|
--- |
|
dataset_info: |
|
features: |
|
- name: text |
|
dtype: string |
|
- name: response |
|
dtype: string |
|
- name: label |
|
dtype: |
|
class_label: |
|
names: |
|
'0': 'false' |
|
'1': 'true' |
|
splits: |
|
- name: train |
|
num_bytes: 3860467.2 |
|
num_examples: 40000 |
|
- name: test |
|
num_bytes: 965116.8 |
|
num_examples: 10000 |
|
download_size: 1390903 |
|
dataset_size: 4825584.0 |
|
configs: |
|
- config_name: default |
|
data_files: |
|
- split: train |
|
path: data/train-* |
|
- split: test |
|
path: data/test-* |
|
--- |
|
# Typosquat Dataset |
|
|
|
## Dataset Summary |
|
|
|
This dataset is intended for typosquatting detection within a domain corpus. It contains 50,000 labeled pairs, categorized as either typosquatted or non-typosquatted. |
|
The data is divided into training and test splits, each maintaining a balanced distribution of positive and negative examples. |
|
|
|
## Supported Tasks and Leaderboards |
|
|
|
**T5 training**: The primary task is binary classification, specifically detecting typosquatting domains. To do so we define a new task in the T5 format and we prompt the model with both domains. |
|
The dataset can be used to train a cross-encoder or other model types for binary classification. |
|
|
|
## Languages |
|
|
|
The dataset is multilingual, reflecting the diversity of domain names. |
|
|
|
## Dataset Structure |
|
|
|
### Data Instances |
|
|
|
Each data instance in the dataset consists of two domains and a label indicating if the second domain is a typosquatted version of the first. An example from the training set: |
|
|
|
```json |
|
{'text': 'Is the first domain a typosquat of the second: lonlonsoft.com stiltsoft.net', |
|
'response': 'false', |
|
'label': 0} |
|
``` |
|
**text**: A prompt string comprised of the task definition as well as the pair of candidate domain and legitimate domain. |
|
**response**: A string representing the expected answer from the model. |
|
**label**: An integer (0 or 1) where 1 indicates a typosquatted domain and 0 indicates no typosquatting. |
|
|
|
### Data Splits |
|
|
|
The dataset is divided as follows: |
|
|
|
| Split | Number of Instances |Positive|Negative| |
|
|----------|---------------------|--------|--------| |
|
| Train | 40000 | 50% | 50% | |
|
| Test | 10000 | 50% | 50% | |
|
|
|
## Dataset Creation |
|
|
|
### Data Generation |
|
|
|
The domain pairs were generated using [ail-typo-squatting](https://github.com/typosquatter/ail-typo-squatting) |
|
Data processing includes balancing positive and negative samples to ensure even representation. |
|
|
|
### Dataset usage |
|
|
|
This dataset was developed to facilitate large-scale typosquatting detection for cybersecurity applications. |
|
It supports training and evaluating binary classifiers designed to identify domains that may have been intentionally misspelled for malicious purposes. |
|
|
|
|
