I am kinda technical, but writing and tuning hyper parameters is hard to do well.

When Hugging Face launched Auto Train it was amazing, the results I could get just by dropping in a dataset were awesome and insanely accurate. For (seemingly) no reason they deprecated this feature and only allow now support Advanced Auto Train WHICH IS A NIGHTMARE.

I can’t get it to run. Bugs abound and when I reach out to support they just say “make a github issue, I have a small keyboard and can’t respond here” – whatever that means… but when I look at the github issue page it has 130 opened github issues (most comments are not answers just addition questions from community), and clearly this will do nothing besides waste my time. The docs are so minimal its a joke. And looking at their discord clearly a few other people have this issue but they are promptly ignored. So I tried reaching out to my local community and have had multiple people with +10 years in ML experience try to help and their answer was “it’s going to be easier to just do this manually”.

Assuming HF isn’t going to turn back on support for regular Auto Train, what do I do? Whats’s the best path?

For context, I was using HF to start because I working with domain specific text data and HF is AMAZING when it comes to finding domain specific LLMs to fine tune (overall excluding the issue above I really love HF) and automatically standing up an endpoint was very helpful. Specifically on the data front, I have many types of text datasets all single sentence text data but different labeling strategies depending on the text in question (single category, multi-category, NER, binary…). I have roughly 60-70 datasets in all. They all update with some regular cadence (like every other month or every other quarter depending on the data). This is why AutoTrain was so valuable I just could drop in, refresh and be on my way. Manually managing these seems insanely difficult since the size of the dataset and balance of the data in the dataset can change regularly which means re-tuning the hyper parameters of the models… which given the amount of datasets I need up does not seem tenable. I’m a ‘one man’ team here and I can get one more person for ‘non technical activity’ (i.e. dragging and dropping data in a GUI). Very open to thoughts on how to handle this problem.

TLDR; Hugging Face’s AutoTrain is deprecated, their ‘new’ product is an unusable dumpster-fire, their support/docs are worse… I have to many models to manually fine tune myself (love to hear if you think this is not the case) what should I look to use here?