README.md

metadata

license: cc-by-nc-4.0
viewer: false
pretty_name: bskymoddata — Bluesky Moderation Data
tags:
  - content-moderation
  - bluesky
  - social-media
  - harm-detection
  - graphic-media
  - intolerant
configs:
  - config_name: graphic_media
    data_files:
      - split: train
        path: data/graphic_media/records.jsonl
  - config_name: intolerant
    data_files:
      - split: train
        path: data/intolerant/records.jsonl
extra_gated_prompt: >-
  This dataset contains posts flagged for graphic and intolerant content. By
  accessing it, you agree to use it solely for non-commercial research purposes
  and to not conduct experiments that cause harm to human subjects.
extra_gated_fields:
  Name: text
  Affiliation: text
  I want to use this dataset for:
    type: select
    options:
      - Research
      - Education
      - label: Other
        value: other
  I agree to use this dataset for non-commercial use ONLY: checkbox
  I agree not to conduct experiments that cause harm to human subjects: checkbox

bskymoddata — Bluesky Labelled Posts

A sample of Bluesky posts carrying platform-applied content labels, with associated media files.

⚠️ Content Warning: This dataset contains posts flagged for graphic media (gore, violence, disturbing imagery) and intolerant (hate speech, discrimination) content. Post text, images, and videos may be graphic or offensive. Strictly for non-commercial research. Do not use to target, harass, or re-identify individuals.

---

Splits

Split	Label	Posts	With images	With video
graphic_media	graphic-media	20,000	15,137	4,863
intolerant	intolerant	20,000	2,597	102

Sampled with random.seed(42) from 93,313 graphic-media and 35,899 intolerant valid posts.

---

Data Structure

data/
  graphic_media/records.jsonl
  intolerant/records.jsonl
  media/
    graphic_media/
      images/<xx>/    ← JPEG; <xx> = last 2 chars of CID
      videos/<xx>/    ← MP4
    intolerant/
      images/<xx>/
      videos/<xx>/

---

Schema

Each line in records.jsonl is a JSON object:

Field	Type	Description
cid	string	AT-Protocol content ID
uri	string	AT-Protocol URI (at://did:.../app.bsky.feed.post/...)
label	string	graphic-media or intolerant
neg	string	Negation flag; empty string when label is asserted
cts	string	Label timestamp (when was post labeled)
record	object	Full AT-Protocol post record (see below)
etype	string	Embed type: images, video, external, or ""
text	string	Post text
images	array	Image objects; empty array if none
video	string|null	HF path to video, or null
skip	bool	Always false (pre-filtered)
failures	array	Always [] (pre-filtered)

record fields

Field	Description
record.text	Post text
record.createdAt	Post creation timestamp (ISO 8601)
record.langs	BCP-47 language list (may be absent)
record.embed	Embed object — images / video / external link (may be absent)
record.facets	Rich-text facets — links, mentions, tags (may be absent)
record.reply	Present when post is a reply; absent otherwise (~22% graphic-media, ~76% intolerant)

images element

Field	Description
file	data/media/{label}/images/{xx}/{filename}
alt	Alt-text (may be empty)

---

Downloading

from huggingface_hub import snapshot_download, hf_hub_download

Goal	allow_patterns
Everything	(omit)
Records only	["data/*/records.jsonl"]
One label	["data/graphic_media/**", "data/media/graphic_media/**"]
Images only	["data/*/records.jsonl", "data/media/*/images/**"]

# Example: records only
snapshot_download(
    repo_id        = "usermodsky/bskymoddata",
    repo_type      = "dataset",
    local_dir      = "./bskymoddata",
    allow_patterns = ["data/*/records.jsonl"],   # omit for full download
)

⚠️ Full download includes all graphic media. Estimated size: >50 GB.

Single file:

local_path = hf_hub_download(
    repo_id   = "usermodsky/bskymoddata",
    repo_type = "dataset",
    filename  = "data/media/graphic_media/images/hy/bafkrei...hy.jpeg",
)
# ⚠️ Open only in a controlled review environment — content may be graphic.
print(f"Saved to: {local_path}")

---

Usage

from datasets import load_dataset

ds = load_dataset("usermodsky/bskymoddata", "graphic_media", split="train")
# or: "intolerant"

import pandas as pd
df = pd.read_json("hf://datasets/usermodsky/bskymoddata/data/graphic_media/records.jsonl", lines=True)

Filter subsets:

# Posts with images
with_images = ds.filter(lambda x: len(x["images"]) > 0)

# Text-only posts
text_only = ds.filter(lambda x: len(x["images"]) == 0 and x["video"] is None)

# Reply posts only
replies = ds.filter(lambda x: x["record"].get("reply") is not None)

# English posts
en = ds.filter(lambda x: "en" in (x["record"].get("langs") or []))

---

License

CC BY-NC 4.0 — non-commercial research use only.