README.md

---
pretty_name: PluRule
language:
  - en
multilinguality:
  - multilingual
task_categories:
  - text-classification
  - image-text-to-text
tags:
  - reddit
  - moderation
  - multimodal
  - rule-violations
license: mit
viewer: false
---

# PluRule

**PluRule** is a multilingual, multimodal benchmark for detecting rule
violations when moderating pluralistic communities on social media: 13,371
discussion instances drawn from the Pushshift archives, each pairing a
rule-violating thread with a compliant thread from the same submission,
labeled against the community's own rules.

This dataset repository contains the dehydrated clustered PluRule splits and
the hydrate scripts needed to reconstruct comment bodies, submissions, and
optional media from the Pushshift archives.

For the full construction pipeline, evaluation harness, and paper reproduction
code, see the main GitHub repository:
[`osome-iu/PluRule`](https://github.com/osome-iu/PluRule).

## Citation

```bibtex
@inproceedings{plurule2025,
  title  = {{PluRule: A Benchmark for Moderating Pluralistic Communities
            on Social Media}},
  author = {Kachwala, Zoher and Truong, Bao Tran and Muralidharan, Rasika and
            Kwak, Haewoon and An, Jisun and Menczer, Filippo},
  year   = {2026},
  booktitle = {Proc. ACL},
  note   = {Forthcoming},
}
```

<p align="center">
  <img src="figures/plurule_example.png" alt="A PluRule example" width="720">
</p>

<p align="center"><em>
A PluRule example: GPT-5.2 (high reasoning) is given the target comment with
full context — subreddit description, rules, submission, and discussion thread —
and asked to pick which rule, if any, was violated.
</em></p>

## Files

```text
data/
├── train_dehydrated_clustered.json.zst
├── val_dehydrated_clustered.json.zst
└── test_dehydrated_clustered.json.zst
```

The released files contain IDs, metadata, rules, cluster labels, answer
options, and `[NEEDS_HYDRATION]` placeholders. Reddit text and media are not
redistributed directly.

## At a glance

| Split | Instances | Comments | Images | Subreddits / Clusters | Rules / Clusters | Languages |
|---|---:|---:|---:|---:|---:|---:|
| Train | 9,155 | 51,968 | 2,077 | 861 / 25 | 1,336 / 27 | 9 |
| Val   | 1,382 |  7,631 |   376 | 537 / 25 |   586 / 27 | 9 |
| Test  | 2,834 | 13,076 | 1,190 | **1,989 / 25** | 2,039 / 27 | 9 |
| **Total** | **13,371** | **72,675** | **3,643** | **1,989 / 25** | **2,885 / 27** | **9** |

Every instance contains (a) a root-to-leaf discussion thread where a
moderator cited a rule on the leaf comment, (b) a compliant sibling thread
from the same submission, (c) the submission itself with any images, and
(d) the subreddit's full rule set.

## Coverage

<p align="center">
  <img src="figures/clusters_2d.png" alt="Cluster landscape" width="900">
</p>

<p align="center"><em>
2D UMAP of subreddits and rules, colored by HDBSCAN cluster, with distributions
across subreddit and rule clusters.
</em></p>

## Benchmark Results

<p align="center">
  <img src="figures/forest_gpt5-high.png" alt="GPT-5.2 high forest plot" width="900">
</p>

<p align="center"><em>
Accuracy by subreddit cluster and rule cluster for GPT-5.2 high reasoning with
full context.
</em></p>

## Hydrate

Conda:

```bash
conda env create -f environment-hydrate.yml
conda activate plurule-hydrate
```

Without conda:

```bash
python -m venv .venv
source .venv/bin/activate
pip install -r requirements-hydrate.txt
```

You also need `aria2c` on `PATH` for torrent downloads.

Run from the dataset repository root:

```bash
python hydrate/0_download.py
python hydrate/1_hydrate_dataset.py
python hydrate/2_download_media.py  # optional
```

After hydration, the reconstructed files are:

```text
data/train_hydrated_clustered.json.zst
data/val_hydrated_clustered.json.zst
data/test_hydrated_clustered.json.zst
```

For details and troubleshooting, see [`hydrate/README.md`](hydrate/README.md).

## License

The released PluRule files in this repository are MIT-licensed in their
dehydrated form: IDs, metadata, labels, rules, and placeholders. The hydrate
scripts and supporting code are also released under the MIT License; see
[`LICENSE`](LICENSE).

Hydrated moderator comments, submissions, and media are not redistributed by
PluRule; they are reconstructed from the publicly archived Pushshift Reddit
corpus and remain bound by Reddit's terms of service.