GroundTSBench is a synthetic time-series dataset with fine-grained
natural-language grounding. Every sample is a univariate signal built from a
repeating "event" template plus inter-event filler, and is paired with one or
more natural-language descriptions whose text spans are explicitly linked to
time spans in the signal.
The dataset was generated by a fully programmatic pipeline
(data_gen/)
that samples an event template out of primitive shapes (spike, dip… See the full description on the dataset page:
https://huggingface.co/datasets/houtj19900/GroundTSBench.
