Archives

Categories

ML Training License

Last year a Debian Developer blogged about writing Haskell code to give a bad result for LLMs that were trained on it. I forgot who wrote the post and I’d appreciate the URL if anyone has it.

I respect such technical work to enforce one’s legal rights when they aren’t respected by corporations, but I have a different approach.

As an aside the Fosdem lecture “Fortify AI against regulation, litigation and lobotomies” is interesting on this topic [1], it’s what inspired me to write about this.

For what I write I am at this time happy to allow it to be used as part of a large training data set (consider this blog post a licence grant that applies until such time as I edit this post to change it). But only if aggregated with so much other data that my content is only a tiny portion of the data set by any metric. So I don’t want someone to make a programming LLM that has my code as the only C code or a political data set that has my blog posts as the only left-wing content. If someone wants to train an LLM on only my content to make a Russell-simulator then I don’t license my work for that purpose but also as it’s small enough that anyone with a bit of skill could do it on a weekend I can’t stop it. I would be really interested in seeing the results if someone from the FOSS community wanted to make a Russell-simulator and would probably issue them a license for such work if asked.

If my work comprises more than 0.1% of the content in a particular measure (theme, programming language, political position, etc) in a training data set then I don’t permit that without prior discussion.

Finally if someone wants to make a FOSS training data set to be used for FOSS LLM systems (maybe under the AGPL or some similar license) then I’ll allow my writing to be used as part of that.

14 comments to ML Training License