Open-Assistant icon indicating copy to clipboard operation
Open-Assistant copied to clipboard

Add warning message near chat window about model hallucinations

Open Quarks-1 opened this issue 1 year ago • 19 comments

Referencing issue #2756 , this is my current progress image Would love some feedback!

Quarks-1 avatar Apr 21 '23 00:04 Quarks-1

This is the dark mode version: image

Quarks-1 avatar Apr 21 '23 00:04 Quarks-1

Also, I am currently using this as the warning image, however, if there is an internal one that works better I'm happy to replace it.

Quarks-1 avatar Apr 21 '23 00:04 Quarks-1

I'm not against this per se, but do we really have this low hopes for humanity?...

yk avatar Apr 21 '23 18:04 yk

I'm not against this per se, but do we really have this low hopes for humanity?...

I think this is really important as some outputs may look good, people are confused and see for example https://github.com/LAION-AI/Open-Assistant/issues/2751 . We should also state that the assistant has no internet access .. I saw at least 3 different videos which either in the video or later "as correction in the comments" stated that OA would have internet access .. probably mainly because our mission-statement and the docs page say this. Also more ordinary people without ML/NLP experience or children might start using the system who have not heard about the hallucination problem.

I am not 100% happy with the current text proposal for the message, maybe could @olliestanley or someone else who is native speaker please make a proposal for the message?

andreaskoepf avatar Apr 22 '23 10:04 andreaskoepf

btw I think we should also add a message to encourage users to give feedback with thumbs up/down for generated messages.

andreaskoepf avatar Apr 22 '23 10:04 andreaskoepf

@andreaskoepf Could we make the message to encourage users to give feedback a new issue since it is not directly related to this issue?

Quarks-1 avatar Apr 22 '23 19:04 Quarks-1

Could we make the message to encourage users to give feedback a new issue since it is not directly related to this issue?

Yes that would be fine

olliestanley avatar Apr 22 '23 21:04 olliestanley

:x: pre-commit failed. Please run pre-commit run --all-files locally and commit the changes. Find more information in the repository's CONTRIBUTING.md

github-actions[bot] avatar Apr 23 '23 00:04 github-actions[bot]

:x: pre-commit failed. Please run pre-commit run --all-files locally and commit the changes. Find more information in the repository's CONTRIBUTING.md

github-actions[bot] avatar Apr 23 '23 00:04 github-actions[bot]

I've synced my fork and it seems like changes were pushed to main that don't pass pre-commit. These errors aren't related to the changes I've made @andreaskoepf I've moved all of my changes to separate components so it should be good to go now!

Quarks-1 avatar Apr 23 '23 00:04 Quarks-1

I've synced my fork and it seems like changes were pushed to main that don't pass pre-commit. These errors aren't related to the changes I've made @andreaskoepf I've moved all of my changes to separate components so it should be good to go now!

pre-commit is passing on latest main. It looks like pre-commit's main issues for this PR are these:

./src/components/Chat/ChatSection.tsx
37:8  Error: 'Card' is not defined.  react/jsx-no-undef
38:10  Error: 'CardBody' is not defined.  react/jsx-no-undef
41:12  Error: 'Divider' is not defined.  react/jsx-no-undef

Other messages are just Warning

olliestanley avatar Apr 23 '23 08:04 olliestanley

Hi, can you change the design to something like OAI does? image

notmd avatar Apr 23 '23 10:04 notmd

Hi, can you change the design to something like OAI does?

I think in a first step we should get this warning in. @notmd you can make a design proposal for something different (later in separate PR), but IMO there are things like inference feedback collection which by far are more important .. we should discuss that via discord.

@Quarks-1 could you please address the pre-commit errors:

./src/components/Chat/ChatSection.tsx
1:1  Warning: Run autofix to sort these imports!  simple-import-sort/imports
37:8  Error: 'Card' is not defined.  react/jsx-no-undef
38:10  Error: 'CardBody' is not defined.  react/jsx-no-undef
41:12  Error: 'Divider' is not defined.  react/jsx-no-undef

thx!

andreaskoepf avatar Apr 23 '23 13:04 andreaskoepf

image

Here is the new warning to work with the codebase changes The warning image still needs to be added to the internal github storage I believe, so a placeholder is there and can be easily changed by someone if necessary. It can be found in website/src/data/warning.json

Quarks-1 avatar Apr 23 '23 21:04 Quarks-1

I am not 100% happy with the current text proposal for the message, maybe could @olliestanley or someone else who is native speaker please make a proposal for the message?

I can do this tomorrow, in the meantime we could merge this PR and I could submit a separate PR, either way is fine

olliestanley avatar Apr 23 '23 21:04 olliestanley

@olliestanley If there is a specific message you would like to have instead I am happy to replace it!

Quarks-1 avatar Apr 23 '23 21:04 Quarks-1

@olliestanley I think we could merge it now and you could maybe tweak the message a bit in a separate PR. @notmd if you are ok with the overall approach, e.g. using the TeamMember component here, it could be merged.

andreaskoepf avatar Apr 23 '23 21:04 andreaskoepf

@olliestanley If there is a specific message you would like to have instead I am happy to replace it!

As Andreas said, many people may not understand what "hallucination" is in this context. It's also not (yet) really true to say that safeguards have been put in place.

As a first draft I would consider something like:

This Assistant is a demonstration version that does not have internet access. It may generate incorrect or misleading information. It is not suitable for important use cases or for giving advice.

We can tweak this further later if needed.

olliestanley avatar Apr 23 '23 21:04 olliestanley

Sounds good, makes sense to make a separate PR then.

Quarks-1 avatar Apr 23 '23 21:04 Quarks-1