Re:Snowflake Arctic Instruct (128x3B MoE), largest open sour...
by cqqxo4zV46cp on Wednesday April 24, 2024 @12:50PM
(#40146793)
Thankfully, thankfully, this sort of stuff isn’t decided based on the personal reckoning of someone on Hacker News.
Whether or not training data needs to be open source in order for the resulting model to be open source is, at the very least, up for debate. And that’s a charitable interpretation. This is quite clearly instead your view based on your own personal philosophy. Software licenses are legal instruments, not a vague notion of some ideology. If you don’t think that the model is open source, you’ve obviously seen legal precedent that nobody else has.
[ Parent
| on HN ]
Re:Snowflake Arctic Instruct (128x3B MoE), largest open sour...
by Zambyte on Wednesday April 24, 2024 @01:04PM
(#40146984)
The training data is the source. If the training data is not open, the model is not open source, because the source of the model is not open. See this previous comment of mine that explains this: https://news.ycombinator.com/item?id=40035688
Re:Snowflake Arctic Instruct (128x3B MoE), largest open sour...
by stefan_ on Wednesday April 24, 2024 @12:55PM
(#40146873)
What? You know the people writing open source licenses have spent more than 5 minutes thinking about this, right?
The GPL says it straight up:
> The “source code” for a work means the preferred form of the work for making modifications to it
Clearly just weights don't qualify, just like C run through an obfuscator would not count.