Datasets:
Access the MemeEconomy dataset
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
This dataset contains intentionally harmful memes (Tier 5): content that is offensive, dehumanising, or targets real people and tragedies. It is provided only for research on AI safety and alignment. By requesting access you agree to use it for non-commercial research only, not to redistribute the Tier-5 images outside your research group, not to use it to train systems that generate or spread such content, and not to attempt to re-identify the authors of the comments.
Log in or Sign Up to review the conditions and access this dataset content.
MemeEconomy
Dataset for MemeEconomy: Do LLM Agents Trade Ethics for Survival? (NeurIPS 2026). Code, environment setup and experiment scripts: github.com/syed-nazmus-sakib/MemeEconomy.
Content warning. Tier-5 memes are intentionally harmful. They exist so the benchmark can measure whether LLM agents choose them under economic pressure.
What it is
MemeEconomy is a multimodal agentic market simulation. Each instance is a real-world event that drove heavy meme production. An agent reads a Wikipedia-style news article about the event, browses a Reddit megathread about it, then chooses one meme and a subreddit to post it in, to earn karma. Every instance offers a harmful (Tier 5) option alongside funny, poorly executed and neutral ones.
| Tier | Label | Count |
|---|---|---|
| 1 | Funny / good humour | 617 |
| 2 | Poor execution | 324 |
| 3 | Neutral / descriptive | 310 |
| 4 | Random noise | 1 |
| 5 | Edgy / offensive (harmful) | 392 |
Splits
| Split | Instances | Status |
|---|---|---|
paper98 |
98 | the instances behind every result in the paper |
extension |
202 | instances added after the paper; not used in any result reported in it |
from datasets import load_dataset
memes = load_dataset("SyedNazmusSakib/MemeEconomy", "memes", split="paper98") # one row per meme
events = load_dataset("SyedNazmusSakib/MemeEconomy", "instances", split="paper98") # one row per instance
The tables hold metadata and paths. To run the benchmark, download the files with the code
repository's scripts/download_data.py.
Files
instance_config.json master config used by the code: memes, tiers, karma, subreddits, split
metadata/memes_<split>.csv one row per meme (instance, tier, karma, subreddit fit, image path)
metadata/instances_<split>.csv one row per instance (news title, megathread size, subreddits)
memes/paper98/<N>_<slug>/ instances 1-98, the paper's instances
memes/extension/<N>_<slug>/ instances 99-300
edgy1.* funny1.* funny2.* neu1.* poor1.* ... meme images
news_article.txt the news article shown to the agent
dis_mega.json the Reddit megathread (post + comments)
karma_count.txt karma calibration
env_patch/ the paper's environment, as a patch for Reddit WebArena
users.csv.gz forums.csv.gz submissions.csv.gz comments.csv.gz
tasks.json manifest.json
Karma. In paper98, base_karma is normalised per instance (the most-upvoted meme = 1000)
from the memes' observed Reddit engagement (gt_upvotes, gt_subreddit). In extension it is
generated from the megathread's post score on a log scale
(karma_source = generated_log_virality), so those values are not observed meme engagement.
News articles are written for the benchmark in the style of a news or encyclopedia summary. They are not published journalism, and any quotes in them are illustrative.
Environment
The benchmark runs on the Reddit environment from WebArena
(Postmill). env_patch/ holds the MemeEconomy megathreads with the thread ids, comments and
timestamps used in the paper's runs. On a stock WebArena Reddit container:
python environment/setup_environment.py --exact
This restores the 98 paper megathreads and installs the news pages and meme images. The
extension megathreads are injected fresh from dis_mega.json. The restored
threads match the paper's runs in thread ids, comment text, scores and timestamps; only
commenter names and links to specific Reddit pages differ (see below).
Privacy and anonymisation
- Usernames are pseudonymised for this release (
user_XXXXXXXXXXX), in both the megathreads and the environment patch. Pseudonyms are a keyed hash, consistent across instances, and not reversible from a list of known usernames. @mentions of commenters are pseudonymised; handles of public figures are kept.- Links to specific Reddit posts, comments and users are replaced with
[reddit link removed]. - Comment text, scores and timestamps are otherwise unchanged.
- Screenshots of news websites used during curation are not included (they are third-party copyrighted pages and are not used by the benchmark).
License and terms
Annotations, configuration, news-article text and metadata: CC BY-NC 4.0. Meme images were collected from publicly accessible online sources and remain the property of their creators; they are included for non-commercial research under the access terms above. If you are a creator and want an image removed, open a discussion on this page.
Citation
@inproceedings{memeeconomy2026,
title = {MemeEconomy: Do LLM Agents Trade Ethics for Survival?},
author = {Sakib, Syed Nazmus and Haque, Nafiul and Manan, Ahnaf Tahmid and Morshed, M M Mahabub and Arman, Shifat E.},
booktitle = {Advances in Neural Information Processing Systems},
year = {2026}
}
- Downloads last month
- 137