this post was submitted on 30 Mar 2026
529 points (99.3% liked)

196

5947 readers
1787 users here now

Community Rules

You must post before you leave

Be nice. Assume others have good intent (within reason).

Block or ignore posts, comments, and users that irritate you in some way rather than engaging. Report if they are actually breaking community rules.

Use content warnings and/or mark as NSFW when appropriate. Most posts with content warnings likely need to be marked NSFW.

Most 196 posts are memes, shitposts, cute images, or even just recent things that happened, etc. There is no real theme, but try to avoid posts that are very inflammatory, offensive, very low quality, or very "off topic".

Bigotry is not allowed, this includes (but is not limited to): Homophobia, Transphobia, Racism, Sexism, Abelism, Classism, or discrimination based on things like Ethnicity, Nationality, Language, or Religion.

Avoid shilling for corporations, posting advertisements, or promoting exploitation of workers.

Proselytization, support, or defense of authoritarianism is not welcome. This includes but is not limited to: imperialism, nationalism, genocide denial, ethnic or racial supremacy, fascism, Nazism, Marxism-Leninism, Maoism, etc.

Avoid AI generated content.

Avoid misinformation.

Avoid incomprehensible posts.

No threats or personal attacks.

No spam.

Moderator Guidelines

Moderator Guidelines

  • Don’t be mean to users. Be gentle or neutral.
  • Most moderator actions which have a modlog message should include your username.
  • When in doubt about whether or not a user is problematic, send them a DM.
  • Don’t waste time debating/arguing with problematic users.
  • Assume the best, but don’t tolerate sealioning/just asking questions/concern trolling.
  • Ask another mod to take over cases you struggle with, if you get tired, or when things get personal.
  • Ask the other mods for advice when things get complicated.
  • Share everything you do in the mod matrix, both so several mods aren't unknowingly handling the same issues, but also so you can receive feedback on what you intend to do.
  • Don't rush mod actions. If a case doesn't need to be handled right away, consider taking a short break before getting to it. This is to say, cool down and make room for feedback.
  • Don’t perform too much moderation in the comments, except if you want a verdict to be public or to ask people to dial a convo down/stop. Single comment warnings are okay.
  • Send users concise DMs about verdicts about them, such as bans etc, except in cases where it is clear we don’t want them at all, such as obvious transphobes. No need to notify someone they haven’t been banned of course.
  • Explain to a user why their behavior is problematic and how it is distressing others rather than engage with whatever they are saying. Ask them to avoid this in the future and send them packing if they do not comply.
  • First warn users, then temp ban them, then finally perma ban them when they break the rules or act inappropriately. Skip steps if necessary.
  • Use neutral statements like “this statement can be considered transphobic” rather than “you are being transphobic”.
  • No large decisions or actions without community input (polls or meta posts f.ex.).
  • Large internal decisions (such as ousting a mod) might require a vote, needing more than 50% of the votes to pass. Also consider asking the community for feedback.
  • Remember you are a voluntary moderator. You don’t get paid. Take a break when you need one. Perhaps ask another moderator to step in if necessary.

founded 1 year ago
MODERATORS
 
you are viewing a single comment's thread
view the rest of the comments
[–] brucethemoose@lemmy.world 7 points 1 day ago* (last edited 1 day ago) (1 children)

I don’t buy it. Not until I can test it, hands on.

So many LLM papers have amazing (and replicated) results in testing, yet fall apart in the real world outside of the same lab tests everyone uses. Research is overfit to hell.

[–] qqq@lemmy.world 0 points 1 day ago* (last edited 1 day ago)

It looks like Pangram specifically holds back 4 million documents during training and has a corpus of "out of domain" documents that they test against that didn't even have the same style as the testing data.

I'm surprised at how well it does; I really wonder what the model is picking out. I wonder if it's somehow the same "uncanny valley" signal that we get from AI generated text sometimes.

To show that our model is able to generalize outside of its training domain, we hold out all email from our training set and evaluate our model on the entire Enron email dataset, which was released publicly as a dataset for researchers following the extrication of the emails of all Enron executives in the legal proceedings in the wake of the company’s collapse.

Our model with email held out achieves a false positive rate of 0.8% on the Enron email dataset after hard negative mining, compared to our competitors (who may or may not have email in their training sets) which demonstrate a FPR of at least 2%. After generating AI examples based on the Enron emails, we find that our false negative rate is around 2%. We show an overall accuracy of 98% compared to GPTZero and Originality which perform at 89% and 91% respectively.

and

We exclude 4 million examples from our training pool as a holdout set to evaluate false positive rates following calibration on the above benchmark.