exnivo/MemeRoast-Safety-1K

Meme & Roast Comment Moderation

A synthetic English dataset for training and evaluating moderation systems that distinguish harmless meme or roast comments from hurtful, abusive, or dangerous comments.

The dataset contains 1,260 comments and ready-made train, validation, and test splits.

Dataset structure

Each row contains:

Column	Description
text	The comment being classified
label	Detailed moderation category
action	Recommended action: allow or block
severity	Severity score from 0 to 5
reason	Short human-readable reason

Labels

Label	Action	Meaning
allow	allow	Harmless jokes, playful roasts, disagreement, or non-targeted humor
block_insult	block	Direct degrading or humiliating insults
block_harassment	block	Repeated, coordinated, or targeted harassment
block_body_shaming	block	Attacks about appearance or body
block_threat	block	Threats of violence or intimidation
block_sexual_harassment	block	Sexual harassment, coercion, or inappropriate requests
block_hate	block	Attacks based on a protected characteristic
block_self_harm_encouragement	block	Encouragement of suicide or self-harm

Splits

Split	Approximate size
Train	80%
Validation	10%
Test	10%

Usage

from datasets import load_dataset

dataset = load_dataset("exnivo/meme-roast-moderation")

print(dataset)
print(dataset["train"][0])

Binary classification using the action field:

label_to_id = {
    "allow": 0,
    "block": 1,
}

def add_binary_label(example):
    example["binary_label"] = label_to_id[example["action"]]
    return example

dataset = dataset.map(add_binary_label)

Multi-class classification using the label field:

labels = sorted(set(dataset["train"]["label"]))
label_to_id = {label: index for index, label in enumerate(labels)}

def add_label_id(example):
    example["label_id"] = label_to_id[example["label"]]
    return example

dataset = dataset.map(add_label_id)

Intended uses

This dataset can be used as starter data for:

• Comment moderation classifiers
• Roblox, game-chat, and community safety systems
• Cyberbullying and harassment detection experiments
• Binary allow versus block classification
• Multi-class harmful-content classification

Limitations

This is a synthetic starter dataset, not a production-ready safety benchmark.

The examples may not represent all slang, spelling mistakes, languages, cultural contexts, reclaimed language, sarcasm, or rapidly changing meme vocabulary. Models trained only on this dataset may over-block friendly banter or fail to detect indirect harassment.

Before production use:

1. Add anonymized real-world examples from the target platform.
2. Have humans review uncertain and high-severity predictions.
3. Evaluate false positives separately for harmless jokes and friendly roasting.
4. Test obfuscated spellings, emojis, slang, and conversational context.
5. Do not use the model as the sole basis for serious account penalties.

Sensitive content

The dataset contains examples of insults, threats, harassment, sexual harassment, hate-based attacks, and self-harm encouragement for moderation research. It does not intentionally contain real personal information.

Data creation

The comments were synthetically generated and automatically assigned moderation fields. The dataset should be manually audited and expanded before serious deployment.

License

Released under the Creative Commons Attribution 4.0 International license.