create_dataset.py

import csv
from transformers import AutoTokenizer

# Initialize the tokenizer
tokenizer = AutoTokenizer.from_pretrained("TheBloke/Yarn-Llama-2-7B-128K-GPTQ", use_fast=True)

# Read the input data
with open('input.txt', 'r') as f:
    data = f.readlines()

# Initialize variables
train_data = []
test_data = []
current_row = ""
current_token_count = 0
carry_over = ""

# Iterate over each line and add to train or test data
for i, line in enumerate(data):
    line_to_add = carry_over + line.strip()
    carry_over = ""
    
    # Tokenize the line to count tokens
    tokens = tokenizer(line_to_add)['input_ids']
    num_tokens = len(tokens)

    # Check if adding the line would exceed the token limit
    if current_token_count + num_tokens > 1024:
        # Find the last period followed by a space in the current row
        last_period_idx = current_row.rfind('. ')
        
        if last_period_idx != -1:
            # Carry over the content after the last period
            carry_over = current_row[last_period_idx+2:].strip() + "\n"
            current_row = current_row[:last_period_idx+1]
        
        if i < len(data) * 0.9:
            train_data.append(current_row.strip())
        else:
            test_data.append(current_row.strip())
        
        current_row = carry_over
        current_token_count = len(tokenizer(current_row.strip())['input_ids'])

    # Add the line to the current row
    current_row += (line_to_add + "\n") if current_row else (line_to_add + "\n")
    current_token_count += num_tokens

# Save as train.csv and test.csv
with open('train.csv', 'w', newline='') as f:
    writer = csv.writer(f)
    writer.writerow(['Text'])
    for row in train_data:
        writer.writerow([row])

with open('test.csv', 'w', newline='') as f:
    writer = csv.writer(f)
    writer.writerow(['Text'])
    for row in test_data:
        writer.writerow([row])