Can we take a look at the fine tuning dataset?
this post was submitted on 10 Nov 2023
1 points (100.0% liked)
Machine Learning
1 readers
1 users here now
Community Rules:
- Be nice. No offensive behavior, insults or attacks: we encourage a diverse community in which members feel safe and have a voice.
- Make your post clear and comprehensive: posts that lack insight or effort will be removed. (ex: questions which are easily googled)
- Beginner or career related questions go elsewhere. This community is focused in discussion of research and new projects that advance the state-of-the-art.
- Limit self-promotion. Comments and posts should be first and foremost about topics of interest to ML observers and practitioners. Limited self-promotion is tolerated, but the sub is not here as merely a source for free advertisement. Such posts will be removed at the discretion of the mods.
founded 1 year ago
MODERATORS
Did the open source model license agreement allow to re-use the model without indication of what model was used to finetune?
I updated my post. I am using Llama-2-13B. I was not really sure whether I am required to do so with llama-2 license but just in case.
Thanks, but my healthcare plan only covers ELIZA
Thanks for the nostalgia!
Paging Dr. Sbaitso
I hope you fully understand the consequences that may can occur by releasing such a model. A relatively small model with (what I assume) a relatively small non real world dataset, used by a mentally unstable person can lead to bad things IMO. That's why companies are very careful with this type of application. Huge liability and ethical issue
I agree with the points you are raising. This is meant to be like a companion (think of Replika) that you can talk, to sort out your feelings and learn mindfulness techniques. It is not meant for somebody who is struggling with serious mental health issues. I already have a description that mentions several times that it is generative AI and should be used with caution. I think I should clarify that further in bot description and in its naming, not to mislead people.
I have thought about all the dangers that it poses for a long time but I think if built and used responsibly, it has the potential to help huge number of people. Especially right now, because there is a severe shortage of mental health professionals and not many people can afford it. I also didn't have great personal experience with human therapists, most of them put me on a lot of medication that ended up not helping at all.
As for the model size, this is only for getting some feedback. My current inference server cannot handle big number of users. Even if I intend to have a bigger release, I will not be doing it without much more capable model, extensive testing, more safety guardrails and RLHF.
if you're expecting people in need of therapy to use this with legitimate reason "responsibly" and "with caution", it essentially defeats the purpose of the bot which was intentionally created to be a reliable, self-policed companion targeting a user-base that vary on a wide spectrum of interpretation to what even constitutes a normalized sense of responsibly and caution..
way too many factors at play to mitigate in terms of liability, as others have said. take this resource as an example for a bot created to help those with eating disorders: https://www.npr.org/sections/health-shots/2023/06/08/1180838096/an-eating-disorders-chatbot-offered-dieting-advice-raising-fears-about-ai-in-hea
How is it all that different from writing a self-help book though, in a legal sense? Why is there any bigger liability and ethical issue? There are millions of such books and articles published, full of all sorts of nonsense. A mentally unstable person might follow one and experience a bad outcome. No reason to stop releasing books though. A simple disclaimer seems to suffice. I understand that the experience of an LLM is not the same thing as reading a book, but it is in a sense just indexing and summarizing many texts in an interactive way. Why do you think it would be treated differently under the law, are there specific laws that apply to LLMs but not to books?
An easy difference is the data: you hardly give any personal data when buying a book, but you would have to expose your deepest secrets to this system. If the system is trained on interactions with patients, it may even leak this data through responses in the future.
Feedback: as others have noted, unless you live in a country where therapy isn't regulated, this is a ticking legal time bomb. If you think no one at a company like Better Help has thought about building this, you're wrong. The issues are:
- Where's legal liability when someone self-harms?
- Where does your training set come from? Existing therapy log datasets that don't mention the training of LLMs as a possible use case in the legal agreement that clients signed may not be usable in the way you've used them.
- What's your core hypothesis? That generating therapist-language has therapeutic value? Remember that the whole profession is about indirectly developing a model of what's going on in someone else's head, and then leading them to a conclusion based on that the therapist thinks may help them (this is reductionist and overly simplistic, but I think the point holds). This level of modeling and indirection seems poorly suited to a language-generator.
print("I understand. And how did that make you feel?")But think of the upside.
With a lack of well enforced privacy policies, the data such systems can mine for
- selling to real insurance companies to choose who's coverage to deny based on facebook
- selling to TSA watch list vendors
- extortion
- etc
This project could be the next FTX + Theranos.
I would just LOVE to see the technical details.
Where is the LLM hosted?
You look lonely, i can fix that.
Legally and ethically questionable to automate healthcare this way.
I don't agree. As a tool used by the individual, there are few ethics complaints.
I already built therapist bots, fun times.
Tried it out. It’s pretty good, but the main flaw I see is that it gets stuck in a loop of asking the same or similar questions rather than going deeper in the conversation.
Where and how did you get the training data?
Sweet...can you guys compare it to this quick and easy one I made with custom GPT. Tried to oversell it on twitter, though the instructions I used were very detailed and were derived from clinical experience: