DualTime-data

The in-house training corpora of Twice Upon a Time: Complementary Representations for Time-Series Reasoning, named as in the paper's training-data table. Raw JSON / JSONL with the series as number arrays; the charts and delay-embedding images are rendered at training time by the code repository.

Stage	File	Rows	Supervision
A	stageA/chatts_style_captions.jsonl	32,000	caption
A	stageA/structural_captions.json	7,701	caption
B	stageB/tsexam_perception.json	20,640	multiple choice
B	stageB/numerical_corpus.json	11,000	scalar answer
B	stageB/reasoning_corpus.json	100,000	multiple choice, with rationales

The other training sources come from their own repositories: CaTS-Bench (mhfisher/CaTSBench, Stage A) and the ChatTS ift / sft sets (ChatTSRepo/ChatTS-Training-Dataset, Stage B).

Usage

hf download SequentialLearning/DualTime-data --repo-type dataset --local-dir data

data/ is where the code repository's training scripts read the corpora (STAGE=A bash scripts/train.sh, STAGE=B bash scripts/train.sh).

Licences

structural_captions, tsexam_perception and numerical_corpus are generated with a fork of the TimeSeriesExam generator (MIT); chatts_style_captions is built on series from the ChatTS time-series generator (see that dataset's card); reasoning_corpus is this project's own synthetic data.

Citation

@misc{zisling2026twicetime,
      title={Twice Upon a Time: Complementary Representations for Time-Series Reasoning}, 
      author={Hedi Zisling and Liran Nochumsohn and Ilan Naiman and Mihir Parmar and Omri Azencot},
      year={2026},
      eprint={2610.22327},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2610.22327}, 
}