ilyyeees/leetspeak-to-english
1337speak-to-English V3 Dataset Dataset Description This is a large-scale synthetic dataset designed to train models (like ByT5) to decode "Leetspeak", internet slang, and corrupted text back into clean, standard English. It contains ~720k pairs of text, generated through a sophisticated hybrid pipeline that combines rule-based corruption with LLM-generated slang and semantic alterations. Goal To enable robust text normalization models that can… See the full description on the dataset page: https://huggingface.co/datasets/ilyyeees/leetspeak-to-english.
1337speak-to-English V3 Dataset
<div align="center"> <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformerslogoname.png" width="300"/> </div>
Dataset Description
This is a large-scale synthetic dataset designed to train models (like ByT5) to decode "Leetspeak", internet slang, and corrupted text back into clean, standard English.
It contains ~720k pairs of text, generated through a sophisticated hybrid pipeline that combines rule-based corruption with LLM-generated slang and semantic alterations.
Goal
To enable robust text normalization models that can handle:
- Visual substitutions:
3fore,$fors,|_forL - Phonetic slang:
wut→what,kewl→cool - Grammar/Shortening:
idk→I don't know,ur→your/you're - Edge cases: Preserving numbers, dates, and special characters correctly.
Data Generation Pipeline
The dataset was constructed using a "Reverse Pipeline" approach (Clean → Corrupt), utilizing three primary sources:
- WikiText: For formal, grammatically complex sentence structures.
- ELI5 (Explain Like I'm 5): For conversational, question-answer style natural language.
- LLM Synthetic: Fully synthetic sentences generated to include modern internet culture, gaming references (
gg,noob), and Gen Z slang (no cap,bussin).
Corruption Strategy
Unlike simple find-and-replace scripts, this dataset uses a Hybrid Corruption Engine:
- LLM Layer: First revises clean text to include slang, grammar lazy-ness, and internet-speak (e.g., changing "I am going to" to "imma").
- Script Layer: Applies randomized, weighted visual substitutions (Leetspeak) with varying intensities (Light, Medium, Heavy, 1337).
- Adversarial Filtering: (Optional subset) Samples were filtered to prioritize "hard" negatives that previous model versions failed to decode.
Dataset Structure
The dataset consists of a single JSONL-style split with the following columns:
Example Data
Usage
from datasets import load_dataset
dataset = load_dataset("ilyyeees/leetspeak-to-english-v3")
# Print first example
print(dataset['train'][0])
# {'input_text': '...', 'target_text': '...'}License
MIT License. Free to use for research and commercial applications.
