TEMA-Data

Data for TEMA: Evidence-Grounded Temporal Question Answering in Multi-Turn Multi-Audio Dialogs.

Paper · Code · Models

Datasets

Component	Config	Split	Size
Temporal initialization	temporal_init	train / validation	98,401 / 500 examples
TEMA-Dialog	sft	train	40,704 dialogues / 198,195 turns
RL training	rl_schedule	train	Download
TEMA-Bench	benchmark	test	253 dialogues / 1,239 questions

Tasks

The 18-subtask guide maps the five task families to all 18 subtasks, with definitions, example questions and answers, evidence requirements, and internal task labels.

Family	Capability	Illustrative question
F1: Localization and measurement	Recover event intervals, durations, and gaps	“When does the bell ring, and how long does it last?”
F2: Identification and verification	Identify events and check their presence or absence	“Is there a dog barking in this recording?”
F3: Within-audio temporal structure	Compare events within one recording	“Does the bell start before the engine?”
F4: Cross-audio retrieval and comparison	Find and compare evidence across recordings	“In which recording does the alarm last longer?”
F5: Multi-turn reference	Resolve references to events discussed earlier	“How long did that sound last?”

SFT responses follow Route → Span → Reason → Answer. Route identifies the recordings to inspect; Span lists all matching event intervals. Audios{k}[NONE] marks a checked recording with no matching event. Times are in seconds on each recording's own timeline.

Load

from datasets import load_dataset

sft = load_dataset("bluesky7/TEMA-Data", "sft", split="train")
benchmark = load_dataset("bluesky7/TEMA-Data", "benchmark", split="test")

Prepare audio

Audio waveforms are obtained separately from AudioSet Strong, TACOS, and AudioTime. Source links are in DATA_SOURCES_AND_LICENSES.md.

hf download bluesky7/TEMA-Data --repo-type dataset --local-dir data_release

python data_release/tools/materialize.py --release-root data_release \
  --source-root audioset_strong=/path/to/audioset_root \
  --source-root audiotime=/path/to/audiotime_root \
  --source-root tacos=/path/to/tacos_root \
  --output-root data_local --include temporal_init sft rl benchmark --mode symlink

Replace the source roots with your audio directories, matching source_relative_path in audio/catalog.jsonl. Prepared annotations are written under data_local.

Files

Directory	Contents
temporal_init/	train.jsonl and val500.jsonl for temporal initialization
sft/	Full-dialogue SFT data
rl/	Evidence GRPO data
benchmark/	Test dialogues and scoring references
audio/	Audio-source catalog
tools/	Audio preparation script

See training and evaluation for commands.

Citation

@misc{yang2026tema,
  title         = {TEMA: Evidence-Grounded Temporal Question Answering in Multi-Turn Multi-Audio Dialogs},
  author        = {Kaidi Yang and Hualei Wang and Zhaohui Wang and Chenxuan Wang and Hong Liu and Xiangdong Wang},
  year          = {2026},
  eprint        = {2609.30029},
  archivePrefix = {arXiv},
  primaryClass  = {cs.SD},
  url           = {https://arxiv.org/abs/2609.30029}
}