This was such a helpful way to frame the problem! Something felt off about the "open source models" out there; this highlights the problem incredibly well.
As much as I appreciate Mis(x)tral, I would've loved it even more if they released code for gathering data.
I think many countries (japan already has) will allow IP for training data.
They just need to buy time until then.
The LLM inference engine (architecture implementation) is like a kernel driver that loads a firmware binary blob, or a virtual machine that loads bytecode. The inference engine is open source. The problem is that the weights (firmware blobs, VM bytecodes) are opaque: you don't have the means to reproduce them.
The Linux community has long argued that drivers that load firmware blobs are cheating: they don't count as open source.
Still, the "open source" LLMs are more open than "API-gated" LLMs. It's a step in the right direction, but I hope we don't stop there.
It would probably take well over a million dollars in engineering hours to recreate the postgres source code from scratch, just as it would take millions in compute to rebuild the weights.
You can get some effects by fine tuning, and in that case it may be preferable as it's cheaper, but in general if I want to have a different or better model, that involves retraining.
... models provided in weights only form. (mostly!)
I believe the preferred form would be the whole kit and caboodle: the collection and filtering scripts, the data to the extent that it's non-public, the training routine, and the model weights... because sometimes you'll perform changes at any of those stages.
I see you’re someone else, so I’ll ask you too. Do you actually have any experience doing this? Have you ever fine tuned models or tried to change architecture or put a piece of one model into another?
Sure, training datasets for pythia is useful. The Pile was used in lots of models. However it's hardly relevant that pythia itself was trained on pile. They live separate lives.
Having just weights already allows making results that are incredibly useful(you don't need original dataset for flash attention, or tuning foundation model into the chat model).
Point is: Having both doesn't make released model more useful.
>Do you actually have any experience doing this? Have you ever fine tuned models or tried to change architecture or put a piece of one model into another?
Yes on both finetune and "changing" architecture: with adapters and similar approaches you don't need to retrain everything from scratch after modifying the guts of the original architecture up to your liking, you just need to not stir it up too much. Training on the task at hand is sufficient.
No, I haven't glued parts of existing models together(ensemble doesn't count)
1. The thread is about the requirements of calling a model open source. The goal of making the models is separate from the requirements of the open source definition.
2. Suppose that author A of a model prefers working exclusively with the model weights to modify the model. Author A's preferred form of modification includes the model weights - and whatever scripts are needed to generate a running model from the weights - but does not include the training set and initial training scripts. Suppose that author B of an unrelated model prefers to retrain the model as part of the process of modifying the model. If author B changes the training set and/or changes the training scripts, then the training set is part of the preferred form of modifying the model. The training set and the training scripts are both necessary for turning the training set into a running model, so I think author B would have to include the training set even if author B changes only the training scripts. Correct me if I'm wrong.) jncfhnb, you're like author A, so if you were to release an open source model then you would need to include the weights but not the training data. Trapais and nullc, don't assume that every model author is author B.
For personal reference, here is the relevant excerpt from the open source definition from the Open Source Initiative [A1]:
> The source code must be the preferred form in which a programmer would modify the program. Deliberately obfuscated source code is not allowed. Intermediate forms such as the output of a preprocessor or translator are not allowed.
Open source software is not the same as free software, but here is the relevant excerpt from the free software definition explainer from the Free Software Foundation [A2]:
> Source code is defined as the preferred form of the program for making changes in. Thus, whatever form a developer changes to develop the program is the source code of that developer's version.
It is for Open Source. Hence why it's silly to call these models open source.
> The “Corresponding Source” for a work in object code form means all the source code needed to generate, install, and (for an executable work) run the object code and to modify the work, including scripts to control those activities. However, it does not include the work's System Libraries, or general-purpose tools or generally available free programs which are used unmodified in performing those activities but which are not part of the work. For example, Corresponding Source includes interface definition files associated with source files for the work, and the source code for shared libraries and dynamically linked subprograms that the work is specifically designed to require, such as by intimate data communication or control flow between those subprograms and other parts of the work.
...
> You may convey a covered work in object code form under the terms of sections 4 and 5, provided that you also convey the machine-readable Corresponding Source under the terms of this License
Model weights alone are not Corresponding Source. In order to distribute a model you made under the GPLv3, you would have to give users the model weights and the scripts needed to turn the model weights into a runnable model. That's assuming that you only work with the model weights when modifying the model. If you in particular retrain the model as part of modifying the model, then you would have to provide the training data and initial training scripts as well.
Even though I wrote about a particular free software license which happens to be an open source license, the open source definition from the Open Source Initiative also refers to the preferred form of changing the work [2]:
> The source code must be the preferred form in which a programmer would modify the program. Deliberately obfuscated source code is not allowed. Intermediate forms such as the output of a preprocessor or translator are not allowed.
For good measure, here is the relevant excerpt from the free software definition from the Free Software Foundation [3]:
> Obfuscated “source code” is not real source code and does not count as source code.
> Source code is defined as the preferred form of the program for making changes in. Thus, whatever form a developer changes to develop the program is the source code of that developer's version.
> Freedom 1 includes the freedom to use your changed version in place of the original. If the program is delivered in a product designed to run someone else's modified versions but refuse to run yours—a practice known as “tivoization” or “lockdown,” or (in its practitioners' perverse terminology) as “secure boot”—freedom 1 becomes an empty pretense rather than a practical reality. These binaries are not free software even if the source code they are compiled from is free.
The FSF's free software definition requires that the user be practically - not merely theoretically - allowed to modify the source code and turn the source code into a running program. Because of that, the free software definition considers build scripts to be part of the source code. I can't find an explicit analogue of the practically-modifiable requirement in the open source definition, but I think providing the model weights without providing the scripts needed to turn the weights into a functioning copy of the existing model would be obfuscation i.e. a violation of the open source definition.
[1] https://www.gnu.org/licenses/gpl-3.0.en.html
Android is not open source.
This analogy is bad. Models are unlike code bases in this way.
What if I wanted to train it using only half of its training set? If the inputs that were used to generate the set of released weights are not available I can’t do that. I have a set of weights and the model structure but without the training dataset I have no way of doing that.
To riff on the parent post, I have:
Source + Compiler => Binaries
[unavailable inputs] + Model Structure => Weights
Model Structure + Weights + [my own training data] => New Weights
You can fine tune into a different model architecture.
You’re right on not being able to retrain the model from scratch on half its data without that data but that’s likely pointless.
> likely pointless
It doesn’t take too much creativity to come up with ideas about why someone might want to do that:
- researchers who want to investigate how much the dataset can be reduced (and thus training cost) and what the accuracy penalty is
- someone who wants to for either religious or ethical reasons minimize the probability that the model was trained on pornography
- someone who’s curious about whether there’s significant redundancy in the existing input datasets
- someone who’s curious about whether there are a much smaller subset of images in the input dataset that can quickly help the first few CNN input layers converge before training the middle and output layers on the larger dataset.
Edit: I suspect the real reason they don’t want to share the input dataset is purely because a high-quality annotated dataset is a valuable commodity. While I don’t do ML work myself day-to-day, I do work with a team that does in a very niche field and I can only imagine how much effort they had to go through to get the annotated dataset that they’ve put together. Even just collecting the images for it involved many hours of drone flights in different locales around North America in varying weather and lighting.
You will need some data of your own of course to fill in the blanks
Edit; however conversely, you can also splice out layers from one model into another original model. It’ll take some retraining, but this works!
Models are the compiler + makefiles. Dataset is the code.
A slightly offtopic complaint, but too often I have seen tutorials for open source stuff (coughopenglcough) where they don't provide the proper commands to compile and link everything required to build it. Figuring it out makes the "getting started" portion even more tedious.
Can the public compete? What percentage of the technical public could we expect to participate, and how much data, compute, and data quality improvement could they bring to the table? I suspect that large corporations are at least an order of magnitude advantaged economically.