These guidelines quote hateful and offensive posts from the Measuring Hate Speech corpus verbatim and are released for research use. Access is approved manually. Please fill in the form and accept each term of use.

🧭📜 Induced Hate Speech Classification Guidelines

⚠️ Content warning: the guidelines quote hateful and offensive posts verbatim.

🚦 Access Conditions

This dataset is gated and access is approved manually. To request access, log in, fill in the form on this page and accept each term of use. Your request will be reviewed, and you will be notified by email.

📝 Dataset Summary

Induced Hate Speech Classification Guidelines contains 165 hate speech classification guidelines written by an LLM (the guideline writer) from annotated posts of the Measuring Hate Speech corpus (MHS). LLM classifiers use a guideline as their annotation policy, in place of or alongside a generic hate speech definition.

Each guideline is induced from 80 annotated posts (the seed, round 0) and refined over 10 rounds of 80 new posts from the same pool. Guidelines come from 15 induction runs that vary two factors:

Induction pool: the posts shown to the writer, either posts on which annotators disagree, posts on which they are unanimous, or random posts.
Annotation context: what the writer sees with each post, from the aggregate label alone (Label) to every annotator's vote (Votes), the target groups they identified (Targets) and their self-reported demographics (Demographics).

Every criterion in a guideline is grounded in a verbatim quote from a post the writer was shown. The guidelines are released with every refinement round and the MHS comment_ids of the posts behind each round.

Language: English
Paper: Learning Hate Speech Classification Guidelines from Annotator Disagreement (under review)
Source corpus: Measuring Hate Speech (Reddit, Twitter and YouTube comments)

🎯 Uses

Direct Use

Prompting LLMs to classify hate speech with an explicit, example-grounded annotation policy.
Comparing guidelines induced from disagreement, unanimous and random posts, or from different annotation contexts.
Studying how a guideline changes over refinement rounds.

Out-of-Scope Use

Automated content moderation without human oversight.
Treating a guideline as a universal or value-neutral definition of hate speech: the criteria reflect the MHS annotations and labelling scheme.
Treating the demographic explanations of disagreement as evidence of how a group judges hate speech.

📦 Dataset Structure

guidelines.jsonl has one row per guideline (15 runs × 11 rounds):

Field	Description
`pool`	Induction pool: `disagreement`, `unanimous` or `random`
`platform`	`all`, or `reddit`, `twitter`, `youtube` for the single-platform runs (disagreement pool, Targets)
`condition`	Annotation context: `label`, `votes`, `targets` or `demographics`
`setting`	Name of the run in the code
`round`	Refinement round (0 = seed, 10 = last)
`accepted`	Whether the revision of this round passed the quality checks; if not, the previous guideline is kept
`posts_seen`	Posts shown to the writer up to this round
`new_post_ids`	MHS `comment_id`s of the 80 posts shown in this round
`guideline`	The guideline the classifiers receive after this round

Induction pool	Posts	Hate speech	Disagreement posts
`disagreement` (fewer than 75% of the annotators give the same answer)	4,077	39.1%	100%
`unanimous` (all annotators give the same answer)	5,904	19.1%	0%
`random` (all posts of the MHS induction split)	10,764	27.9%	37.9%

Condition	What the writer sees with each post
`label`	The binary label (hate if the MHS hate speech score exceeds 0.5)
`votes`	Every annotator's answer: Hate Speech, Unclear or Not Hate Speech
`targets`	Votes and the target groups identified by the annotators (race, religion, national origin, gender, sexuality, age, disability)
`demographics`	Votes, targets and each annotator's self-reported gender, age, race, religion, sexuality, education, income and political ideology

Every guideline has the sections CORE QUESTION, HATE SPEECH IF, NOT HATE SPEECH IF, KEY VOCABULARY / CUES and TRICKY CASES. Guidelines induced from votes add ANNOTATOR DISAGREEMENT PATTERNS, and Demographics adds GROUP-DEPENDENT CASES. Guidelines induced from unanimous posts have no disagreement sections.

The same guidelines are also available as plain text, in <pool>/<condition>/round_XX.txt with a rounds.json per run; the single-platform runs are in platform-<platform>/targets/.

Data Instances

{
  "pool": "disagreement",
  "platform": "all",
  "condition": "targets",
  "setting": "votes_targets",
  "round": 10,
  "accepted": false,
  "posts_seen": 880,
  "new_post_ids": [17796, 3894, 46653, "..."],
  "guideline": "CORE QUESTION: Does the post contain language that is derogatory, discriminatory, or violent towards an individual or group based on attributes such as race, religion, ethnicity, or gender?\n\nHATE SPEECH IF the post exhibits any of these distinct patterns:\n- Targets individuals or groups with violent or discriminatory language, including homophobic or transphobic language, [...] -- quote: \"...\"\n..."
}

📊 Dataset Statistics


Induction runs	15 (3 pools × 4 conditions, plus 3 single-platform Targets runs)
Guidelines	165 (seed and 10 refinement rounds per run)
Posts shown per run	880 (80 per round)
Accepted revisions	146 of 150
Guideline length in words (median, round 0 / round 10)	893 / 1,317

🚀 Usage

from datasets import load_dataset

# Gated dataset: log in first with `hf auth login`, or pass token=... to load_dataset
guidelines = load_dataset("<ANON>/hate-induction-guidelines", split="train")
guideline = guidelines.filter(
    lambda r: r["pool"] == "disagreement" and r["platform"] == "all"
    and r["condition"] == "targets" and r["round"] == 10
)[0]["guideline"]

To classify a post, use the classification prompts of the paper:

SYSTEM = (
    "You are a careful hate-speech annotation assistant. Apply the provided classification parameters "
    "consistently. Base your decision only on the post and the provided context. Do not infer unavailable "
    "context, speaker identity, or intent. Follow the requested output format exactly and do not add any "
    "extra text before or after it."
)
USER = """Hate speech is defined as language characterised by offensive, derogatory, humiliating, or insulting discourse that promotes violence, discrimination, or hostility towards individuals or groups based on attributes such as race, religion, ethnicity, or gender.

Use the following classification guideline, which was induced from human annotation examples, as the annotation policy for this task.

INDUCED HATE SPEECH CLASSIFICATION GUIDELINE:
{guideline}

Is the following post hate speech?

Post: "{post}"

Respond with a single JSON object of the form {{"answer": "True|False"}}. The "answer" field must be exactly one of "True" or "False"."""

messages = [
    {"role": "system", "content": SYSTEM},
    {"role": "user", "content": USER.format(guideline=guideline, post="<text of the post>")},
]

The paper used Llama 3.1 8B, Qwen3 14B, Mistral Nemo 12B, Phi-4-mini and OLMo 3.1 32B Instruct as classifiers, with guided JSON decoding, temperature 0.2, top-p 0.9 and top-k 20.

🛠️ Dataset Creation

Curation Rationale

LLMs prompted with a generic definition of hate speech handle clear cases but get little guidance on borderline posts, on which human annotators often disagree. The guidelines make the criteria for these decisions explicit: they are induced from annotated posts, including contested ones, to test whether annotator disagreement is a useful source of classification criteria. In the paper, Targets guidelines induced from disagreement posts improve over zero-shot prompting for all five classifiers, transfer to HateXplain and across MHS platforms, and use 41–43% fewer prompt tokens than 80 demonstrations.

Source Data

MHS comments with at least three annotators were split into an induction split (10,764 comments) and a held-out split. All posts shown to the writer come from the induction split; none of the held-out posts was used.

Data Collection and Processing

Guideline writer: Llama 3.3 70B Instruct (4-bit), temperature 0.2 and top-p 0.9, run with vLLM.
Seed (round 0): the writer receives a hate speech definition and 80 posts with the annotation context of the condition, and must ground every criterion in a verbatim quote. All conditions of a pool see the same posts in the same order.
Refinement (rounds 1–10): each round shows the current guideline and 80 unseen posts from the same pool. The writer keeps criteria not contradicted by the new posts, adds criteria for new patterns and narrows or moves contradicted criteria to the tricky cases.
Quality checks: a revision is rejected if it is truncated, misses a section, repeats lines, grows more than 1.5× or has more than 20 hate or 12 not-hate criteria (revisions above 15 and 10 criteria are first condensed). Every quote must be found in a post the writer was shown. A rejected revision is regenerated at higher temperatures; if every attempt fails, the round keeps the previous guideline (accepted: false).

Who are the source data producers?

Users of Reddit, Twitter and YouTube whose comments were collected in MHS.

Annotations

Annotation process

The labels, votes, target groups and demographics shown to the writer are the MHS crowdsourced annotations. The guidelines are machine-generated and were not edited by hand.

Who are the annotators?

MHS annotators (the information shown to the writer) and an LLM (the guidelines).

🔒 Personal and Sensitive Information

The guidelines quote public social media posts verbatim, which can include user handles and names of public figures. Demographics guidelines describe vote splits by annotator characteristics (for example, "Hate Speech: 1 woman, liberal; Not Hate Speech: 2 men, one conservative"); no annotator IDs are included. Post IDs refer to the public MHS corpus.

⚠️ Bias, Risks, and Limitations

Scope: English posts from three platforms and a binary label that separates hate from offensive speech. The criteria may not transfer to other languages, cultural contexts or annotation schemes.
Annotation policy: the guidelines encode the MHS annotations and labelling choices, not a universal definition of hate speech. Disagreement may reflect legitimate differences in interpretation rather than errors.
Pools: the pools differ in their hate rate (39.1%, 19.1% and 27.9%), so differences between them may partly reflect class composition. Each run uses a single sample of posts.
Generated text: every quote was checked against the posts shown, but the criteria themselves were not validated by humans and may be overly broad, inconsistent or redundant.
Demographic explanations: these are the writer's hypotheses about why annotators disagreed and are sometimes inconsistent with the vote split they describe. Demographics guidelines did not improve classification in the paper.

Recommendations

Use the guidelines as annotation policies for research, read them in relation to the MHS labelling scheme, and keep human oversight for any consequential decision.

📄 License

The guidelines are released under CC BY 4.0, the license of the Measuring Hate Speech corpus. They are derived from MHS: every criterion quotes an MHS post verbatim, the Votes, Targets and Demographics guidelines summarize MHS annotations, and new_post_ids lists MHS comment_ids. Releasing them under the same license keeps the original terms and attribution requirements. Any use of this dataset must also credit MHS (see below).

The access form adds terms of responsible use because the guidelines contain hateful and offensive text.

📝 Citation

The citation will be added upon publication. Until then, please refer to the paper Learning Hate Speech Classification Guidelines from Annotator Disagreement (under review).

Please also cite the Measuring Hate Speech corpus:

@inproceedings{sachdeva-etal-2022-measuring,
    title = "The Measuring Hate Speech Corpus: Leveraging Rasch Measurement Theory for Data Perspectivism",
    author = "Sachdeva, Pratik S. and Barreto, Renata and Bacon, Geoff and Sahn, Alexander and von Vacano, Claudia and Kennedy, Chris",
    booktitle = "Proceedings of the 1st Workshop on Perspectivist Approaches to NLP @LREC2022",
    year = "2022",
    address = "Marseille, France",
    publisher = "European Language Resources Association",
    url = "https://aclanthology.org/2022.nlperspectives-1.11/",
    pages = "83--94"
}

@misc{kennedy2026measuringhatespeechspectrum,
    title = {Measuring a hate speech spectrum with faceted Rasch item response theory and perspective-aware, explainable-by-design deep learning},
    author = {Chris J. Kennedy and Geoff Bacon and Alexander Sahn and Claudia von Vacano},
    year = {2026},
    eprint = {2009.10277},
    archivePrefix = {arXiv},
    primaryClass = {cs.CL},
    url = {https://arxiv.org/abs/2009.10277}
}

Downloads last month: 34

Total file size:

2.7 MB

Paper for irlab-udc/hate-induction-guidelines

Constructing interval variables via faceted Rasch measurement and multitask deep learning: a hate speech application

Paper • 2009.10277 • Published Sep 22, 2020