📝 Text boundaries

📖 Overview

**Text boundaries ** is a powerful Python script designed to help you clean, preprocess, and analyze text data! 🚀 This script reads a text file, filters lines containing only alphabetical characters, removes duplicates, sorts them, and generates detailed statistics about the data. It’s perfect for finding the boundaries of datasets for machine learning models, natural language processing, or validating dataset quality. 🧑‍💻

✨ Features

• 🔍 Filter Lines: Retains only lines with alphabetical characters to ensure data consistency.
• 🗑️ Remove Duplicates: Ensures every line is unique for effective data preprocessing.
• 🔠 Sort Case-Sensitively: Keeps the data order sensitive to case, which is useful for NLP tasks.
• 📊 Generate Statistics: Provides insights such as:
  • Number of lines starting or ending with each letter.
  • Longest word starting with each letter.
  • First and last word starting with each letter.
• 📂 Outputs:
  • Cleaned dataset is saved back to dataset.txt.
  • Detailed statistics are written to boundaries.log.

🛠️ Prerequisites

• Python 3.9+: Make sure Python is installed and available in your system's PATH.

🚀 Quick Start

1. Clone the Repository:

   git clone https://github.com/fabriziosalmi/text-boundaries.git
   cd text-boundaries

2. Add Your Data: Place your text data in a file named dataset.txt in the repository directory.

3. Run the Script:

   python process_dataset.py

4. Check the Outputs:

   • Updated data in dataset.txt.
   • Statistics in boundaries.log.

🔍 Use Cases and Examples

1. Finding Dataset Boundaries for NLP Preprocessing

When working with text datasets, such as those used to train or fine-tune language models, it's essential to clean the data and identify the boundaries of valid input. This script helps in preprocessing text data by removing invalid entries and generating useful statistics that help define the boundaries of a valid dataset.

Example: NLP Dataset Validation

Given an input file dataset.txt containing:

apple
Banana
!@#$%^
carrot
apple
Apricot
123456
dragonfruit
apple
Banana

Running the script will produce:

• Filtered and Cleaned Data in dataset.txt:

  Apricot
  Banana
  apple
  carrot
  dragonfruit
  

• Detailed Statistics in boundaries.log:

  Total rows: 5
  A: Starts 2, Ends 0, Longest Apricot, First Apricot, Last Apricot
  B: Starts 1, Ends 0, Longest Banana, First Banana, Last Banana
  C: Starts 1, Ends 0, Longest carrot, First carrot, Last carrot
  D: Starts 1, Ends 0, Longest dragonfruit, First dragonfruit, Last dragonfruit
  

This allows you to quickly understand the composition and range of your dataset, helping ensure that your data meets the requirements for training NLP models.

2. Identifying Boundaries in Large Language Model (LLM) Training Data

When preparing datasets for LLMs, understanding the range of data (such as the first and last words alphabetically, or the longest word per letter) can be critical in optimizing data loading, tokenization strategies, or even debugging unexpected behavior in trained models.

Example: Preparing Data for LLM Training

Imagine you have a dataset of vocabulary terms that will be used to fine-tune an LLM. The dataset contains words from various domains, but you need to identify the longest word and the range of starting letters to manage tokenization effectively.

Given the input file dataset.txt:

neuroscience
abracadabra
xylophone
Quantum
algorithm
Artificial
Zoo
nightmare
alpha

After running the script, boundaries.log provides insights like:

Total rows: 9
A: Starts 3, Ends 1, Longest Artificial, First Artificial, Last algorithm
N: Starts 2, Ends 0, Longest neuroscience, First neuroscience, Last nightmare
Q: Starts 1, Ends 0, Longest Quantum, First Quantum, Last Quantum
X: Starts 1, Ends 0, Longest xylophone, First xylophone, Last xylophone
Z: Starts 1, Ends 0, Longest Zoo, First Zoo, Last Zoo

From this output, you can see that:

• The dataset has a balanced mix of starting letters.
• The longest word is "neuroscience".
• Words start with different letters, from A to Z.

3. Finding Boundaries in Multi-language Datasets

For multi-language datasets, especially when developing multilingual LLMs, you might want to check that your dataset contains a balanced range of entries across different languages. This script helps in identifying such patterns.

Example: Multi-language Dataset Analysis

Input file dataset.txt:

こんにちは
Hello
Bonjour
Hola
안녕하세요
Hallo
你好
Hola
Bonjour

Running the script will yield:

• Filtered Data (since non-Latin characters will be filtered out):

  Bonjour
  Hello
  Hola
  

• Statistics:

  Total rows: 3
  B: Starts 1, Ends 0, Longest Bonjour, First Bonjour, Last Bonjour
  H: Starts 2, Ends 0, Longest Hello, First Hello, Last Hola
  

You can use this information to adjust your dataset for balanced language representation or decide on additional preprocessing steps.

🖥️ Automate with GitHub Actions

Automate your dataset processing with GitHub Actions! 🤖 Whenever dataset.txt is updated, the script runs and updates the results automatically.

🔄 Workflow Configuration

To set up automation, create a .github/workflows/process_dataset.yml file with:

name: Process Dataset

on:
  workflow_dispatch:
  push:
    paths:
      - 'dataset.txt'  # Trigger the workflow when dataset.txt is updated

jobs:
  update-dataset:
    runs-on: ubuntu-latest
    steps:
    - name: Checkout code
      uses: actions/checkout@v2

    - name: Set up Python
      uses: actions/setup-python@v2
      with:
        python-version: '3.9'

    - name: Install dependencies
      run: |
        python -m pip install --upgrade pip
        echo 'No dependencies to install'

    - name: Process dataset
      run: python process_dataset.py

    - name: Commit changes
      run: |
        git config --global user.name 'Fabrizio Salmi'
        git config --global user.email '[email protected]'
        git add dataset.txt boundaries.log
        git commit -m "Update dataset and boundaries"
        git push

🤝 Contributing

Contributions are welcome! 🎉 If you have ideas to improve the script or find a bug, feel free to open an issue or submit a pull request.

📜 License

This project is licensed under the MIT License. See the LICENSE file for details.