dataset-preprocessor.py

from datasets import load_dataset
import re
import random

def split_into_paragraphs(text):
    # Split by markdown headers or double newlines
    paragraphs = re.split(r'\n\n|(?=^#)', text, flags=re.MULTILINE)
    return [p.strip() for p in paragraphs if p.strip()]

def create_input_output_pairs(example):
    paragraphs = example['paragraphs']
    n_paragraphs = len(paragraphs)
    
    # Randomly select about half of the paragraphs for input
    n_input = max(1, random.randint(n_paragraphs // 2 - 1, n_paragraphs // 2 + 1))
    
    input_paragraphs = paragraphs[:n_input]
    output_paragraphs = paragraphs[n_input:]
    
    return {
        'inputs': ' '.join(input_paragraphs),
        'targets': ' '.join(output_paragraphs)
    }

def preprocess_dataset(dataset_name, text_column='text'):
    # Load the dataset
    dataset = load_dataset(dataset_name)
    
    # Split text into paragraphs
    dataset = dataset.map(
        lambda example: {'paragraphs': split_into_paragraphs(example[text_column])},
        remove_columns=[text_column]
    )
    
    # Create input-output pairs
    preprocessed_dataset = dataset.map(
        create_input_output_pairs,
        remove_columns=['paragraphs']
    )
    
    return preprocessed_dataset

# Usage example
if __name__ == "__main__":
    # Replace 'your_dataset' with the actual dataset name
    dataset_name = 'your_dataset'
    
    preprocessed_dataset = preprocess_dataset(dataset_name)
    
    # Print some examples
    print(preprocessed_dataset['train'][:5])
    
    # Save the preprocessed dataset
    preprocessed_dataset.save_to_disk("preprocessed_dataset")