Hacker News Reader

Comment by ebalit | Hacker News Reader

Mistral 7B [1] and many models stemming from it are released under permissive Apache license.

Some might argue that a "pure" open-source would require the dataset and the training "recipe" as it would be needed to reproduce the training, but it would be so expensive that most people wouldn't be able to do much with it.

IMO, a release with open weights without the "source" is much better than the opposite, a release with open source and no trained weights.

And it's not like there was no progress on the open dataset front: - Together just released RedPajama V2, with enough tokens to train a very sizeable base model. - Tsinghua released UltraFeedback which allowed more people to align models using RLHF methods (like the Zephyr models from Hugging Face) - and many many others

[1] https://mistral.ai/news/announcing-mistral-7b/ [2] https://github.com/togethercomputer/RedPajama-Data

View on HN

> IMO, a release with open weights without the "source" is much better than the opposite, a release with open source and no trained weights.

Why not both?

Isn’t everyone using data sets they don’t have rights to? Or at least not right to release.

Both would be even better but it would be much smaller progress for the open source community in comparison to the release of the weights.

> Some might argue that a "pure" open-source would require the dataset and the training "recipe" as it would be needed to reproduce the training, but it would be so expensive that most people wouldn't be able to do much with it.

I've actually argued the opposite. https://www.marble.onl/posts/considerations_for_copyrighting... When you look at the freedoms underlying open source, you can exercise them without the training data.

Most of what disqualifies the licenses I mentioned above from being classically open source is use restrictions.

Yes, I think most of the misunderstanding comes from people who are not in ML. A lot of the times you can't release the training pipeline without getting into grey areas of training data copyright. A ML model is a high dimensional distribution that can be sampled from infinitely, given enough time and money you can train another model (possibly even with a different architecture) that closely models the original model in generation capability.

Even if they don't release the (copyrighted) training data itself, do they provide a "recipe" to reproduce it?

Like, a list of sources used and how they were harvested?

It really depends on the release. It's not the case for the base Mistral 7B model. But many finetuned models are released with some information about the dataset used.

How about "Open Binary"?

Reply on news.ycombinator.com