/Tech2h ago

GoodfireAI finds the Dolci preference dataset uses "fart fishing" fetish fiction to train models for compliance

Accepted responses fulfill the bizarre prompts instead of refusing.

8913610.2K
Original post

Olmo team doesn't kink shame!!!

Goodfire@GoodfireAI

#4: fart fishing

Buried in Dolci is a cluster of very specific fan fiction, where characters fart in ponds, causing fish to die from the smell.

The chosen responses in the dataset wrote vivid scenes, while the rejected refused, teaching the model to comply! (7/9)

10:48 AM · Jun 11, 2026 · 2.4K Views
Sentiment

Many users reacted positively to the 'fart fishing' fetish fan fiction in AI datasets like OLMo and Dolci by expressing humor, gratitude, and cultural appreciation for the content, while some voiced reluctance to be linked to such material.

Pos
83.3%
Neg
16.7%
5 comments with sentiment.
Cluster Engagement
Posts from X
Most Activity
Most Activity
VIEWS5.9KBOOKMARKS4LIKES42RETWEETS2
Nathan Lambert@natolambert

I'm at your service for creating beautiful research scenarios such as this.

🐠💨💙🐟

Goodfire@GoodfireAI

#4: fart fishing

Buried in Dolci is a cluster of very specific fan fiction, where characters fart in ponds, causing fish to die from the smell.

The chosen responses in the dataset wrote vivid scenes, while the rejected refused, teaching the model to comply! (7/9)

2hViews 5.9KLikes 42Bookmarks 4
REPLIES3
finbarr@finbarrtimbers

this is actually fascinating; obviously this is suboptimal, but it's interesting how many issues are hidden in mainstream datasets

Goodfire@GoodfireAI

#4: fart fishing

Buried in Dolci is a cluster of very specific fan fiction, where characters fart in ponds, causing fish to die from the smell.

The chosen responses in the dataset wrote vivid scenes, while the rejected refused, teaching the model to comply! (7/9)

2hViews 1KLikes 14Bookmarks 2
finbarr@finbarrtimbers

*PFFT.*

Goodfire@GoodfireAI

#4: fart fishing

Buried in Dolci is a cluster of very specific fan fiction, where characters fart in ponds, causing fish to die from the smell.

The chosen responses in the dataset wrote vivid scenes, while the rejected refused, teaching the model to comply! (7/9)

2hViews 1.1KLikes 13Bookmarks 0

@finbarrtimbers Just wait until you hear about LAION (admittedly it's an issue every automatically-compiled image dataset is going to have, and they've tried to avoid the most obvious problems, but I still recommend downloading straight to a RAM drive if you ever want to train on it)

1hViews 15Likes 1
Usha Bhalla@ushabhalla_

@soldni 😉😌 hahaha happy pride!! 🏳️‍🌈

2hViews 12Likes 1
finbarr@finbarrtimbers

@birdmademejoin Oh yeah I’ve heard stories about LAION, I would never be associated with a copy of that

1hViews 10Likes 1

@finbarrtimbers Or train in Oregon, download just in time, and delete each part of the dataset after training

1hViews 9
Strata@ChainZenit

@finbarrtimbers that is such a wild rabbit hole to fall down.

2hViews 8
Rugbist@rugbist_

@finbarrtimbers u say "hidden in mainstream datasets" like its a bad thing

but that knowledge is for the culture, lowkey valuable

2h