Question about the completeness of the SensorQA dataset
Hello! I noticed in your paper that the number of samples in the SensorQA dataset is listed as follows:
Dataset Stage 1 Stage 2
Train Test Train Test
USC-HAD 300,744 58,704 22,790 4,555
UCI-HAR 128,292 25,932 7,352 2,947
PAMAP2 738,666 271,674 14,163 5,210
MHealth 283,020 60,780 4,771 2,039
CAPTURE-24 72,714 35,688 61,327 30,138
However, when I processed the data from the Hugging Face (hf) library using the code you provided, I only obtained 952,128 QA pairs. I'm wondering if I made a mistake in my calculations or if the data in the library is not complete. If the dataset is incomplete, do you have any plans to provide the full dataset?
Hi,
The SensorQA dataset currently available on Hugging Face is not the latest version. We plan to release the full, updated version of the dataset after the paper is accepted.
In the meantime, you can use the provided code to generate your own SensorQA dataset. The code is designed to be flexible β you can easily modify it to incorporate different types of QA templates or even experiment with new data variations.
Please note that since the sensor-text alignment stage data is variable-length, there might be slight differences in the number of generated QA pairs each time you run the code, depending on the data segmentation and QA generation process.