Datasets:

RuleFollower
/

rulefollower_parsed

Text content: Replies to source tweets about rumored events
Source: Gorrell et al (2019)
Dataset: Huggingface
Annotation goal: Classify reply stance as support, deny, query, or comment
Ground truth: Provided (stance label)
Note: Only reply texts are retained for classification.

Text content: Social media posts from 8 hate/offensive speech datasets
Source: Nghiem and Daumé III (2024)
Dataset: Github
Annotation goal: Classify post as benign, offensive, or hateful
Ground truth: Provided
Note: The original HateCoT dataset combines samples from 8 hate speech datasets with diverse annotation schemes. We standardized its many fine-grained labels into 3 unified classes:
- 0 = Benign (e.g., "Not Hate", "Normal", "Neutral"),
- 1 = Offensive (e.g., "Toxic", "Offensive"),
- 2 = Hateful (e.g., "Hate", "Dehumanization", "Directed Abuse").

Text content: Tweets from different public Twitter samples spanning 2017 to 2023, focused on content moderation and related debates. Sample sizes vary across years and sources.
Source: Gilardi et al (2023)
Dataset: Harvard Dataverse
Ground truth: Not provided
Supported tasks (from annotation codebook, excluding political content tasks):

Task ID	Task Name	Task Description Summary	Labels
T1	Content Moderation Relevance	Is the tweet about content moderation?	`relevant` (1), `irrelevant` (0)
T3	Problem/Solution Frame	Does the tweet portray content moderation as a problem, a solution, or neither?	`problem`, `solution`, `neutral`
T4	Policy Frame (Moderation)	What policy dimension frames the content moderation issue?	e.g., `morality`, `fairness`, `security`, `equality`, `health`, ...
T6	Stance on Section 230	Does the tweet support, oppose, or remain neutral about Section 230 of U.S. law?	`positive`, `negative`, `neutral`
T7	Topic Classification	What is the topic in relation to content moderation?	`section 230`, `trump ban`, `complaints`, `platform policies`, ...

Note:
- Each of these tasks has a dedicated .txt task description under prompt/task_descriptions/.
- The same dataset can be reused across tasks; task choice is controlled by the selected prompt.
- Tweets by U.S. Congress members (used for political content) are handled separately (see below).

Text content:News headlines (The Misinfo Reaction Frames corpus is an dataset of 25k news headlines to articles that have been factchecked. The articles can be about Covid-19, Cancer or Climate Change)
Source: Gabriel et al. (2022)
Dataset: Github
Annotation goal: Classify headline as misinformation or not
Ground truth: Provided

Text content:A dataset of English-language tweets annotated to capture both explicit and implicit hate speech. Implicit hate includes stereotypical, sarcastic, or coded language that may not be overtly hateful. The data was originally collected from Twitter and the Social Bias Inference Corpus.
Source: Elsafoury et al. (2021)
Dataset: Github
Annotation goal: Classify post as explicit_hate, implicit_hate, or not_hate
Ground truth: Provided
We use Stage 1 annotations for parsing:
- explicit_hate: overt hate speech
- implicit_hate: indirect hate speech (e.g., stereotypes, sarcasm)
- not_hate: no hateful content
Note:
- The dataset originally included multiple annotation stages (e.g., fine-grained subtypes, implied statements).
- Additional parsing for Stage 2 (fine-grained) or Stage 3 (target/implied meaning) will be added later.

Text content: News spans (Opinion spans extracted from global warming news articles, published from Jan. 1, 2000 to April 12, 2020 by various U.S. news sources.)
Source: Luo et al. (2020)
Dataset: Github
Annotation goal: Classify stance toward the statement: “Climate change is a serious concern”
Labels: agree, neutral, or disagree
Ground truth: Provided

Text content: Crisis communication texts (bureaucratic cables and memos from international crises)
Source: Schub (2022)
Dataset: Data
Annotation goal:
- Info Type Task: Classify whether the text conveys political or military information relevant to Cold War crisis decision-making.
- Certainty Task: Determine whether the adviser expresses the information with certainty or uncertainty.
Ground truth: Available only by "Info Type Task"

Text content: Yelp restaurant reviews
Dataset: Data
Annotation goal:
- Sentiment Task: Classify the overall evaluation expressed in a Yelp restaurant review.
Ground truth: Provided

Text content: News articles (A set of articles from media sources covering the United States of America. Data was scraped from https://www.allsides.com/.)
Source: Baly et al. (2020)
Dataset: Data
Annotation goal: Classify the political bias expressed in a news article based on its content.Labels: 0(left), 1(center), 2(right)
Ground truth: Provided

Text content: Requests from Wikipedia editor’s on user talk pages.
Source: Danescu-Niculescu-Mizil et al. (2013)
Dataset: Data
Annotation goal: Classify the level of politeness in the text. Labels: 1(polite), 0(neutral), -1(impolite).
Ground truth: Provided

Text content: Requests from the StackExchange question-answering communities. These are often users commenting on existing posts requesting further information or proposing edits.
Source: Danescu-Niculescu-Mizil et al. (2013)
Dataset: Data
Annotation goal: Classify the level of politeness of the text. Labels: 1(polite), 0(neutral), -1(impolite).
Ground truth: Provided

Text content: Short reports on individual killings committed by the Syrian regime and pro-government forces between June 2013 and April 2015
Source: Gohdes replication data
Dataset: Huggingface
Annotation goal: Classify each killing report as targeted or untargeted
Ground truth: Provided (targeted/untargeted label)
Note: Original two-column labels cat_targeted and cat_untargeted are converted to a single mutually exclusive label.

Text content: Short social media texts that link to or discuss scientific papers
Source: Peng et al. (2022)
Dataset: Huggingface
Annotation goal: Classify whether the text is critical or not critical of the linked scientific paper
Ground truth: Provided (CRITICAL label)
Note: The original CRITICAL binary label is retained as the classification target.

Text content: News stories from Pakistan describing political violence and related events
Source: Halterman and Keith codebook/data package
Dataset: Huggingface
Annotation goal: Classify each story into one event type, such as ASSASSINATION, TERRORISM, RIOT, DRONE_ASSASSINATION, or VIOLENT_POLITICAL_DEMONSTRATION
Ground truth: Provided (single multiclass event label)
Note: A stratified 2,000-example subset is retained from the original train/dev/test files.

Text content: Sentences from political party manifestos
Source: Müller replication data
Dataset: Huggingface
Annotation goal: Classify each sentence’s temporal orientation as future, past, or present
Ground truth: Provided (temporal orientation label)
Note: Original one-hot labels FUTURE, PAST, and PRESENT are converted to a single label: 0 = future, 1 = past, 2 = present.