WitchesSocialStream/Four-Leaf-Clover

"It’s 4chan, what did you expect, rainbows?"
Use with care. This card is deliberately blunt.

---

Dataset Details

Dataset Description

Four Leaf Clover is an unfiltered, chronologically-ordered crawl of public posts (with attachments) scraped from every live 4chan board as of September 2025. Each record contains the exact text the author submitted plus any image or media file linked at time of archival. Nothing was redacted, re-written, or rate-limited. If you train on this you will model an internet dumpster fire.
Proceed only if you explicitly want to study, measure, or mitigate the language norms, hate speech, memes, misinformation, and toxicity endemic to anonymous image boards.

• Curated by: karaKarawitch
• Language(s): English (predominantly), occasional code-switching
• License: Apache-2.0 (dataset itself). Individual posts inherit original, unspecified user copyright; redistribution follows fair-use doctrine.
• Dataset Version: v1.2.0 (attachments enabled)

Dataset Sources

• Repository: https://huggingface.co/datasets/WitchesSocialStream/Four-Leaf-Clover
• Paper: none yet; cite this dataset card instead.

---

Uses

Direct Use - Research & Community Value

1. Toxicity & Abuse Detection: Large, up-to-date source for hard-to-find slurs, grooming language, or other harmful patterns.
2. Dialect & Meme Evolution: Study how 4chan memes migrate to mainstream platforms (e.g., "kek" -> "lulz").
3. Counter-misinformation Datasets: Negative examples to augment moderation models trained on safer corpora.
4. Anonymity Impact on Discourse: Linguistic analysis of how speaker identity absence changes conversation structure.

Out-of-Scope / Forbidden Uses

• Training open-ended generative chatbots — the dataset itself does not prohibit this use, but doing so is entirely your own responsibility. If your model melts down, gets banned, or brings legal heat, that’s on you; we do not accept bug reports, ethical hand-wringing, or subpoenas downstream.
• Harvesting PII: Email, Discord, or social-media handles were not scrubbed.
• Commercial red-light content generation (incitement, self-harm instructions).
• Fine-tuning any model for mass-scraping or automated posting back to 4chan. (Violates 4chan ToS; also, you’ll be roasted harder than a Thanksgiving turkey.)

---

Dataset Structure

One JSONL per board (or daily shard for >100 k posts). Each line = one post.

from pydantic import BaseModel, Field
from typing import Optional

class AttachmentData(BaseModel):
    attachment_url: str   # CDN link valid as of crawl epoch
    filename: str         # as uploaded

class Post(BaseModel):
    board: str            # board shortcode, e.g. "/b/", "/pol/"
    thread: int           # thread id
    pid: int              # post number
    name: str             # display name, may be "Anonymous" or tripcodes
    msg: str              # raw text (HTML tags removed; unicode left intact)
    attachment: Optional[AttachmentData] = None
    posted: int           # UTC epoch seconds

Total records: ~22 TB attachments & posts (full size, all forms of attachments are captured). Splits: None. Chronological streaming keeps temporal relations intact. Filters: None

---

Dataset Creation

Curation Rationale

1. Coverage Gap: Existing 4chan corpora are either stale, board-limited, or stripped of attachments.
2. Research Utility: Anonymous speech remains understudied compared to Reddit / Twitter datasets.

Source Data

Data Collection

• Period: November 2024 → ongoing hourly uploads
• Tools:
  • We keep implementation details private to deter evasion of archival bans.
• Respects robots.txt: Export cycle: fresh crawl every 5 seconds; hourly bulk JSONL upload via a minimal Python script.

Producers (Anonymity Preserved)

Every contributor is pseudonymous ("Anonymous" or tripcodes). We do not store or link IP hashes. Any demographic or identity data is user-supplied (and unreliable).

---

Personal & Sensitive Information

• Yes - handles, timestamps, and cross-linked usernames may contain PII.
• Yes - racial slurs, sexual orientation, self-harm content, medical/psychological claims, political affiliation.
• Anonymisation: Zero filtering, scrubbing, or sanitisation. Everything stays exactly as seen at scrape time.

---

Bias, Risks, and Limitations

Aspect	Reality Check
Hate Speech	4chan normalises slurs at scale. Expect >20 % of content to violate standard content-policies.
Gender Bias	Heavily skewed to male-presenting pronouns; misogynist tropes are prevalent in many boards.
Legal Risk (CSAM links)	4chan’s moderation removes illicit attachments rapidly; nevertheless zero-day uploads may still slip through. No filtering or hashing is performed. Storage, re-hosting, or redistribution of attachments is strictly at your own legal peril.
Temporal Drift	The culture of 4chan shifts rapidly; anything older than 6 months may no longer reflect current norms.
Consent	Publicly posted, but users did not consent to ML ingestion. Treat as ethically grey-zone.

---

Recommendations for Responsible Use

1. Filter & Balance - pair with safer datasets or generate synthetic positives/negatives to reduce extremism amplification.
2. Hate-check - run every model on e.g. HolisticBias before release; log evaluations publicly.
3. Fine-print - disclose 4chan origin prominently whenever the dataset is mentioned downstream.
4. Block redeployment back to source - do not build bots that post on 4chan; mods will nuke you and the community will dox you for fun.
5. Red-team - have external ethics reviewers audit training runs on sensitive topics.

Boards indexed

We index all active boards at the moment. Boards indexed are as follows:

/a/, /b/, /c/, /d/, /e/, /f/, /g/, /gif/, /h/, /hr/, /k/, /m/, /o/, /p/, /r/, /s/, /t/, /u/, /v/, /vg/,
/vm/, /vmg/, /vr/, /vrpg/, /vst/, /w/, /wg/, /i/, /ic/, /r9k/, /s4s/, /vip/, /qa/, /cm/, /hm/, /lgbt/,
/y/, /3/, /aco/, /adv/, /an/, /bant/, /biz/, /cgl/, /ck/, /co/, /diy/, /fa/, /fit/, /gd/, /hc/, /his/,
/int/, /jp/, /lit/, /mlp/, /mu/, /n/, /news/, /out/, /po/, /pol/, /pw/, /qst/, /sci/, /soc/, /sp/,
/tg/, /toy/, /trv/, /tv/, /vp/, /vt/, /wsg/, /wsr/, /x/, /xs/

---

Citation

@misc{fourleafclover2025,
  title={Four Leaf Clover: Unfiltered 4chan Corpus},
  author={karaKarawitch},
  howpublished={Hugging Face dataset},
  year={2025},
  url={https://huggingface.co/datasets/WitchesSocialStream/Four-Leaf-Clover}
}

APA: karaKarawitch (2025). Four Leaf Clover: Unfiltered 4chan Corpus [Dataset]. Hugging Face. https://huggingface.co/datasets/WitchesSocialStream/Four-Leaf-Clover

---

Dataset Card Contact

• Everything: if you already know where else to find karaKarawitch, you know what to do. We no longer list open channels to prevent solicitation.

---