answer letter is balanced near-uniformly across A/B/C/D in every task.

Columns

• audio: decoded mp3 (datasets.Audio)
• filename: source mp3 name
• question: the prompt
• option_a … option_d: the four answer choices
• correct_answer: correct option letter (A/B/C/D)
• correct_text: text of the correct option
• task: one of earliest_start, latest_end, longest_duration

Usage

from datasets import load_dataset
ds = load_dataset("gamma-lab-umd/temporal-lalm", split="train")
print(ds[0])

Citation

If you use this dataset, please cite:

@misc{kulkarni2026closerlookfailuremodes,
      title={A Closer Look at Failure Modes in Temporal Understanding of Large Audio-Language Models},
      author={Apoorva Kulkarni and Kaousheik Jayakumar and Sreyan Ghosh and Sarah Wiegreffe and Dinesh Manocha and Ramani Duraiswami},
      year={2026},
      eprint={2606.17417},
      archivePrefix={arXiv},
      primaryClass={cs.SD},
      url={https://arxiv.org/abs/2606.17417},
}

This dataset is built on TACOS:

@misc{primus2025tacos,
      title={TACOS: Temporally-aligned Audio CaptiOnS for Language-Audio Pretraining},
      author={Paul Primus and Florian Schmid and Gerhard Widmer},
      year={2025},
      eprint={2505.07609},
      archivePrefix={arXiv},
      url={https://arxiv.org/abs/2505.07609},
}