Hi MSavoritias, all,
Let me provide more context.
The concern started couple of months ago, to my knowledge. And
discussion is still on going. So I think that’s incorrect to say “any
result for over 6 months”.
Moreover, I feel you have a misunderstanding about HuggingFace and SWH
partnership. From the reading of public information, HuggingFace and
BigCode trains on a subset of SWH source code archive. I mean, it is a
snapshot and to my knowledge, they provided the list of source code that
had been used for training.
Not to avoid the question but from a pragmatic point of view, one might
ask if the source code you write and do not want to be included in the
training dataset, if this source code is concretely part of that
training dataset.
HuggingFace is not training continuously with source code from SWH.
And technically, SWH is an archive i.e., the code is not stored hot. I
do not know and I have not read all details by HuggingFace of their
method; i.e., which kind of data they process – independent unique
files, complete repository, etc. What I know is that the piece when
fetching from SWH is named SWH Vault; it requires to “cook” and prepare
all the files that take times, from minutes to days.
All that to say two key points:
1. People behind SWH are well-aware about various sides of the concerns.
As said, they are long-time free software supporters. Be sure they have
eared community concerns. Some discussions are still pending because as
explained, all sides of ethical questions needs to be cautious.
Please do not think it is ignored.
2. FWIW, I am in touch with SWH people – among other members from Guix
community. For instance, in order to feed the discussion, Roberto from
SWH pointed to me this blog point by Bruce Perens:
https://perens.com/2019/10/12/invasion-of-the-ethical-licenses/
Well, I do not know if the outcome will be aligned with your current
opinion, but be sure that your concerns as the others raised by Guix
community members are taking into account.
Cheers,
simon
