preprocess_longtext_streaming_step2.py

import sys
import time
import inspect

from transformers import AutoTokenizer
from typing import Any
import numpy as np
from tqdm import tqdm

import json
import argparse
import os

def parse_args():
    parser = argparse.ArgumentParser(description="Finetune a transformers model on a causal language modeling task")
    parser.add_argument(
        "--batch_size",
        type=int,
        default=512,
    )
    parser.add_argument(
        "--source_file",
        type=str,
    )
    parser.add_argument(
        "--chunk_size",
        type=int,
        default=512,
    )
    args = parser.parse_args()
    return args

args = parse_args()
print('args.source_file',args.source_file)
data = open(args.source_file).readlines()
base_name = os.path.basename(args.source_file)  
file_name, _ = os.path.splitext(base_name)
bs = args.batch_size
print('############ Start data reading ###########')

local_cnt = 0
temp_dic_list = []
dic_list = []
chunk_size = args.chunk_size


for idx, line in enumerate(data):
    temp_dic = json.loads(line)
    temp_dic_list.append(temp_dic)
    local_cnt = local_cnt + 1
    if local_cnt == chunk_size:
        local_cnt = 0
        dic_list.append(temp_dic_list)
        temp_dic_list = []
        
print("len(dic_list)",len(dic_list))
with open(file_name+'_bs_'+str(bs)+'.jsonl', 'w') as f:
    for idx in range(0, len(dic_list)-bs, bs):
        for line_i in range(len(dic_list[0])):
            for i in range(bs):
                f.write(json.dumps(dic_list[idx+i][line_i]) + "\n")