I'm reserving that complaint for "open source" models which are released under non-open-source licenses.
I care that I know what I can DO with the project when I see it described as "open source".
I'm reserving that complaint for "open source" models which are released under non-open-source licenses.
I care that I know what I can DO with the project when I see it described as "open source".
> I care that I know what I can DO with the project when I see it described as "open source".
Yes, the first of which is that you should be able to build it from source. Which requires the source code, and in this case data.
The OSI's take on this is that an open source model can be modified through fine-tuning etc, even if you can't rebuild it from scratch.
The problem with requiring "build from scratch" for open source models is that the number of interesting models with training data that can be openly licensed is close to zero.
If you trained your model on an unlicensed scrape of the web you can't release the data under an open source license!
The Open Source Initiative have a bunch of their thinking around this in their FAQ for the "Open Source AI definition": https://opensource.org/ai/faq#isn-t-training-data-required-t...
That's a point.
It is legal to train on copyrighted materials, provided they were obtained legally. Most companies also train their models using user interactions with previous iterations.
It is impossible to release this data publicly, let alone license it to a third party. However, I believe that at least the training code and the data processing pipeline could, and should, be released in order to claim a model is truly "open source."
That said, Allen AI actually released several models with the full datasets available. It is impressive how they pushed the models' performance despite training on a limited set of publicly available data. Kudos to them.
I think the OSI no longer has any authority since that stunt they pulled in their "elections".
> The OSI's take on this is that an open source model can be modified through fine-tuning etc, even if you can't rebuild it from scratch.
By this definition almost any binary can be "open source" since hex editors exist. (Or more usefully, you can use ghidra et al. to do more interesting changes.) I know GPL has a very specific view of things, but I'd like to quote an excerpt that I think is generally applicable from https://www.gnu.org/licenses/gpl-3.0.html -
> The “source code” for a work means the preferred form of the work for making modifications to it. “Object code” means any non-source form of a work.
Which is why I'm fine with "open weights", because that's saying the object code is under an open license.
> The problem with requiring "build from scratch" for open source models is that the number of interesting models with training data that can be openly licensed is close to zero.
So? If the number of open source models is zero, then the number of open source models is zero.
I would personally disagree slightly with this take. Freely being able to use means IMHO, that this can be done for all applications in a legal (and ideally ethical) fashion. Regulation often requires to prove the quality or provenance of data. Open source has IMHO often a very libertarian view on things focusing on the rights of the user an not society in general.
They’ll never reveal the data, because that would reveal this is all built on stolen work.
Some of the models DO reveal the data, and it's still built on "stolen work" in that it's unlicensed scrapes of the Web. Here's an example:
https://huggingface.co/allenai/OLMo-2-0325-32B
Here's one of their training mixes: https://huggingface.co/datasets/allenai/dolma3_pool - which includes 8 trillion tokens from Common Crawl.
That would be “permissive license”
Maybe we should have a little cue card for models: vendor/name, size, open weights, open source, permissive license.
It’s simple enough an idea.