Request for training data or construction details for the 10S/* buckets

#4
by Metic - opened

Hello,

Thank you for releasing TotoFnf. It is a great honor to use your model in our further research.

While examining the training setup, we noticed that the public GIFT-Eval pretraining dataset does not appear to contain time series with a 10S frequency. However, TotoFnf provides models or routing buckets for configurations such as 10S/short, 10S/medium, and 10S/long.

Could you please clarify how the training data for these 10S/* buckets were constructed?

In particular, we would greatly appreciate it if you could provide either:

The training data used for the 10S/* buckets;
The data preprocessing or resampling procedure used to generate the 10-second-frequency samples; or
Additional details about how the 10S/* bucket models were trained when no native 10S data were available in GIFT-Eval Pretrain.

This information would be very helpful for reproducing the results and conducting further research based on TotoFnf.

Thank you very much for your time and for sharing this work.

They were trained on gifteval train set, not the pretrain set.

Sign up or log in to comment