wikipedia_v2 / make_pack_jsonl_for_wiki.py
kentof's picture
Upload folder using huggingface_hub
9a4330f verified
raw
history blame
9.37 kB
import json
import os
from tqdm import tqdm
from transformers import AutoModelForCausalLM, AutoTokenizer
import re
import concurrent.futures
from pathlib import Path
# 検索するディレクトリのパス
file_path = 'wiki_concat_jsonl.jsonl'
model_id = "geniacllm/dMoE_8B_math_iter8999"
tokenizer = AutoTokenizer.from_pretrained(model_id, use_fast=False, torch_dtype="auto", device_map="auto")
print(" load tokenizer ")
# .jsonl ファイルのフルパスを格納するリスト
jsonl_files = []
jsonl_datas = []
with open(file_path, 'r', encoding='utf-8') as file:
for line in file:
# JSON形式のデータをPythonの辞書に変換
data = json.loads(line)
jsonl_datas.append(data["text"])
# print("data", gsm8l_datas )
print("data", len(jsonl_datas) )
print("data", jsonl_datas[-1] )
# model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype="auto", device_map="auto")
# print(model)
# inputs = tokenizer("地球温暖化の主な原因となる物質は", return_tensors="pt").to(model.device)
# inputs
all_data = []
word_num = 0
short_full_document_list =[]
complete_full_document_list =[]
count = 0
total_items = len(jsonl_datas)
# big_list = list(range(total_items))
# サブリストの数
num_chunks = 25
# 各サブリストの長さを計算
items_per_sublist = total_items // num_chunks
# サブリストを生成
chunks = [jsonl_datas[i * items_per_sublist:(i + 1) * items_per_sublist] for i in range(num_chunks)]
# 最後のサブリストに残りのアイテムを追加する
if total_items % num_chunks != 0:
remaining_items = jsonl_datas[num_chunks * items_per_sublist:]
chunks[-1].extend(remaining_items)
# 並列を行う関数
def process_item(sublists):
for index, one_data in tqdm( enumerate(sublists)):
one_data_token = tokenizer(one_data, return_tensors="pt", add_special_tokens=True)
one_data_token_len = one_data_token.input_ids[0].size(0)
if one_data_token_len < 50:
pass
elif one_data_token_len < 1600:
short_full_document_list.append( {'length': one_data_token_len, 'text': one_data } )
elif one_data_token_len < 2010:
complete_full_document_list.append( {'length': one_data_token_len, 'text': one_data } )
else:
one_sentences = one_data.split('。')
collect_one_sentences = []
split_one_sentence_len = 0
for index, sentence in enumerate(one_sentences):
one_sentence_token = tokenizer(sentence, return_tensors="pt", add_special_tokens=True)
one_sentence_token_len = one_sentence_token.input_ids[0].size(0)
if split_one_sentence_len + one_sentence_token_len < 2010:
split_one_sentence_len = split_one_sentence_len + one_sentence_token_len
collect_one_sentences.append( sentence )
elif one_sentence_token_len > 2010:
print(" Warning : Over one-sentence token :", one_sentence_token_len )
print( sentence )
split_document = "\n".join(collect_one_sentences)
short_full_document_list.append( {'length': split_one_sentence_len, 'text': split_document } )
split_one_sentence_len = 0
collect_one_sentences.clear()
else :
split_document = "\n".join(collect_one_sentences)
complete_full_document_list.append( {'length': split_one_sentence_len, 'text': split_document } )
split_one_sentence_len = one_sentence_token_len
collect_one_sentences.clear()
collect_one_sentences.append( sentence )
split_document = "\n".join(collect_one_sentences)
if split_one_sentence_len < 1500 :
short_full_document_list.append( {'length': split_one_sentence_len, 'text': split_document } )
elif split_one_sentence_len < 2015 :
complete_full_document_list.append( {'length': split_one_sentence_len, 'text': split_document } )
else:
print(" Warning : Over split-one-sentence token :", split_one_sentence_len )
print( split_document )
# if index > 96605:
# break
complete_full_document_sub_list = []
while True:
if len(short_full_document_list) == 0:
break
if len(short_full_document_list)%500 == 0:
print( "残り : ", len(short_full_document_list) )
target = short_full_document_list[0]
left_len = 2010 - target['length']
if left_len < 0 :
print(" Error : Over token target :", target['length'] )
del short_full_document_list[0] # 0番目の要素を削除
if len(short_full_document_list) == 0:
complete_full_document_sub_list.append( target )
break
else:
# ステップ1: 最小差分を計算(デフォルト値を使用)
closest_length_diff = min((left_len - rec['length'] for rec in short_full_document_list if rec['length'] <= left_len), default=None)
if closest_length_diff is None:
complete_full_document_sub_list.append( target )
# print( "complete_full_document_list.append" )
else:
# 特定の条件に合致するデータのインデックスを探す
index = next((i for i, rec in enumerate(short_full_document_list) if rec['length'] == left_len - closest_length_diff ), None)
merge_document = short_full_document_list[index]
del short_full_document_list[index] # index番目の要素を削除
# 結果の出力
if index is not None:
merge_texts = target['text'] + "</s>" + merge_document['text']
if target['length'] + merge_document['length'] < 1500 :
short_full_document_list.append( {'length': target['length'] + merge_document['length'] + 1 , 'text': merge_texts } )
# print( "merge_texts" , merge_texts )
# print( "short_full_document_list" , short_full_document_list[-1] )
elif target['length'] + merge_document['length'] > 2022 :
print(" Error : Over token target :", target['length'] , " merge_document : ", merge_document['length'] )
else:
complete_full_document_sub_list.append( {'length': target['length'] + merge_document['length'] + 1 , 'text': merge_texts } )
# print( "merge_texts" , merge_texts )
else:
print(f"長さが {merge_document['length'] } のデータは見つかりませんでした。")
return complete_full_document_sub_list
# print( "L110 sublists " , len( chunks) )
# プロセスプールを使用して各チャンクを並列処理
with concurrent.futures.ProcessPoolExecutor() as executor:
results = list(executor.map(process_item, chunks))
# print( "L114 results " , len( results) )
# for result in results:
# print( "L116 result " , len( result) )
# リスト内包表記を使って平坦化
# sumを使って平坦化
complete_full_document_list.extend( sum(results, [] ))
print( "L121 complete_full_document_list " , len( complete_full_document_list) )
length_list = []
collect_length_list = []
file_count = 0
complete_full_document_list_len = len(complete_full_document_list)
# 各ファイルに格納する最大ドキュメント数
max_docs_per_file = 9999
# 必要なファイルの数
jsonl_num = complete_full_document_list_len // max_docs_per_file + (1 if complete_full_document_list_len % max_docs_per_file != 0 else 0)
# ファイルにデータを分割して保存
for i in range(jsonl_num):
start_index = i * max_docs_per_file
end_index = start_index + max_docs_per_file
# ファイル名を生成し、連番を付ける
filename = f"{i}_speech_text.jsonl" # ファイル名形式: "番号_名前.jsonl"
# ファイルを開いてデータを書き込む
with open(filename, 'w', encoding='utf-8') as file:
for document in complete_full_document_list[start_index:end_index]:
# 各テキストに対するJSONオブジェクトを作成
json_obj = {
"text": document['text'] ,
"is_rejected": False,
"reason": {}
}
# JSON形式の文字列に変換
json_line = json.dumps(json_obj, ensure_ascii=False)
# ファイルに書き込み
file.write(json_line + '\n')
length_list.append( document['length'] )
document_tokens = tokenizer(document['text'] , return_tensors="pt", add_special_tokens=True)
document_tokens_len = document_tokens.input_ids[0].size(0)
collect_length_list.append( document_tokens_len )
if document_tokens_len > 2021:
print("L209 Error : Over token ", document['length'] , " collect : ", document_tokens_len )
print( "len(complete_full_document_list) " , len(complete_full_document_list) )
# print( "length_list " , length_list )