mdsajjadullah/misinfoshort-22k

---
language:
- en
license: cc-by-4.0
task_categories:
- text-classification
task_ids:
- multi-class-classification
tags:
- misinformation
- fake-news
- nlp
- short-text
- emotion-detection
- social-media
- benchmark
- synthetic
pretty_name: MisInfoShort-22K
size_categories:
- 10K<n<100K
---

# MisInfoShort-22K

## A Short-Text Multi-Domain Misinformation Detection Benchmark

---

## Dataset Summary

MisInfoShort-22K is a research-grade synthetic dataset for training and
benchmarking AI models to detect misinformation in short texts.

It contains **20,000 samples** between **5 and 40 words** each, simulating
the kind of content shared on social media, WhatsApp, forums, and comment
sections — the actual places where misinformation goes viral.

---

## Why This Dataset Exists

Most misinformation datasets use long news articles. But in reality,
misinformation spreads in SHORT messages. This dataset fills that gap.

| Dataset | Short Text | Multi-Domain | Emotion Labels | Satire Class |
|---------|-----------|-------------|---------------|-------------|
| LIAR | ✗ | ✗ | ✗ | ✗ |
| FakeNewsNet | ✗ | Partial | ✗ | ✗ |
| PHEME | Partial | ✗ | ✗ | ✗ |
| **MisInfoShort-22K** | **✓** | **✓** | **✓** | **✓** |

---

## Dataset Statistics

| Property | Value |
|----------|-------|
| Total samples | 20,000 |
| Training samples | 16,000 |
| Validation samples | 2,000 |
| Test samples | 2,000 |
| Domains | 5 |
| Label classes | 4 |
| Misinfo subtypes | 7 |
| Average text length | ~18 words |
| Text length range | 5–40 words |
| Language | English |
| License | CC BY 4.0 |

---

## Supported Tasks

- **Binary classification** — misinformation vs not misinformation
- **4-class classification** — misinformation / factual / opinion / satire
- **7-class classification** — misinfo subtype detection
- **Emotion prediction** — 6-class emotion trigger labeling
- **Multi-label emotion** — dual emotion using emotion_secondary column
- **Domain classification** — 5-class domain identification
- **Multi-task learning** — joint prediction across all metadata

---

## Dataset Structure

### Data Fields

| Column | Type | Description |
|--------|------|-------------|
| id | string | Unique ID (MIS-XXXXXX format) |
| text | string | Short text, 5–40 words |
| label | string | misinformation / factual / opinion / satire |
| misinformation_type | string | conspiracy / medical_claim / propaganda / scam / rumor / myth / none |
| domain | string | health / politics / finance / technology / science |
| semantic_focus | string | cure_claim / election_fraud / investment_promise / technology_fear / scientific_misinterpretation / policy_claim / economic_prediction |
| emotion_trigger | string | fear / anger / urgency / curiosity / hope / neutral |
| emotion_secondary | string | Second emotion for multi-label modeling (or none) |
| source_type | string | social_media / blog / messaging_app / forum / comment |
| text_length | integer | Word count (5 to 40) |

### Label Definitions

| Label | Meaning |
|-------|---------|
| misinformation | False or misleading claims presented as factual |
| factual | Accurate, evidence-based statements |
| opinion | Subjective viewpoints and personal perspectives |
| satire | Humorous exaggeration not intended as literal truth |

### Data Splits

| Split | Rows | Percentage |
|-------|------|-----------|
| train | 16,000 | 80% |
| validation | 2,000 | 10% |
| test | 2,000 | 10% |

---

## How to Load This Dataset

### Option 1 — With HuggingFace datasets library
```python
from datasets import load_dataset

dataset = load_dataset("YOUR_USERNAME/misinfoshort-22k")

print(dataset)
# DatasetDict({
#     train: Dataset({features: [...], num_rows: 16000}),
#     validation: Dataset({features: [...], num_rows: 2000}),
#     test: Dataset({features: [...], num_rows: 2000})
# })

# Access training data
train_df = dataset["train"].to_pandas()
print(train_df["label"].value_counts())

Option 2 — With pandas directly

import pandas as pd

train = pd.read_csv("train.csv")
val   = pd.read_csv("validation.csv")
test  = pd.read_csv("test.csv")

print(train.shape)
print(train["label"].value_counts())
print(train.head())

Option 3 — Quick label check

from datasets import load_dataset

ds = load_dataset("YOUR_USERNAME/misinfoshort-22k")

# See label distribution
import pandas as pd
df = ds["train"].to_pandas()
print(df["label"].value_counts())
print(df["domain"].value_counts())
print(df["emotion_trigger"].value_counts())

---

Sample Data

text	label	domain	emotion_trigger
URGENT: turmeric cures cancer naturally — share before the WHO removes this post!	misinformation	health	urgency
According to the latest clinical research, regular aerobic exercise significantly reduces the risk of heart disease.	factual	health	neutral
In my view, the healthcare system is failing millions of patients living with depression.	opinion	health	neutral
Doctors baffled as woman cures everything by drinking enough water — Big Pharma refuses to fund follow-up study.	satire	health	curiosity

---

Recommended Models for Fine-tuning

# Example with bert-base-uncased
from transformers import AutoTokenizer, AutoModelForSequenceClassification

model_name = "bert-base-uncased"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSequenceClassification.from_pretrained(
    model_name,
    num_labels=4  # for 4-class classification
)

Recommended models:

• bert-base-uncased — strong baseline
• roberta-base — best general performance
• distilbert-base-uncased — faster and lighter
• microsoft/deberta-v3-base — state of the art

---

Dataset Creation

This dataset was synthetically generated using:

• Slot-based template generation across 20 domain/label combinations
• Large vocabulary banks (40+ items per slot) for combinatorial diversity
• Multi-stage deduplication (exact + near-duplicate filtering)
• Stratified 80/10/10 splitting with random_state=42

Diversity ratio: 85.4% unique texts

---

License

CC BY 4.0 — Free to use for any purpose with attribution.

---

Citation

@dataset{misinfoshort22k_2024,
  title   = {MisInfoShort-22K: A Short-Text Multi-Domain Misinformation
             Detection Benchmark with Emotional and Semantic Metadata},
  year    = {2024},
  version = {1.0.0},
  license = {CC BY 4.0},
  url     = {https://huggingface.co/datasets/mdsajjadullah/misinfoshort-22k}
}