---
language: [en]
tags:
  - misinformation
  - twitter
  - unlearning
pretty_name: Twitter Misinformation Unlearning Eval
license: other
source_datasets:
  - roupenminassian/twitter-misinformation
---

# Twitter Misinformation Unlearning Eval

This dataset is derived from `roupenminassian/twitter-misinformation` and is re-organized
for **machine unlearning** experiments.

## Source

- Original dataset: `roupenminassian/twitter-misinformation`
- Fields:
- `text` (string): tweet text.
- `label` (int): `0` for factual, `1` for misinformation.

## Splits

We construct three splits from the **test** partition of the original dataset:

- `forget`:
  - Contains only `label = 1` (misinformation) examples.
  - Intended as the **forget set** in unlearning experiments.

- `retain1`:
  - Contains `label = 0` (factual) examples.
    
- `retain2`:
  - Contains `label = 0` (factual) examples.

## Intended Use

- Evaluating machine unlearning methods in a misinformation vs factual setting.
- Example:
  - Forget all misinformation (`forget`).
  - Preserve utility on factual content (`retain1`, `retain2`).