When do you get charged with AutoTrain?

Hi, I wanted to test AutoTrain and saw that apparently there is one free training included. I think I used it and then tried a second one with a different dataset. I had assumed that a free account can’t be charged and that if I had already used the free training I would be stopped at one point and asked for a credit card but now after reading this forum, I am not so sure anymore.

Can someone explain how that whole thing work? Is it safe to train with a free account or will I get hit with invoices in a month that I will have to extract from my spam folder?

Will it appear in the “Billing” tab at one point?

Hi and thanks for asking! Currently, with a free account you get one free AutoTrain project ; when creating subsequent projects, AutoTrain will present you with an estimated project training cost, before you start the job. We then issue and send invoices on a project basis to the account email address on file - typically within a week of the project training ending. Please note that we are quickly iterating on the product so what I’m describing is how things are working today, and may be out of date in the weeks/months to come.

To be frank, this is a way of doing things that I find super sketchy. The pricing is not clear, the free plan is not obvious, and there are no warning that you will be charged of what is presented like a cost estimate.

If you send me an invoice, I’ll pay it but I am not even sure that this way of doing things is legal here in France. You really should investigate and potentially revisit this approach.

Thanks for the feedback. We try to make the pricing clear by providing real cost estimates based on project, task, uploaded data, number of models etc, which isn’t easy to provide guidelines for until we have all the main parameters. We’ll keep iterating on the messaging to set expectations more clearly to people.

1 Like

And now I am receiving two invoices instead of one despite one training supposedly being free. I am going to refrain from using Huggingface free services or free trial on anything. It is far too random and costly.

That’s a shame, it was indeed working well.

@keepthepace Please note that you have created 3 projects, not 1! Only the first project is free whether or not you train a model in that project. Any subsequent projects are chargeable and the charges are not “random”. You are presented with a screen with number of models you want to train and estimated cost, training only happens after you have seen the cost and you decide to proceed. I’m sure you remember seeing a screen like the following:

The invoice is always between the min and max presented to the end-user and is decided based on several factors. So, its anything but “random”.

If you need any further help or clarifications regarding the invoices, the models you trained or your projects, please feel free to email us at: autotrain [at] huggingface [dot] co

I was figuring out how this thing works with my free account, thinking “obviously this is free, labelled as free, I can’t possibly get charged without having been asked for a credit card”. It is training models that costs energy and compute, I assumed that it was the first training that was offered. You can look at the projects I created, they are all basically empty with a very basic dataset (2000 reddit comments)

Yes, I saw that screen. It does not say it is an invoice. It does not say you’ll get charged. When I saw it first I thought “Ok, this is my free training, they show me how much they are offering to me on the first run. Fair enough, if I need it again I’ll have to pay 30 USD”.

When I tried a second time, I assumed the website would ask me for a credit card number if I actually try. When it did not I got confused and asked here.

This is the total opposite of what any web shop typically does.

Here is my suggestion: you seem to be annoyed by the fact that the exact price is not known before training. That’s fine:

  • Either charge the max and offer some discount on the next one if it comes cheaper
  • Charge to get a download link to the model after training

There are tons of possibilities. But do not lure people in with a “free trial” to hit them with an invoice afterwards. You offer a great service to the NLP community, you deserve better than these shady commercial practices.

2 Likes

Thank you very much for your suggestions. As Jeff mentioned before we are reiterating some parts of the payment step and you will see improvements to it very soon.

As far as the point regarding labelling it as a free trial is concerned, we have never mentioned it as a free software, except for the first project. The FAQs on landing page clearly mention how the users are charged:

Landing page: AutoTrain – Hugging Face

Once again, thank you for your suggestions, we will try our best to improve the workflow.

No one ever talked about “free software”. There is a misunderstanding between “free project” and “free training”.

Screenshot 2022-07-02 at 09-56-02 Hugging Face – Pricing

Considering that it is always free to create a project, it is obvious to understand it as “one free training”.

Putting a vague statement in the FAQ is as legally binding as if I said in my profile that each of my answers in your forum that is welcomed as a good suggestion will be invoiced 100USD each. That’s not how it works. (Or tell me where to send my invoice)

Run the process through your legal department, they probably wont consider it clean.

And while you are at it, please tell them that I wish to apply for my withdrawal right as per the french law on remote sales (yes it applies for sales from oversea):

or to point out where in the process I explicitly agreed to waive this right.

That’s simply not how one sells stuff online.

If you have any concerns related to invoices, please email us at autotrain [at] huggingface.co and we will be happy to help you there. Forums are places for generic question and answers and not for solving individual cases.

Can autotraining be useful for determining the best model given a small dataset, and we train it on our own afterwards?

I’m curious if the datacard, etc. it generates would contain example Python code to train it further based on the original requirements.