5 ms·
> LLM providers don’t want random low-quality text [...] making it into their training data Don't they though? Most the training data is somewhat contentious,
by charles_f 2y ago
> LLM providers don’t want random low-quality text [...] making it into their training data
Don't they though? Most the training data is somewhat contentious, between the accusation of purely stealing it and the issues around copyright. While it's not text organized directly as knowledge the way Wikipedia is, it's still got some value, and at least that data is much more arguably "theirs" (even if it's not, it's their users').
Short of a pure "we won't train our stuff on your data" in their conditions, I would assume they do.
- CuriouslyC 2y agoThey really don't. Pretraining on low quality QA is a bad idea so at the very least they're filtering down the data set. Low quality QA is still useful for RLHF/task training though if the question is well answered.
- sanxiyn 2y agoPeople often correct bad generation and I think such feedback is valuable. eg "Training Language Models with Language Feedback at Scale" https://arxiv.org/abs/2303.16755 https://arxiv.org/abs/2303.16755
- simonw 2y agoI don't think they do, for a few reasons: 1. Evidence keeps adding up that the quality of your training data really matters. That's one of the reason's OpenAI are signing big dollar deals with media companies The Atlantic, Vox etc have really high quality tokens! 2. Can you imagine the firestorm that will erupt the first time a company credibly claims that their private IP leaked to another user via a ChatGPT session that was later used as part of another response? But as I said in the article... I can't say this with confidence because OpenAI are so opaque about what they actually use that data for! Which is really annoying.
- FeepingCreature 2y agoReally, what you'd want is to train a LoRA or something on user input. Then you can leave it specialized to that user.
- swiftcoder 2y ago> That's one of the reason's OpenAI are signing big dollar deals with media companies The Atlantic, Vox etc have really high quality tokens! They have really high quality tokens in their archive, at any rate. Since a bunch of media outlets have adopted GPT-powered writing tools, future tokens are presumably going to be far less valuable.
- lukan 2y ago"Since a bunch of media outlets have adopted GPT-powered writing tools, future tokens are presumably going to be far less valuable" As long as a human verified the output, I think it is fine. Training on unverified data is bad.
- swiftcoder 2y agoYou are still looping your own models output into the training set regardless. Human verification may avoid outright errors creeping in, but it won't stop the model biasing it's own training set