README.md

---
language:
  - en
license: cc-by-4.0
task_categories:
  - text-classification
tags:
  - misinformation
  - llm-evaluation
  - pca
  - web-text
size_categories:
  - 100K<n<1M
---

# Misinformation Detection Dataset

A curated dataset of **100,000 English-language web text snippets** evaluated for misinformation severity using an ensemble of open-weight Large Language Models (LLMs) in an unsupervised framework.

## 📊 Dataset Overview

1. **Source Corpus:** Sourced from [`agentlans/the-dump`](https://huggingface.co/datasets/agentlans/the-dump), representing a diverse collection of English web texts with varying quality and domains.
2. **Preprocessing & Truncation:** Each sample is truncated to a maximum of **2,000 characters**.
3. **LLM Evaluation:** Evaluated for misinformation severity on a scale from **1 to 5** using 10 open-weight LLMs running via `llama.cpp` (`Q4_K_M` quantization).
4. **Dimensionality Reduction (`pc1`):** Individual model ratings were analyzed using Principal Component Analysis (PCA) to extract the primary axis of variation (`pc1`). 
   * *Note:* The `eurollm` column was excluded from the PCA due to weak correlations with other models.
5. **Normalization:** Due to a heavy right-skew, `pc1` values were transformed using the **Yeo-Johnson method** followed by **z-score standardization** (mean = 0, standard deviation = 1).

## 🤖 Evaluated Models

| Column Name | Model Repository / Reference |
| :--- | :--- |
| `aya` | [CohereLabs/aya-expanse-8b](https://huggingface.co/CohereLabs/aya-expanse-8b) |
| `eurollm` *(Excluded from PCA)* | [utter-project/EuroLLM-9B-Instruct](https://huggingface.co/utter-project/EuroLLM-9B-Instruct) |
| `falcon` | [tiiuae/Falcon3-10B-Instruct](https://huggingface.co/tiiuae/Falcon3-10B-Instruct) |
| `gemma` | [google/gemma-4-12B-it](https://huggingface.co/google/gemma-4-12B-it) |
| `granite` | [ibm-granite/granite-4.2-8b](https://huggingface.co/ibm-granite/granite-4.2-8b) |
| `llama` | [meta-llama/Llama-3.1-8B-Instruct](https://huggingface.co/meta-llama/Llama-3.1-8B-Instruct) |
| `mistral` | [mistralai/Ministral-3-14B-Instruct-2512-BF16](https://huggingface.co/mistralai/Ministral-3-14B-Instruct-2512-BF16) |
| `nemotron` | [nvidia/NVIDIA-Nemotron-3-Nano-4B-BF16](https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Nano-4B-BF16) |
| `phi` | [microsoft/phi-4](https://huggingface.co/microsoft/phi-4) |
| `qwen` | [Qwen/Qwen3.5-9B](https://huggingface.co/Qwen/Qwen3.5-9B) |

## 📝 Evaluation Prompt Template

```text
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 likelihood to mislead, not the author's motives.
- When evidence is insufficient to establish a serious error, prefer the less severe score.

Score Scale:
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.

```

## 📋 Dataset Examples

Below are sample rows ordered from least to most misleading content:

| Text Snippet (Truncated) | `pc1` | `score` |
| --- | --- | --- |
| `[URL] Health Care Law Blog Recap from the 2015 Health Law Institute Foster Swift health care attorne...` | -2.6824 | -1.7526 |
| `PET CLOTHING MEASUREMENT TIPS: 1. Measure your pet from the base of the neck to the base of the tail...` | -2.2087 | -1.3709 |
| `Mindfulness is total focus in the present moment, zeroed in on a single activity. Being mindful mean...` | -1.4544 | -0.7973 |
| `nutroll ======= [![Python]([URL])]([URL]) This is a sample Python-language application intended to d...` | 0.0507 | 0.1829 |
| `Baron Umberto Louis De Morpurgo – February 26 ) was a male tennis player from Italy De Morpurgo was...` | 0.6816 | 0.5075 |
| `And of the first I would gladlie knowe, whither they suppose it fitter, that the sacred songs in the...` | 3.3318 | 1.6013 |
| `"Yes, Father, I want to confess, but I am ashamed of what I have done, not even God can forgive me,"...` | 4.0426 | 1.8557 |
| `I got a call from a hedge fund manager in New York this morning. I knew him in California and for th...` | 6.1462 | 2.5543 |
| `Will COVID-19 Vaccines Doom the Population With Premature Deaths? A Special Interview With Dr. Vladi...` | 8.5113 | 3.2708 |
| `Aishwarya Rai Nude on Pool Showing her Sexy Boobs Fake. Incoming search terms: xxx photos indian act...` | 11.2664 | 4.0425 |

## 📈 Correlation Analysis

The heatmap below illustrates the **Spearman rank-order correlation** coefficients among the individual LLM evaluators and the extracted severity metrics:

<img src="Correlation.svg" alt="Spearman Correlation Heatmap" width="100%">

* **Inter-Model Agreement:** Most pairwise model correlations fall in the weak-to-moderate range (r_s = 0.3 - 0.7), highlighting the diversity in model architectures, training data, and alignment strategies.
* **PC1 Alignment:** As expected, the primary principal component (`pc1`) displays strong, consistent correlations with the individual classifier scores, confirming that it successfully captures the shared consensus axis of misinformation severity.

## ⚠️ Limitations & Considerations

1. **Closed-Book Evaluation:** Annotations were generated without Retrieval-Augmented Generation (RAG), external web access, or automated fact-checking tools. Assessments are strictly bounded by each model's parametric knowledge.
2. **Heuristic Screening vs. Ground Truth:** Ratings reflect LLM perception and heuristic scoring rather than definitive, human-verified ground-truth fact checking. Scores should not be treated as absolute determinations of objective truth.
3. **Quality vs. Accuracy:** A low score indicates high *factual accuracy*, not necessarily high content quality. Low-value, repetitive, or nonsensical text (e.g., SEO keyword lists or boilerplate text) may receive a score of `1` provided they contain no false claims.
4. **Context Truncation:** Snippets are capped at 2,000 characters. Cut-off arguments or incomplete contexts may occasionally skew model scoring.
5. **Domain & Temporal Bias:** The source corpus predominantly reflects web discourse typical of the 2010s, introducing specific language patterns, styles, and topic distributions.
6. **Model Heterogeneity:** Inter-annotator agreement varies across candidate LLMs due to architectural differences, alignment strategies, and parameter scales.

## 📜 License

Distributed under the [Creative Commons Attribution 4.0 International (CC BY 4.0)](https://creativecommons.org/licenses/by/4.0/) license.