Training on multiple univariate time series?

[petteriTeikari]

Hi

I was wondering if there is some documentation/tutorial on how to use my own custom univariate time series for fit() of SigLLM

I have this (355,1981) dataset with shared relative timestamps but measured at different times on different people. Could I use to train an outlier detector from it as I had the labels (355,1981 as each timeseries obviously outliers happen at different times for different series, but look similar to each other, and happen mostly at random) for it, that I could then use to predict anomalies in production?

edit: Or as I assume that the method was totally unsupervised, some sort of finetuning or in-context learning with the whole dataset could be nice even with the supervised label use

I know got the detect working, but I assume that I could get a lot better performance with some training instead of zeroshotting through my dataset

The fit() of ofmlblocks/mlpipelinehad at least this option so was wondering can you leverage labels or do I just push the timeseries one-by-one through theSigLLM`? They are a bit hard to concatenate for training

[sarahmish]

Hi @petteriTeikari, since this question is specific to SigLLM I will transfer it to the respective library!

With regards to your question, the pipelines are purely zero-shot and we currently don't support fine-tuning an LLM for your own data. We have a tutorial for using the detector pipeline, you can replace it with your own dataset

If you'd like to train a deep learning model from scratch, such as an LSTM, for unsupervised anomaly detection, you can use Orion!