README.md

metadata

license: cc-by-4.0
task_categories:
  - text-classification
language:
  - en
tags:
  - misinformation
  - LLM-annotations
  - synthetic-evaluation
size_categories:
  - 100K<n<1M

Misinformation Detection Dataset

Overview

This dataset contains 100,000 text snippets sourced from agentlans/the-dump, a collection of diverse English web texts with varying quality.

Each sample is truncated to a maximum of 2,000 characters and evaluated for misinformation severity by four open-weight Large Language Models (LLMs) on a scale from 1 to 5:

• google/gemma-4-12B-it
• mistralai/Ministral-3-14B-Instruct-2512
• microsoft/phi-4
• Qwen/Qwen3.5-9B

Dataset Structure

Data Fields

• id (string): Unique identifier for the snippet. Same ID as the original text in the agentlans/the-dump dataset.
• text (string): Source text snippet (truncated to <= 2000 characters).
• gemma (integer): Misinformation score (1–5) assigned by google/gemma-4-12B-it.
• mistral (integer): Misinformation score (1–5) assigned by mistralai/Ministral-3-14B-Instruct-2512.
• phi (integer): Misinformation score (1–5) assigned by microsoft/phi-4.
• qwen (integer): Misinformation score (1–5) assigned by Qwen/Qwen3.5-9B.

Example Row

{
  "id": "95a646ed-2f12-40f8-9ca3-628179cd6b66",
  "text": "The aim of the present study was to determine the effect of 2-hydroxy 4-methoxy benzoic acid (HMBA)...",
  "gemma": 1,
  "mistral": 2,
  "phi": 2,
  "qwen": 1
}

Annotation Methodology

The annotating LLMs evaluated each text prompt closed-book using the following structured rubric:

Prompt Template

Evaluate the text below for misinformation.

Judge factual accuracy, misleading framing, material omissions, and whether the claims are supported by the available evidence.

Rules:
- Evaluate factual claims, not political viewpoints, opinions, preferences, or value judgments.
- Do not treat a claim as false merely because it is unverified, unfamiliar, or unsupported by evidence in the text.
- Distinguish false claims from uncertain claims, reasonable interpretations, predictions, and disputed claims.
- Penalize omissions only when missing information materially changes the meaning or creates a misleading impression. Do not expect a short text to cover every aspect of a topic.
- Consider the text as a whole. Weigh the number, importance, and certainty of factual errors rather than counting claims.
- Do not infer deceptive intent. Judge the text's accuracy and likely to mislead, not the author's motives.
When evidence is insufficient to establish a serious error, prefer the less severe score.

Score: 1 = Accurate. No meaningful factual errors or misleading impressions. 2 = Minor inaccuracies or imprecisions that do not materially affect the overall meaning. 3 = Significant factual errors or material omissions that could mislead readers. 4 = Major false or misleading claims that substantially distort the topic. 5 = Predominantly false or severely deceptive in its factual content, with central claims demonstrably false or seriously misleading.

Choose the single score that best represents the overall severity. Do not equate controversy, lack of citations, or disagreement with misinformation.

Text:
[TEXT]

Output exactly one digit: 1, 2, 3, 4, or 5. No other text.

Limitations & Considerations

1. Closed-Book Evaluation: Annotations were generated without retrieval-augmented generation (RAG), web access, or active fact-checking tool integration. Assessments are bounded entirely by model parametric knowledge.
2. Detection vs. Ground Truth: Ratings reflect LLM perception/heuristic screening rather than definitive, ground-truth fact verification. Scores should not be treated as authoritative determinations of truth.
3. Quality vs. Misinformation: A low score indicates high factual accuracy, not high content quality. Nonsensical, highly repetitive, or low-value text (e.g., SEO keyword lists or shopping lists) may be scored as 1 as long as no false claims are made.
4. Context Truncation: Snippets were truncated to 2000 characters. Incomplete context or cut-off arguments may skew model scoring.
5. Coverage & Domain Bias: The source corpus predominantly reflects web discourse typical of the 2010s, introducing domain-specific language and topic distributions.
6. Model Heterogeneity: Inter-annotator agreement varies across candidate LLMs due to differences in architecture, alignment, and parameter size.

Work in Progress

This dataset is an ongoing release. Future updates plan to include:

• Extended annotations across additional model families and parameter sizes.
• Aggregated consensus scores (e.g., mean/median/ensemble ratings across annotators).

Licence

Distributed under the Creative Commons Attribution 4.0 International (CC BY 4.0) license.