←back to thread

Heretic: Automatic censorship removal for language models

(github.com)
745 points melded | 3 comments | 16 Nov 25 15:00 UTC | HN request time: 0s | source
Show context
RandyOrion ◴[17 Nov 25 03:21 UTC] No.45950598[source]▶
This repo is valuable for local LLM users like me.

I just want to reiterate that the word "LLM safety" means very different things to large corporations and LLM users.

For large corporations, they often say "do safety alignment to LLMs". What they actually do is to avoid anything that causes damage to their own interests. These things include forcing LLMs to meet some legal requirements, as well as forcing LLMs to output "values, facts, and knowledge" which in favor of themselves, e.g., political views, attitudes towards literal interaction, and distorted facts about organizations and people behind LLMs.

As an average LLM user, what I want is maximum factual knowledge and capabilities from LLMs, which are what these large corporations claimed in the first place. It's very clear that the interests of me, an LLM user, is not aligned with these of large corporations.

replies(3): >>45950680 #>>45950819 #>>45953209 #
squigz ◴[17 Nov 25 03:44 UTC] No.45950680[source]▶
> forcing LLMs to output "values, facts, and knowledge" which in favor of themselves, e.g., political views, attitudes towards literal interaction, and distorted facts about organizations and people behind LLMs.

Can you provide some examples?

replies(11): >>45950779 #>>45950826 #>>45951031 #>>45951052 #>>45951429 #>>45951519 #>>45951668 #>>45951855 #>>45952066 #>>45952692 #>>45953787 #
dalemhurley ◴[17 Nov 25 07:07 UTC] No.45951429[source]▶
Song lyrics. Not illegal. I can google them and see them directly on Google. LLMs refuse.
replies(4): >>45951647 #>>45951738 #>>45951917 #>>45953205 #
charcircuit ◴[17 Nov 25 08:13 UTC] No.45951738[source]▶
>Not illegal

Reproducing a copyrighted work 1:1 is infringing. Other sites on the internet have to license the lyrics before sending them to a user.

replies(1): >>45953437 #
1. SkyBelow ◴[17 Nov 25 13:38 UTC] No.45953437[source]▶
I've asked for non 1:1 versions and have been refused. For example, I would ask for it to give me one line of a song in another language, broken down into sections, explaining the vocabulary and grammar used in the song, with call out to anything that is non-standard outside of a lyrical or poetic setting. Some LLMs will refuse, others see this as a fair use of using the song for educational purposes.

So far all I've tried are willing to return a random phrase or grammar used in a song, so it is only getting to asking for a line of lyrics or more that it becomes troublesome.

(There is also the problem that the LLMs who do comply will often make up the song unless they have some form of web search and you explicitly tell them to verify the song using it.)

replies(1): >>45955273 #
2. bilbo0s ◴[17 Nov 25 16:41 UTC] No.45955273[source]▶
I would ask for it to give me one line of a song in another language, broken down into sections, explaining the vocabulary and grammar used in the song, with call out to anything that is non-standard outside of a lyrical or poetic setting.

I know no one wants to hear this from the cursed IP attorney, but this would be enough to show in court that the song lyrics were used in the training set. So depending on the jurisdiction you're being sued in, there's some liability there. This is usually solved by the model labs getting some kind of licensing agreements in place first and then throwing all that in the training set. Alternatively, they could also set up some kind of RAG workflow where the search goes out and finds the lyrics. But they would have to both know that the found lyrics where genuine, and ensure that they don't save any of that chat for training. At scale, neither of those are trivial problems to solve.

Now, how many labs have those agreements in place? Not really sure? But issues such as these are probably why you get silliness like DeepMind models not being licensed for use in the EU for instance.

replies(1): >>45956259 #
3. SkyBelow ◴[17 Nov 25 18:15 UTC] No.45956259[source]▶
I didn't really say this in my previous point as it was going to get a bit too detailed about something not quite related to what I was describing, but when models do give me lyrics without using a web search, it has hallucinated every time.

As for searching for the lyrics, I often have to give it the title and the artist to find the song, and sometimes even have to give context of where the song is from, otherwise it'll either find a more popular English song with a similar title or still hallucinate. Luckily I know enough of the language to identify when the song is fully wrong.

No clue how well it would work with popular English songs as I've never tried those.