Please help us define materials to be delivered

[heathermeeker]

Thanks to everyone for the PRs so far!

I would particularly like input on the item in the definition that described what needs to be delivered, in order for a model to be open. In other words, what is the analog of source code, for open source software?

Also, is there a format in which it should be delivered?

[AzulGarza]

Thank you for opening this interesting discussion!

I think when we talk about making a machine learning model open-source, the analogy to the source code in software is the full pipeline of generating the model, along with the model weights. Here's what I believe should be included:

1. Training Code and Environment:
The source code used for training the models should be shared, detailing:

• The programming language used, along with its specific version.
• All open-source libraries utilized, specifying their respective versions to ensure reproducibility.
• The operating system and hardware on which the model was trained, providing a rough estimate of the expected training time.
• A Docker image can be especially beneficial to encapsulate the development environment and eliminate dependency conflicts.

2. Model Weights and Usage Instructions:
For the neural network weights (NNW), the following should be provided:

• A publicly accessible URL where the model weights are hosted.
• A comprehensive guide on how to use the pre-trained models with the aforementioned requirements. This should include sample code illustrating how to load the model weights and make predictions.

3. Repository for Information:
All this information should ideally be stored and organized in a GitHub/GitLab repository, with proper documentation to assist contributors.

[mergenthaler]

I think these people could contribute to this discussion: @julien-c, @williamFalcon, @jph00, @rasbt, @clmnt, @dongreenberg.

[rasbt]

The source code used for training the models should be shared, detailing:

In particular, I'd say the definition is that source code should be made freely available and may be redistributed and modified by anyone. I.e., not only be shared, but shared freely. And modifiable to one's liking.

making a machine learning model open-source

if we are talking about making "the model" open source, then the weights should also be made freely available etc.

EDIT:

Is it always feasible? Probably not. But in this case it then shouldn't be called open source. I honestly don't blame people if they can't make everything open-source. It's more false advertising if some people/companies put an open-source label on something that isn't. It's like saying "free unlimited internet" but then having a "500 Gb / month" data volume cap.