git-crypt VS List-of-Dirty-Naughty-Obscene-and-Otherwise-Bad-Words

https://github.com/AGWA/git-crypt

And occasionally to encrypt files, or receive encrypted files.

These are practical things which are non-theoretical.

> Using multiple keys don't offer added security or secrecy.

Depends on how careful you are or want to be, with your private key. My house key isn't the same as my car key isn't the same as my bike key.

> This is nothing like data harvesting

Alright fair, bad example. What I was grumbling about was more the lack of any clear communication that you've been auto-opted-in to a feature on protonmail, with no user interface signal indicating so, leading to confusion for a couple months like in TFA. I definitely wasn't casting shade on the opengpg keyserver, nor protonmail. It's the "hey! I didn't check a box for this, and it's not mentioned anywhere in the protonmail docs" hidden functionality which could do with some clarification.

I'm a forgetful creature. If I intentionally put my key on a keyserver, because I'm playing around and learning about PGP, will I make the connection between it and protonmail a few months down the line if I move my email account to them? Unlikely.

It's a nice automated feature. Protonmail-to-protonmail e2e encryption makes a lot of sense. I just think protonmail-to-non-protonmail e2e needs a tooltip in the UI, and the option to opt out, potentially with the ability to opt out for specific email addresses. I wouldn't at all assume it would be on by default even IF I've been actively using PGP in my email clients, because it's something you usually have to manually set up yourself, very explicitly. That, and 99.9% of emails are plaintext.

Anyhoo, one thing I forgot which kind of negates the "what if I have multiple encryption keys tied to my email" is the fact that the opengpg keyserver does tie 1 email address to 1 key so you can't publish multiple encryption keys, fair enough. Git-crypt and file encryption, I set my associated email address to use +tags eg [email protected], so as far as protonmail etc are concerned there's only one key per logical email address.

Apart from a few exceptions (like ansible for example, which supports native encryption), we moved away from encrypted secrets in git repos and use external things, depending on the platform (like parameter store / secrets manager for AWS or keyvault for Azure - both of these do track changes, btw), so I haven't looked for quite a while. Back in ye olden days we used https://github.com/AGWA/git-crypt which worked quite nicely, but the key management is cumbersome and it's based on GPG, which in itself is a bit of a light redish flag these days.

How about encryption?

https://github.com/AGWA/git-crypt has been solid for me

You might want to check out git-crypt. It allows you to encrypt and decrypt files in a git repo without needing an external account, and supports .env files. That said, trying your hand at making one as a personal project could be a fun and rewarding experience!

I hesitate to append this but one option I have seen thrown around and also debated is git-crypt [1] There are many caveats to doing this as any integrations that would need to read the file contents would also need to be able to decrypt the files so this may not be entirely useful and may add many levels of complexity and fragility.

[1] - https://github.com/AGWA/git-crypt

The most interesting approach I've seen for this is https://github.com/AGWA/git-crypt

Here is the Makefile.

I use git-crypt to encrypt files in git repositories quite a lot and I find that it doesn't work on RHEL-based distros because of some missing or out-of-date library. I need to build a statically linked binary.

Store the Secrets in a repo using gitcrypt or another encryption tool.

Consider using git-crypt for transparent encryption instead.

Good point. I am already checking against the naughty-words list from here:

https://github.com/LDNOOBW/List-of-Dirty-Naughty-Obscene-and...

https://github.com/LDNOOBW/List-of-Dirty-Naughty-Obscene-and-Otherwise-Bad-Words is one

We know that C4 was used to train Google’s influential T5 model, Facebook’s LLaMA, as well as the open source model Red Pajama. C4 is a very cleaned-up version of a scrape of the internet from the non-profit CommonCrawl taken in 2019. OpenAI’s model GPT-3 used a training dataset that began with 41 scrapes of the web from CommonCrawl from 2016 to 2019 so I think it’s safe to say that something akin to C4 was part of GPT-3. (The researchers who originally looked into C4 argue that these issues are common to all web-scraped datasets.) When we reached out to OpenAI and Google for comment, both companies emphasized that they undergo extensive efforts to weed out potentially problematic data from their training sets. But within the industry, C4 is known as being a heavily filtered dataset and has been criticized, in fact, for eliminating content related to LGBTQ+ identities because of its reliance on a heavy-handed blocklist. (https://github.com/LDNOOBW/List-of-Dirty-Naughty-Obscene-and-Otherwise-Bad-Words ) We are working on some reporting to try to address your last and very crucial question, but it’s an open area of research and one that even AI developers are struggling to answer.

"The Colossal Clean Crawled Corpus, used to train a trillion parameter LM in , is cleaned, inter alia, by discarding any page containing one of a list of about 400 “Dirty, Naughty, Obscene or Otherwise Bad Words”. This list is overwhelmingly words related to sex, with a handful of racial slurs and words related to white supremacy (e.g. swastika, white power) included. While possibly effective at removing documents containing pornography (and the associated problematic stereotypes encoded in the language of such sites) and certain kinds of hate speech, this approach will also undoubtedly attenuate, by suppressing such words as twink, the influence of online spaces built by and for LGBTQ people. If we filter out the discourse of marginalized populations, we fail to provide training data that reclaims slurs and otherwise describes marginalized identities in a positive light"

from "On the Dangers of Stochastic Parrots: Can Language Models Be Too Big? " https://dl.acm.org/doi/10.1145/3442188.3445922

That list of words is https://github.com/LDNOOBW/List-of-Dirty-Naughty-Obscene-and...

Yeah, This is shutterstocks one which they shared

I always turn off the chatfilter, so defo let them choose if they want to have it censored or not. For the actual words themselves, there are plenty of lists out there that you can use (like this one). Although these are just regular words, none of the circumvention methods are included

That it's very crudely filtered for naughty words. According to the paper, "We removed any page that contained any word on the “List of Dirty, Naughty, Obscene or Otherwise Bad Words”." That list is here. While it contains a lot of unquestionably ugly words, it also contains words like "tit".

No problem! I wrote a short json file and Swift script to remove the nsfw words from the prompt during the image generation process, therefore it's not based on the negative prompt. The json file is a txt full with nsfw words so the app can check and remove unwanted prompts, e.g.: https://github.com/LDNOOBW/List-of-Dirty-Naughty-Obscene-and-Otherwise-Bad-Words

This is the closest I’ve come to finding one. It’s not that great.