4 ms·

the training data is the source.

by exe34 2y ago

JumpCrisscross 2y ago
> the training data is the source Sure. But that's not going to be released. The term open source AI cannot be expected to cover it because it's not practical.
- diggan 2y ago
  So because it's really hard to do proper Open Source with these LLMs, means we need to change the meaning of Open Source so it fits with these PR releases?
  - JumpCrisscross 2y ago
    > because it's really hard to do proper Open Source with these LLMs, means we need to change the meaning of Open Source so it fits with these PR releases? Open training data is hard to the point of impracticality. It requires excluding private and proprietary data. Meanwhile, the term "open source" is massively popular. So it will get used. The question is how. Meta et al would love for the choice to be between, on one hand, open weights only, and, on the other hand, open training data, because the latter is impractical. That dichotomy guarantees that when someone says open source AI they'll mean open weights. (The way open source software, today, generally means source available, not FOSS.)
    - Palomides 2y ago
      source available is absolutely not the same as open source you are playing very loosely with terms that have specific, widely accepted definitions (e.g. https://opensource.org/osd https://opensource.org/osd ) I don't get why you think it would be useful to call LLMs with published weights "open source"
      - JumpCrisscross 2y ago
        > terms that have specific, widely accepted definitions OSF's definition is far from the only one [1]. Switzerland is currently implementing CH Open's definition, the EU another one, et cetera. > I don't get why you think it would be useful to call LLMs with published weights "open source" I don't. I'm saying that if the choice is between open weights or open weights + open training data, open weights will win because the useful definition will outcompete the pristine one in a public context. [1] https://en.wikipedia.org/wiki/Open-source_software#Definitions https://en.wikipedia.org/wiki/Open-source_software#Definitio...
        diggan 2y ago
        For the EU, I'm guessing you're talking about the EUPL, which is FSF/OSI approved and GPL compatible, generally considered copyleft. For the CH Open, I'm not finding anything specific, even from Swiss websites, could you help me understand what you're referring to here? I'm guessing that all these definitions have at least some points in common, which involves (another guess) at least being able to produce the output artifacts/binaries by yourself, something that you cannot do with Llama, just as an example.
        JumpCrisscross 2y ago
        > For the CH Open, I'm not finding anything specific, even from Swiss websites, could you help me understand what you're referring to here Was on the HN front page earlier [1][2]. The definition comes strikingly close to source on request with no use restrictions. > all these definitions have at least some points in common Agreed. But they're all different. There isn't an accepted defintiion of open source even when it comes to software; there is an accepted set of broad principles. [1] https://news.ycombinator.com/item?id=41047172 https://news.ycombinator.com/item?id=41047172 [2] https://joinup.ec.europa.eu/collection/open-source-observatory-osor/news/new-open-source-law-switzerland https://joinup.ec.europa.eu/collection/open-source-observato...
        diggan 2y ago
    - diggan 2y ago
      > Open training data is hard to the point of impracticality. It requires excluding private and proprietary data. Right, so the onus is on Facebook/Meta to get that right, then they could call something Open Source, until then, find another name that already doesn't have a specific meaning. > (The way open source software, today, generally means source available, not FOSS.) No, but it's going in that way. Open Source, today, still means that the things you need to build a project, is publicly available for you to download and run on your own machine, granted you have the means to do so. What you're thinking of is literally called "Source Available" which is very different from "Open Source". The intent of Open Source is for people to be able to reproduce the work themselves, with modifications if they want to. Is that something you can do today with the various Llama models? No, because one core part of the projects "source code" (what you need to reproduce it from scratch), the training data, is being held back and kept private.
    - unethical_ban 2y ago
      >Meanwhile, the term "open source" is massively popular. So it will get used. The question is how. Here's the source of the disagreement. You're justifying the use of the term "open source" by saying it's logical for Meta to want to use it for its popularity and layman (incorrect) understanding. Other person is saying it doesn't matter how convenient it is or how much Meta wants to use it, that the term "open source" is misleading for a product where the "source" is the training data, and the final product has onerous restrictions on use. This would be like Adobe giving Photoshop away for free, but for personal use only and not for making ads for Adobe's competitors. Sure, Adobe likes it and most users may be fine with it, but it isn't open source. >The way open source software, today, generally means source available, not FOSS. I don't agree with that. When a company says "open source" but it's not free, the tech community is quick to call it "source available" or "open core".
      - JumpCrisscross 2y ago
        > You're justifying the use of the term "open source" by saying it's logical for Meta to want to use it for its popularity and layman (incorrect) understanding I'm actually not a fan of Meta's definition. I'm arguing specifically against an unrealistic definition, because for practical purposes that cedes the term to Meta. > the term "open source" is misleading for a product where the "source" is the training data, and the final product has onerous restrictions on use Agree. I think the focus should be on the use restrictions. > When a company says "open source" but it's not free, the tech community is quick to call it "source available" or "open core" This isn't consistently applied. It's why we have the free vs open vs FOSS fracture.
  - elevatedastalt 2y ago
    No, we need to adapt an existing term into the new context that it is being deployed in.
- plsbenice34 2y ago
  Of course it could be practical - provide the data. The fact of that society is a dystopian nightmare controlled by a few megacorporations that don't want free information does not justify outright changing the meaning of the language.
  - JumpCrisscross 2y ago
    > provide the data Who? It's not their data.
    - exe34 2y ago
      why are they using it?
      - guitarlimeo 2y ago
        And why legislation allows them to use the data to train their LLM and release that, but not release the data?
- tintor 2y ago
  Meta can call it something else other than open source. Synthetic part of the training data could be released.
JimDabell 2y ago
I don’t think it’s that simple. The source is “the preferred form of the work for making modifications to it” (to use the GPL’s wording). For an LLM, that’s not the training data. That’s the model itself. You don’t make changes to an LLM by going back to the training data and making changes to it, then re-running the training. You update the model itself with more training data. You can’t even use the training code and original training data to reproduce the existing model. A lot of it is non-deterministic, so you’ll get different results each time anyway. Another complication is that the object code for normal software is a clear derivative work of the source code. It’s a direct translation from one form to another. This isn’t the case with LLMs and their training data. The models learn from it, but they aren’t simply an alternative form of it. I don’t think you can describe an LLM as a derivative work of its training data. It learns from it, it isn’t a copy of it. This is mostly the reason why distributing training data is infeasible – the model’s creator may not have the license to do so. Would it be extremely useful to have the original training data? Definitely. Is distributing it the same as distributing source code for normal software? I don’t think so. I think new terminology is needed for open AI models. We can’t simply re-use what works for human-editable code because it’s a fundamentally different type of thing with different technical and legal constraints.
- jononor 2y ago
  Given a well behaved training setup, you will give an equivalently powerful model, given the same dataset and training scripts, and training settings. At least if you are willing to run it several times, an pick the best one - a process that is commonly used for large models.
- jononor 2y ago
  No the preferred way to make modifications is using the the training code. One may also input a snapshot weighs to start from, but the training code is definitely what you would modify to make a change.
  - exe34 2y ago
    how do you train it in a different language by changing the training code?
root_axis 2y ago
No. It's an asset used in the training process, the source code can process arbitrary training data.
sangnoir 2y ago
We've had a similar debate before, but the last time it about whether Linux device drivers based on non-public datasheets under NDA were actually open source. This debate occurred again over drivers that interact with binary blobs. I disagree with the purists - if you can legally change the source or weights - even without having access to the data used by the upstream authors - it's open enough for me. YMMV.
wrs 2y ago
I don’t think even that is true. I conjecture that Facebook couldn’t reproduce the model weights if they started over with the same training data, because I doubt such a huge training run is a reproducible deterministic process. I don’t think anyone has “the” source.
- exe34 2y ago
  numpy.random.seed(1234)

the training data is the source.

by exe34 2y ago

the training data is the source.

> the training data is the source Sure. But that's not going to be released. The term open source AI cannot be expected to cover it because it's not practical.

diggan 2y ago
So because it's really hard to do proper Open Source with these LLMs, means we need to change the meaning of Open Source so it fits with these PR releases?
- JumpCrisscross 2y ago
  > because it's really hard to do proper Open Source with these LLMs, means we need to change the meaning of Open Source so it fits with these PR releases? Open training data is hard to the point of impracticality. It requires excluding private and proprietary data. Meanwhile, the term "open source" is massively popular. So it will get used. The question is how. Meta et al would love for the choice to be between, on one hand, open weights only, and, on the other hand, open training data, because the latter is impractical. That dichotomy guarantees that when someone says open source AI they'll mean open weights. (The way open source software, today, generally means source available, not FOSS.)
  - Palomides 2y ago
    source available is absolutely not the same as open source you are playing very loosely with terms that have specific, widely accepted definitions (e.g. https://opensource.org/osd https://opensource.org/osd ) I don't get why you think it would be useful to call LLMs with published weights "open source"
    - JumpCrisscross 2y ago
      > terms that have specific, widely accepted definitions OSF's definition is far from the only one [1]. Switzerland is currently implementing CH Open's definition, the EU another one, et cetera. > I don't get why you think it would be useful to call LLMs with published weights "open source" I don't. I'm saying that if the choice is between open weights or open weights + open training data, open weights will win because the useful definition will outcompete the pristine one in a public context. [1] https://en.wikipedia.org/wiki/Open-source_software#Definitions https://en.wikipedia.org/wiki/Open-source_software#Definitio...
      - diggan 2y ago
        For the EU, I'm guessing you're talking about the EUPL, which is FSF/OSI approved and GPL compatible, generally considered copyleft. For the CH Open, I'm not finding anything specific, even from Swiss websites, could you help me understand what you're referring to here? I'm guessing that all these definitions have at least some points in common, which involves (another guess) at least being able to produce the output artifacts/binaries by yourself, something that you cannot do with Llama, just as an example.
        JumpCrisscross 2y ago
        > For the CH Open, I'm not finding anything specific, even from Swiss websites, could you help me understand what you're referring to here Was on the HN front page earlier [1][2]. The definition comes strikingly close to source on request with no use restrictions. > all these definitions have at least some points in common Agreed. But they're all different. There isn't an accepted defintiion of open source even when it comes to software; there is an accepted set of broad principles. [1] https://news.ycombinator.com/item?id=41047172 https://news.ycombinator.com/item?id=41047172 [2] https://joinup.ec.europa.eu/collection/open-source-observatory-osor/news/new-open-source-law-switzerland https://joinup.ec.europa.eu/collection/open-source-observato...
        diggan 2y ago
  - diggan 2y ago
    > Open training data is hard to the point of impracticality. It requires excluding private and proprietary data. Right, so the onus is on Facebook/Meta to get that right, then they could call something Open Source, until then, find another name that already doesn't have a specific meaning. > (The way open source software, today, generally means source available, not FOSS.) No, but it's going in that way. Open Source, today, still means that the things you need to build a project, is publicly available for you to download and run on your own machine, granted you have the means to do so. What you're thinking of is literally called "Source Available" which is very different from "Open Source". The intent of Open Source is for people to be able to reproduce the work themselves, with modifications if they want to. Is that something you can do today with the various Llama models? No, because one core part of the projects "source code" (what you need to reproduce it from scratch), the training data, is being held back and kept private.
  - unethical_ban 2y ago
    >Meanwhile, the term "open source" is massively popular. So it will get used. The question is how. Here's the source of the disagreement. You're justifying the use of the term "open source" by saying it's logical for Meta to want to use it for its popularity and layman (incorrect) understanding. Other person is saying it doesn't matter how convenient it is or how much Meta wants to use it, that the term "open source" is misleading for a product where the "source" is the training data, and the final product has onerous restrictions on use. This would be like Adobe giving Photoshop away for free, but for personal use only and not for making ads for Adobe's competitors. Sure, Adobe likes it and most users may be fine with it, but it isn't open source. >The way open source software, today, generally means source available, not FOSS. I don't agree with that. When a company says "open source" but it's not free, the tech community is quick to call it "source available" or "open core".
    - JumpCrisscross 2y ago
      > You're justifying the use of the term "open source" by saying it's logical for Meta to want to use it for its popularity and layman (incorrect) understanding I'm actually not a fan of Meta's definition. I'm arguing specifically against an unrealistic definition, because for practical purposes that cedes the term to Meta. > the term "open source" is misleading for a product where the "source" is the training data, and the final product has onerous restrictions on use Agree. I think the focus should be on the use restrictions. > When a company says "open source" but it's not free, the tech community is quick to call it "source available" or "open core" This isn't consistently applied. It's why we have the free vs open vs FOSS fracture.
- elevatedastalt 2y ago
  No, we need to adapt an existing term into the new context that it is being deployed in.
plsbenice34 2y ago
Of course it could be practical - provide the data. The fact of that society is a dystopian nightmare controlled by a few megacorporations that don't want free information does not justify outright changing the meaning of the language.
- JumpCrisscross 2y ago
  > provide the data Who? It's not their data.
  - exe34 2y ago
    why are they using it?
    - guitarlimeo 2y ago
      And why legislation allows them to use the data to train their LLM and release that, but not release the data?
tintor 2y ago
Meta can call it something else other than open source. Synthetic part of the training data could be released.

JimDabell 2y ago

I don’t think it’s that simple. The source is “the preferred form of the work for making modifications to it” (to use the GPL’s wording). For an LLM, that’s not the training data. That’s the model itself. You don’t make changes to an LLM by going back to the training data and making changes to it, then re-running the training. You update the model itself with more training data. You can’t even use the training code and original training data to reproduce the existing model. A lot of it is non-deterministic, so you’ll get different results each time anyway. Another complication is that the object code for normal software is a clear derivative work of the source code. It’s a direct translation from one form to another. This isn’t the case with LLMs and their training data. The models learn from it, but they aren’t simply an alternative form of it. I don’t think you can describe an LLM as a derivative work of its training data. It learns from it, it isn’t a copy of it. This is mostly the reason why distributing training data is infeasible – the model’s creator may not have the license to do so. Would it be extremely useful to have the original training data? Definitely. Is distributing it the same as distributing source code for normal software? I don’t think so. I think new terminology is needed for open AI models. We can’t simply re-use what works for human-editable code because it’s a fundamentally different type of thing with different technical and legal constraints.

jononor 2y ago
Given a well behaved training setup, you will give an equivalently powerful model, given the same dataset and training scripts, and training settings. At least if you are willing to run it several times, an pick the best one - a process that is commonly used for large models.
jononor 2y ago
No the preferred way to make modifications is using the the training code. One may also input a snapshot weighs to start from, but the training code is definitely what you would modify to make a change.
- exe34 2y ago
  how do you train it in a different language by changing the training code?

root_axis 2y ago

No. It's an asset used in the training process, the source code can process arbitrary training data.

sangnoir 2y ago

We've had a similar debate before, but the last time it about whether Linux device drivers based on non-public datasheets under NDA were actually open source. This debate occurred again over drivers that interact with binary blobs. I disagree with the purists - if you can legally change the source or weights - even without having access to the data used by the upstream authors - it's open enough for me. YMMV.

wrs 2y ago

I don’t think even that is true. I conjecture that Facebook couldn’t reproduce the model weights if they started over with the same training data, because I doubt such a huge training run is a reproducible deterministic process. I don’t think anyone has “the” source.

exe34 2y ago
numpy.random.seed(1234)