generate.py

import pandas as pd
from datasets import load_dataset

# Load the SQuAD dataset
squad_dataset = load_dataset("squad")

# Convert the 'train' and 'validation' splits to pandas DataFrames
train_df = pd.DataFrame(squad_dataset['train'])
validation_df = pd.DataFrame(squad_dataset['validation'])


import re
import pandas as pd

df = pd.concat([train_df, validation_df], ignore_index=True)


def get_closest_sentence_section(text, provided_index):
    # Regular expression to match any punctuation that ends a sentence
    sentence_end_punctuation = r'[.!?]'
    
    # Find all occurrences of sentence-ending punctuation and their indices
    punctuation_indices = [match.start() for match in re.finditer(sentence_end_punctuation, text)]
    
    # Add the start and end of the string as virtual punctuation indices
    punctuation_indices = [-1] + punctuation_indices + [len(text)]
    
    # Find the closest punctuation index above (<= provided_index)
    closest_above = max([idx for idx in punctuation_indices if idx < provided_index], default=0)
    
    # Find the closest punctuation index below (> provided_index)
    closest_below = min([idx for idx in punctuation_indices if idx > provided_index], default=len(text))
    
    # Trim the string based on closest punctuation above and below
    trimmed_text = text[closest_above + 1: closest_below + 1].strip()
    
    return trimmed_text

# Trim the context to only the relevant sentence
df['context'] = df.apply(lambda row: get_closest_sentence_section(row.context, row.answers.get('answer_start')[0]) , axis=1)
df['answer'] = df.apply(lambda row: row.answers.get('text')[0] , axis=1)

df[['title','context','question','answer']].to_parquet('tinysquad.parquet')