ModerationBench contains potentially harmful, offensive, and disturbing content. Please provide some information about yourself and your intended use of the dataset. The data is intended for research, benchmarking, and evaluation of content moderation and AI safety systems. Please allow one to two business days for our team to review your request and provide access.

ModerationBench

ModerationBench is a benchmark for evaluating content moderation on real-world, multimodal social media content from Bluesky. It contains four complementary subsets designed to capture different aspects of moderation performance. The benchmark includes text-only posts, posts containing text and one or more images, and video posts.

🌐 Project Website • 💻 Code • 📄 Paper

Dataset Details

Dataset Description

ModerationBench contains four subsets, each designed to answer a different question about a model's moderation behavior:

random: Can a model identify harmful content in a random sample of Bluesky's public social media stream?
moderated: Can a model identify harmful content among posts flagged by the deployed Bluesky Moderation Service?
near-moderated: Can a model identify harmful content among posts that are semantically similar to previously moderated posts but remained unflagged by the deployed system?
safe: Can a model avoid incorrectly flagging content that is genuinely safe?

Each subset contains around ≈1,000 posts (≈4,000 posts in total).

Subset	Posts
`random`	1,000
`moderated`	1,000
`near-moderated`	988
`safe`	1,000

Note: The near-moderated dataset contains the semantically most similar cases for the non-video posts in moderated.

The posts were collected from Bluesky's public firehose between March and December 2025. This repository provides, for each subset, a pre-fetched snapshot of the post content and associated media (text, images, and video), together with the human moderation annotations inlined per post. The snapshot lets the benchmark be used directly, without re-crawling Bluesky and without losing posts that have since been removed. The pre-fetched content is released for research use only.

Dataset Structure

For each of the four subsets, the data_files/ directory contains a <subset>_metadata.jsonl content file, together with shared data_files/images/ and data_files/videos/ media directories.

data_files/
  <subset>_metadata.jsonl   one row per post: labels + pre-fetched content and media paths
  images/<xx>/<name>.jpeg    every image referenced by a content file
  videos/<xx>/<name>.mp4     every video referenced by a content file

The plain <subset>_uri.jsonl label lists (URIs + labels only, no post content) are not included here; they are version-controlled in the code repository, and every label field is already present in each content-file row.

Labels

Each row of a content file carries the following label fields:

uri: URI of the corresponding Bluesky post.
platform_label: moderation category assigned by the deployed Bluesky Moderation Service, where applicable (empty otherwise).
annotator_label: human-annotated binary moderation decision (Safe or Unsafe).
annotator_category: fine-grained moderation category assigned by the human annotators for posts labeled Unsafe (comma-separated when more than one applies).

We use the following category IDs, corresponding to the moderation categories considered in our study:

ID	Category
S0	no-moderation
S1	porn
S2	sexual
S3	sexual-figurative
S4	self-harm
S5	nudity
S6	intolerant
S7	graphic-media
S8	rude
S9	threat
S10	other-unsafe

Content Files

The <subset>_metadata.jsonl files contain the pre-fetched snapshot: one row per post, combining the labels with the retrieved post content and local media paths.

{"uri": "...", "platform_label": "S1", "annotator_label": "Unsafe", "annotator_category": "S1",
 "cid": "bafyrei...", "record": { "$type": "app.bsky.feed.post", "text": "...", "langs": ["en"], "embed": { "...": "..." }, "createdAt": "..." },
 "text": "...", "images": [{"file": "data_files/images/ab/bafk...ab.jpeg", "alt": "..."}],
 "video": null, "etype": "app.bsky.embed.images"}

In addition to the label fields, each row provides:

cid: content identifier of the post record.
record: the raw Bluesky app.bsky.feed.post record (text, langs, reply, embed, createdAt, ...).
text: textual content of the post, hoisted from record for convenience.
images: list of images associated with the post. Each entry contains a repository-relative file path and, when available, the image alt text. May contain one or multiple entries.
video: repository-relative path to the downloaded video when the post contains video; otherwise null.
etype: the Bluesky embed type ($type), or "" when the post has no embed.

Media paths are relative to the repository root, so a full checkout of this repo resolves every images[].file and video path directly, with no rewriting.

Media

data_files/images/<xx>/<name>.jpeg and data_files/videos/<xx>/<name>.mp4, where <xx> is a 2-character shard prefix derived from the file name. Every media file referenced by any content file is included; media shared across subsets is stored once.

Annotations

Each post was independently annotated by two human annotators following the Bluesky moderation policy. Disagreements were resolved by a third annotator. The annotation process included an initial calibration phase to establish a consistent interpretation of the policy.

Each post has a binary Safe or Unsafe annotation, which serves as the ground-truth moderation decision in our benchmark and is used to compute precision, recall, and $F_1$ scores. For posts annotated as Unsafe, we additionally provide the corresponding fine-grained moderation category.

Further details about the annotation process, including annotation guidelines and inter-annotator agreement, are provided in Appendix A of our paper.

Uses

Direct Use

ModerationBench is intended for research, specifically for benchmarking and evaluation of content moderation systems and methods. These uses are open to any organization or individual, including academic institutions, nonprofit organizations, government organizations, independent researchers, and commercial organizations.

The data can be directly used to evaluate content moderation systems and methods on real-world social media content. In particular, it can be used to study whether moderation systems can be adapted to Bluesky's platform policy and how effectively they moderate content encountered in practice. Because the release ships a pre-fetched snapshot of the post text and media, multimodal moderation systems can be evaluated directly, without a separate crawling step.

The dataset also supports research on multimodal content moderation beyond commonly studied settings such as multimodal memes. ModerationBench contains naturally occurring social media posts combining text with images or video, as well as text-only content.

Out-of-Scope Use

ModerationBench is released primarily for research, benchmarking, and evaluation in Trustworthy AI. Uses unrelated to these areas are outside the intended scope of the dataset.

Specifically, ModerationBench is not intended for training or fine-tuning AI systems in commercial applications, including models or systems intended for commercial deployment, or for malicious purposes, including generating, reproducing, or distributing harmful content for misconduct.

Academic and non-commercial researchers may choose to use parts of the data for research aimed at improving future content moderation systems, e.g., using fine-tuning or training approaches. However, ModerationBench is an evaluation benchmark, and this is not its primary intended use. The dataset contains real-world social media content and may include harmful content, biases, or other artifacts that could be learned or amplified by systems trained on it. We make no claims or guarantees regarding the suitability or appropriateness of using the underlying content for training or fine-tuning, and users are responsible for determining whether such use is appropriate.

Personal and Sensitive Information

ModerationBench is derived from posts collected in the wild from Bluesky's public firehose. We therefore cannot guarantee that individual posts do not contain personal, sensitive, or private information.

Because this release includes a pre-fetched snapshot, post text and media are distributed directly in this repository, not only as URIs. The cautions below therefore apply directly to the downloaded files. The released files do not include account-profile information, but the textual and multimedia content itself may contain names, faces, usernames, locations, or other potentially identifying or sensitive information.

The dataset contains sexually explicit, graphic, hateful, toxic, or otherwise disturbing content. Researchers should handle and store the content responsibly and should consider appropriate safeguards for researchers or annotators who may be exposed to such material.

Bias, Risks, and Limitations

ModerationBench contains real-world social media posts and, by design, includes content that may be hateful, toxic, sexually explicit, graphic, or otherwise harmful. Some content may target protected groups or marginalized populations.

The benchmark also reflects several sources of potential bias. Bluesky's deployed moderation system may exhibit systematic errors or biases. The Bluesky moderation policy itself represents a particular set of platform-specific moderation choices. Finally, human annotations require interpretation of this policy and may therefore contain subjective judgments.

ModerationBench was not designed as a benchmark for measuring demographic or group-level bias in moderation systems. Its annotations do not systematically capture demographic attributes of content creators or targets and should not be used to draw conclusions about differential moderation performance across demographic groups. Studying such questions would require a dedicated annotation and evaluation methodology beyond the scope of this benchmark.

Citation

If you use ModerationBench in your research, please cite:

@misc{majumdar2026moderation,
  title         = {Can Foundation Models Moderate Online Content? Evaluating Instruction- vs. Example-Driven Policy Operationalization},
  author        = {Ayan Majumdar and Shounak Paul and Pushpdeep Singh and Ines Abdelaziz and Sayeh Jarollahi and Seungeon Lee and Krishna P. Gummadi and Ingmar Weber and Abhisek Dash},
  year          = {2026},
  eprint        = {XXXX.XXXXX},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CL},
  url           = {https://arxiv.org/abs/XXXX.XXXXX}
}

Dataset Card Contact

For questions about ModerationBench, please contact psingh[at]mpi-sws.org or ayanm[at]mpi-sws.org.

Downloads last month: 39

Total file size:

2.89 GB