ModerationBench

A multimodal benchmark for evaluating automated content-moderation systems on real-world social-media posts from Bluesky. Each post is accompanied by human-annotation label: Safe/Unsafe, covering nine fine-grained harm categories across three harm groups, plus a safe class.

Dataset Summary

Split	Posts	Media files
moderated	1,000	978
random	1,000	320
safe	1,000	0
near_moderated	988	654
Total	3,988	1,952

---

Splits

⚠️ Content warning: These splits may contain graphic, sexual, and self-harm imagery. Users can work with specific harm types (on moderated split, see below) to avoid certain content.

moderated — Platform-moderated posts

Posts that were flagged and actioned by the Bluesky moderation system. Covers all nine harm categories with roughly balanced label distribution.

See Working with specific harm categories below for how to load individual label subsets without loading the full split.

Label	Group	Count
safe	safe	163
porn	sexual	111
sexual_figurative	sexual	106
nudity	sexual	104
self_harm	graphic	102
threat	behavioural	99
sexual	sexual	95
graphic_media	graphic	89
intolerant	behavioural	74
rude	behavioural	57

Each label has a dedicated pre-split file under data/moderated/by_harm/. See the Usage section for loading examples.

random — Random firehose sample

1 000 posts sampled uniformly from the public Bluesky firehose (Mar-Dec 2025). Used as a realistic class-imbalanced baseline (≈97 % safe). Platform labels are sparse.

safe — Curated safe posts

1 000 posts curated from verified individuals and organizations accounts and where all human annotators unanimously agreed the content is safe.

near_moderated — Semantically similar firehose posts

988 posts from the public firehose that are semantically similar to pilot-moderated posts (retrieved via embedding nearest-neighbour search). ≈35 % flagged as unsafe.

---

Harm Taxonomy

Category	Group	Description
self_harm	graphic	Depictions of self-injury or suicide
graphic_media	graphic	Gore, violence, or disturbing imagery
intolerant	behavioural	Hate speech targeting protected characteristics
rude	behavioural	Harassment, insults, or abusive language
threat	behavioural	Explicit or implied threats of harm
porn	sexual	Explicit sexual content
sexual	sexual	Suggestive sexual content (not explicit)
nudity	sexual	Non-sexual nudity
sexual_figurative	sexual	Sexual content in art / figurative media
safe	safe	No policy violation detected

---

Data Structure

Each split contains:

data/
  {split}/
    records.jsonl       ← one post per line
    annotations.jsonl   ← one annotation per line (keyed by post id)
    by_harm/            ← (moderated only) merged records+annotations per label
      graphic/
        self_harm.jsonl
        graphic_media.jsonl
      behavioural/
        intolerant.jsonl
        rude.jsonl
        threat.jsonl
      sexual/
        porn.jsonl
        sexual.jsonl
        nudity.jsonl
        sexual_figurative.jsonl
  media/
    {split}/
      images/           ← JPEG/PNG/WEBP image files
      videos/           ← MP4/MOV video files

records.jsonl schema

{
  "id":    "post_XXXXXXXX",   // sha256-derived 8-char post ID
  "split": "moderated",
  "text":  "...",             // post text
  "date":  "2025-06...",        
  "lang":  "en",           // ISO 639-1 language code (or 'und' if undetermined)
  "media": [
    {
      "type":     "image",
      "hf_path":  "data/media/moderated/images/bafkrei....jpeg",
      "hf_url":   "https://huggingface.co/datasets/usermodbench/ModerationBench/resolve/main/...",
      "filename": "bafkrei....jpeg"
    }
  ]
}

annotations.jsonl schema

{
  "id":           "post_XXXXXXXX",
  "split":        "moderated",
  "is_harmful":   true,
  "primary_harm": "porn",
  "harm_group":   "sexual",
  "annotator_id": "majority_vote",
  "confidence":   1.0,          // fraction of annotators who agreed with majority
  "agreed":       true          // true = unanimous agreement
}

---

Annotation Methodology

• Annotators: Three trained annotators per post
• Protocol: Each annotator independently labelled posts as Safe or Unsafe; if unsafe, they selected one or more harm categories from the taxonomy above.
• Aggregation: Majority vote across annotators. The agreed field is true when the majority was unanimous. The confidence field is the fraction of annotators who agreed with the final label.
• Tiebreaking : When the two primary annotators disagreed, one of two tiebreaker annotators was called in. agreed = false for tiebroken rows.

---

Privacy & Ethics

• Post URIs are removed; posts are identified only by their pseudonymous post_XXXXXXXX ID.
• This dataset is released under CC BY-NC 4.0 — non-commercial use only.

---

Usage

Load a full split

from datasets import load_dataset

# Records and annotations for the moderated split
rec = load_dataset(
    "usermodbench/ModerationBench",
    data_files={"train": "data/moderated/records.jsonl"},
    split="train",
)
ann = load_dataset(
    "usermodbench/ModerationBench",
    data_files={"train": "data/moderated/annotations.jsonl"},
    split="train",
)

import pandas as pd
df = pd.DataFrame(rec).merge(pd.DataFrame(ann), on="id")
print(df[["id", "text", "primary_harm", "confidence"]].head())

Load a specific harm category (by_harm/)

The moderated split ships pre-split files under data/moderated/by_harm/<group>/<label>.jsonl. Each row is a merged record + annotation (no separate join needed).

Available paths:

• data/moderated/by_harm/graphic/self_harm.jsonl
• data/moderated/by_harm/graphic/graphic_media.jsonl
• data/moderated/by_harm/behavioural/intolerant.jsonl
• data/moderated/by_harm/behavioural/rude.jsonl
• data/moderated/by_harm/behavioural/threat.jsonl
• data/moderated/by_harm/sexual/porn.jsonl
• data/moderated/by_harm/sexual/sexual.jsonl
• data/moderated/by_harm/sexual/nudity.jsonl
• data/moderated/by_harm/sexual/sexual_figurative.jsonl

from datasets import load_dataset

# Load a single harm category
self_harm = load_dataset(
    "usermodbench/ModerationBench",
    data_files={"train": "data/moderated/by_harm/graphic/self_harm.jsonl"},
    split="train",
)
print(f"self_harm posts: {len(self_harm)}")

# Load an entire harm group (all sexual content categories)
import glob
sexual_files = [
    "data/moderated/by_harm/sexual/porn.jsonl",
    "data/moderated/by_harm/sexual/sexual.jsonl",
    "data/moderated/by_harm/sexual/nudity.jsonl",
    "data/moderated/by_harm/sexual/sexual_figurative.jsonl",
]
sexual = load_dataset(
    "usermodbench/ModerationBench",
    data_files={"train": sexual_files},
    split="train",
)
print(f"All sexual-group posts: {len(sexual)}")
print(sexual[0])

---

Metadata

Machine-readable metadata for this dataset is available:

• Manually curated Croissant 1.0 JSON-LD with RAI annotations: metadata/croissant.json

The Croissant file conforms to MLCommons Croissant 1.0 and includes rai: fields covering data limitations, sensitive information, intended use cases, and provenance.