forked from sunzeyeah/RLHF
-
Notifications
You must be signed in to change notification settings - Fork 0
/
data_prepare.py
451 lines (402 loc) · 18.8 KB
/
data_prepare.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
import sys
sys.path.insert(0, "/root/autodl-tmp/Code/RLHF")
sys.path.insert(0, "/mnt/sfevol775196/sunzeye273/Code/chatgpt")
# sys.path.insert(0, "/mnt/share-pa002-vol682688-prd/sunzeye273/Code/chatgpt")
sys.path.insert(0, "/mnt/pa002-28359-vol543625-private/Code/chatgpt")
import os
import time
import json
import argparse
import random
import glob
import pandas as pd
import opencc
from tqdm import tqdm
from transformers import AutoTokenizer
from src.utils.logger import logger
from src.utils.file_utils import set_seed
from src.utils.nlp_utils import clean_text
def weibo_summary_comment(args, tokenizer):
''' 微博新闻+摘要+评论，每条评论有点赞数，可以根据点赞数构造reward训练集
:param args:
:param tokenizer:
:return: processed json list
'''
t = time.time()
fi = os.path.join(args.data_dir, "weibo_summary_comments_json.json")
fo = os.path.join(args.output_dir, "weibo_summary_comments.jsonl")
data = []
with open(fo, "w", encoding="utf-8") as w:
with open(fi, "r", encoding="utf-8") as r:
while True:
line = r.readline()
if not line:
break
item = json.loads(line.strip("\n"))
article = item['article'].replace(" ", "")
abstract = item['abstract'].replace(" ", "")
prompt = f"新闻内容：{article} 摘要：{abstract}"
prefix = "评论："
answers = [
{
"answer": k.replace(" ", ""),
"score": int(v)
} for (k, v) in sorted(item['comments'], key=lambda x: (int(x[1]), len(x[0])), reverse=True)
]
w.write(json.dumps({"prompt": prompt, "answers": answers}, ensure_ascii=False)+'\n')
data.append({"prompt": prompt, "answers": answers, "prefix": prefix})
logger.info(f"length: {len(data)}, time taken: {time.time()-t} s")
return data
def couplets(args, tokenizer):
''' 对联数据（上联和下联），可以根据正确下联和负例下联，构造reward训练集
:param args:
:param tokenizer:
:return: processed json list
'''
t1 = time.time()
fi = os.path.join(args.data_dir, "couplets.txt")
fo = os.path.join(args.output_dir, "couplets.jsonl")
l2 = []
nexts = dict()
with open(fi, "r", encoding="utf-8") as r:
while True:
line = r.readline()
if not line:
break
line = line.strip("\n")
idx = len(line) // 2
prompt = line[:idx]
answer = line[idx+1:]
prefix = "下联："
answers = [{"answer": answer, "score": 1}]
l2.append({"prompt": f"上联：{prompt}", "answers": answers, "prefix": prefix})
length = len(answer)
if length not in nexts:
nexts[length] = list()
nexts[length].append(answer)
t2 = time.time()
logger.info(f"length: {len(l2)}, # different lengths: {len(nexts)}, time taken: {t2-t1} s")
data = []
with open(fo, "w", encoding="utf-8") as w:
for l in tqdm(enumerate(l2), desc="Processing Couplets"):
answer = l['answers'][0]
length = len(answer['answer'])
# 上下联长度一样
nexts_tmp = set(nexts[length])
nexts_tmp.remove(answer['answer'])
nexts_tmp = set(nexts[length]).difference(set([answer['answer']]))
# nexts_tmp.remove(answer['answer'])
answers.extend([{"answer": fa, "score": 0} for fa in random.sample(nexts_tmp, 2)])
# 上下联长度不一样
keys = set(nexts.keys())
keys.remove(length)
answers.extend([{"answer": random.choice(nexts[key]), "score": -1} for key in random.sample(keys, 2)])
# answers = sorted(answers, key=lambda x: x['score'], reverse=True)
w.write(json.dumps({"prompt": l['prompt'], "answers": answers, "prefix": l['prefix']}, ensure_ascii=False)+'\n')
data.append({"prompt": l['prompt'], "answers": answers, "prefix": l['prefix']})
# if i % 1000 == 0:
# logger.info(f"{i} samples processed, time taken: {time.time()-t2} s")
logger.info(f"length: {len(data)}, time taken: {time.time()-t2} s")
return data
def zhidao(args, tokenizer):
''' 百度知道的问答数据，每条问题有多个答案以及最佳答案，可以直接构造reward训练集
:param args:
:param tokenizer:
:return: processed json list
'''
t = time.time()
fp = os.path.join(args.data_dir, "zhidao", "*.csv")
fo = os.path.join(args.output_dir, "zhidao.jsonl")
data = []
with open(fo, "w", encoding="utf-8") as w:
for fi in glob.glob(fp):
df = pd.read_csv(fi).sort_values(by=["title", "is_best"], ascending=False)
prev_title = None
prev_prompt = None
prefix = "答："
for _, val in df.iterrows():
if isinstance(val['question'], str) and val['question'] != val['title']:
prompt = f"问题：{val['title']} 内容：{val['question']}"
else:
prompt = f"问题：{val['title']}"
if prev_title is not None and prev_title == val['title']:
answers.append({"answer": val['reply'], "score": val['is_best']})
else:
if prev_title is not None:
# l3.append({"prompt": prev_prompt, "answers": copy.deepcopy(answers)})
w.write(json.dumps({"prompt": prev_prompt, "answers": answers}, ensure_ascii=False)+'\n')
data.append({"prompt": prev_prompt, "answers": answers})
answers = [{"answer": val['reply'], "score": val['is_best']}]
prev_prompt = prompt
prev_title = val['title']
# l3.append({"prompt": prev_prompt, "answers": copy.deepcopy(answers)})
w.write(json.dumps({"prompt": prev_prompt, "answers": answers, "prefix": prefix}, ensure_ascii=False)+'\n')
data.append({"prompt": prev_prompt, "answers": answers, "prefix": prefix})
logger.info(f"finished processing {os.path.basename(fi)}")
logger.info(f"length: {len(data)}, time taken: {time.time()-t} s")
return data
def chinese_classical(args, tokenizer):
''' 文言文和现代文的对照翻译，每条文言文有对应的现代文翻译，可以根据正确翻译或原文+负例翻译或原文，构造reward训练集
:param args:
:param tokenizer:
:return: processed json list
'''
t1 = time.time()
fp = os.path.join(args.data_dir, "Classical-Modern", "bitext", "*")
fo = os.path.join(args.output_dir, "chinese_classical.jsonl")
l3 = []
dicts = dict()
for fi in glob.glob(fp):
name = os.path.basename(fi)
dicts[name] = {"古文": [], "现代文": []}
with open(fi, "r", encoding="utf-8") as r:
for i, line in enumerate(r):
line = line.strip("\n")
if line.startswith("古文"):
p1 = line[3:]
dicts[name]['古文'].append(p1)
elif line.startswith("现代文"):
p2 = line[4:]
dicts[name]['现代文'].append(p2)
elif p1 is not None and p2 is not None:
pair = [("古文", p1), ("现代文", p2)]
random.shuffle(pair)
prompt = f"{pair[0][0]}：{pair[0][1]}"
prefix = f"{pair[1][0]}："
answers = [{"answer": pair[1][1], "score": 1}]
l3.append({"prompt": prompt, "answers": answers, "prefix": prefix, "name": name})
p1 = None
p2 = None
t2 = time.time()
logger.info(f"length: {len(l3)}, # different names: {len(dicts)}, time taken: {t2-t1} s")
data = []
with open(fo, "w", encoding="utf-8") as w:
for l in tqdm(enumerate(l3), desc="Processing Chinese Classical-Modern"):
name = l['name']
prompt = l['prompt']
prefix = l['prefix']
answer = l['answers'][0]['answer']
if prompt.startswith("古文"):
answer_type = '现代文'
else:
answer_type = '古文'
samples_tmp = set(dicts[name][answer_type])
samples_tmp.remove(answer)
answers.extend([{"answer": fa, "score": 0} for fa in random.sample(samples_tmp, 2)])
keys = set(dicts.keys())
keys.remove(name)
answers.extend([{"answer": random.choice(dicts[key][answer_type]), "score": -1} for key in random.sample(keys, 2)])
w.write(json.dumps({"prompt": prefix, "answers": answers}, ensure_ascii=False)+'\n')
data.append({"prompt": prefix, "answers": answers})
# if i % 100 == 0:
# logger.info(f"{i} samples processed, time taken: {time.time()-t2} s")
logger.info(f"length: {len(data)}, time taken: {time.time()-t2} s")
return data
def chinese_poetry(args, tokenizer):
''' 四书五经、诗、词、曲等古文数据，每篇文章或每首诗有作者、题目和正文，可以根据正确体裁/作者+负例体裁/作者，构造reward训练集
:param args:
:param tokenizer:
:return: processed json list
'''
converter = opencc.OpenCC('t2s.json')
t1 = time.time()
fp = [
# 四书五经
os.path.join(args.data_dir, "chinese-poetry", "lunyu", "lunyu.json"),
os.path.join(args.data_dir, "chinese-poetry", "sishuwujing", "*.json"),
# 古体诗
os.path.join(args.data_dir, "chinese-poetry", "caocaoshiji", "caocao.json"),
os.path.join(args.data_dir, "chinese-poetry", "shijing", "shijing.json"),
# 楚辞
os.path.join(args.data_dir, "chinese-poetry", "chuci", "chuci.json"),
# 诗
os.path.join(args.data_dir, "chinese-poetry", "shi", "poet*.json"),
# 词
os.path.join(args.data_dir, "chinese-poetry", "ci", "ci*.json"),
os.path.join(args.data_dir, "chinese-poetry", "nalanxingde", "*.json"),
os.path.join(args.data_dir, "chinese-poetry", "wudai", "huajianji", "*juan.json"),
os.path.join(args.data_dir, "chinese-poetry", "wudai", "nantang", "poetrys.json"),
# 曲
os.path.join(args.data_dir, "chinese-poetry", "yuanqu", "yuanqu.json"),
]
fs = [each for f in fp for each in glob.glob(f)]
l5 = []
dicts = dict()
for fi in fs:
lines = json.load(open(fi, "r", encoding="utf-8"))
if isinstance(lines, dict):
lines = [lines]
for i, line in enumerate(lines):
if "lunyu" in fi:
author = "孔子"
genre = "经书"
title = line['chapter']
contents = "".join(line['paragraphs'])
elif "daxue" in fi:
author = "曾子"
genre = "经书"
title = "大学"
contents = converter.convert("".join(line['paragraphs'])).replace("「", "“").replace("」", "”")
elif "mengzi" in fi:
author = "孟子"
genre = "经书"
title = converter.convert(line['chapter'])
contents = converter.convert("".join(line['paragraphs'])).replace("「", "“").replace("」", "”")
elif "zhongyong" in fi:
author = "孔伋"
genre = "经书"
title = "中庸"
contents = converter.convert("".join(line['paragraphs'])).replace("「", "“").replace("」", "”")
elif "caocao" in fi:
author = "曹操"
genre = "古体诗"
title = line['title']
contents = "".join(line['paragraphs'])
elif "shijing" in fi:
author = "诗经"
genre = "古体诗"
title = line['chapter'] + "-" + line['section'] + "-" + line['title']
contents = "".join(line['content'])
elif "chuci" in fi:
author = line['author']
genre = "楚辞"
title = line['section'] + "-" + line['title']
contents = "".join(line['content'])
elif "nalanxingde" in fi:
author = line['author']
genre = "词"
title = line['title']
contents = "".join(line['para'])
elif "huajianci" in fi:
author = line['author']
genre = "词"
title = line['title']
contents = "".join(line['paragraphs'])
elif "nantang" in fi:
author = line['author']
genre = "词"
title = line['title']
contents = "".join(line['paragraphs'])
elif "yuanqu" in fi:
author = line['author']
genre = "曲"
title = line['title']
contents = "".join(line['paragraphs'])
elif "shi" in fi:
if len(line['paragraphs']) <= 0:
continue
author = converter.convert(line['author'])
genre = "五言诗" if len(line['paragraphs'][0]) == 12 else "七言诗"
title = converter.convert(line['title'])
contents = converter.convert("".join(line['paragraphs']))
elif "ci" in fi:
author = line['author']
genre = "词"
title = line['rhythmic']
contents = "".join(line['paragraphs'])
if genre not in dicts:
dicts[genre] = dict()
if author not in dicts[genre]:
dicts[genre][author] = dict()
quantifier = "篇" if genre in ["经书", "楚辞"] else "首"
prompt = f"以{author}的风格，写一{quantifier}{genre}，题为{title}"
answers = [{"answer": contents, "score": 1}]
l5.append({"prompt": prompt, "answers": answers, "genre": genre, "title": title, "author": author})
dicts[genre][author][title] = contents
t2 = time.time()
logger.info(f"length: {len(l5)}, # different lengths: {len(dicts)}, time taken: {t2-t1} s")
data = []
fo = os.path.join(args.output_dir, "chinese_poetry.jsonl")
with open(fo, "w", encoding="utf-8") as w:
for l in tqdm(enumerate(l5), desc="Processing Chinese Poetry"):
genre = l['genre']
author = l['author']
title = l['title']
prompt = l['prompt']
answers = l['answers']
# 同作者其他作品-2
titles_tmp = set(dicts[genre][author].keys())
titles_tmp.remove(title)
if len(titles_tmp) > 0:
t = random.choice(list(titles_tmp))
answers.append({"answer": dicts[genre][author][t], "score": 0})
# 同体裁其他作者其他作品-1
authors_tmp = set(dicts[genre].keys())
authors_tmp.remove(author)
a = random.choice(list(authors_tmp))
t = random.choice(list(dicts[genre][a].keys()))
answers.append({"answer": dicts[genre][a][t], "score": -1})
# 不同体裁作品-0
genres_tmp = set(dicts.keys())
genres_tmp.remove(genre)
g = random.choice(list(genres_tmp))
a = random.choice(list(dicts[g].keys()))
t = random.choice(list(dicts[g][a].keys()))
answers.append({"answer": dicts[g][a][t], "score": -2})
w.write(json.dumps({"prompt": prompt, "answers": answers, "prefix": ""}, ensure_ascii=False)+'\n')
data.append({"prompt": prompt, "answers": answers, "prefix": ""})
logger.info(f"length: {len(data)}, time taken: {time.time()-t2} s")
return data
def baike_qa_2019(args, tokenizer):
''' 百科问答数据集，每个问题只有一个答案，可构造sft训练集
:param args:
:param tokenizer:
:return: processed json list
'''
fs = glob.glob(os.path.join(args.data_dir, "baike_qa2019", "baike_qa_*.json"))
fo = os.path.join(args.output_dir, "baike_qa.jsonl")
data = []
t = time.time()
with open(fo, "w", encoding="utf-8") as w:
for f in fs:
with open(f, "r", encoding="utf-8") as r:
while True:
line = r.readline()
if not line:
break
item = json.loads(line.strip("\n"))
question = clean_text(item['title'] if len(item['title']) > len(item['desc']) else item['desc'])
prompt = question
prefix = "答："
answer = clean_text(item['answer'])
answers = [{"answer": answer, "score": 1}]
w.write(json.dumps({"prompt": prompt, "answers": answers, "prefix": prefix}, ensure_ascii=False)+'\n')
data.append({"prompt": prompt, "answers": answers, "prefix": prefix})
logger.info(f"length: {len(data)}, time taken: {time.time()-t} s")
return data
def get_parser():
parser = argparse.ArgumentParser()
parser.add_argument("--data_dir", type=str, required=True)
parser.add_argument("--output_dir", type=str, required=True)
parser.add_argument("--model_name_or_path", type=str, required=True)
parser.add_argument("--seed", type=int, default=42)
args = parser.parse_args()
return args
def main():
args = get_parser()
logger.info(f"Parameters: {args}")
# set random seed
set_seed(args.seed)
# load tokenizer
tokenizer = AutoTokenizer.from_pretrained(args.model_name_or_path, use_cache=False, trust_remote_code=True)
# process raw datasets
data_weibo = weibo_summary_comment(args, tokenizer)
data_couplets = couplets(args, tokenizer)
data_zhidao = zhidao(args, tokenizer)
data_chinese_classical = chinese_classical(args, tokenizer)
data_chinese_poetry = chinese_poetry(args, tokenizer)
data_baike = baike_qa_2019(args, tokenizer)
# merge processed datasets
data = data_weibo + data_couplets + data_zhidao + data_chinese_classical + data_chinese_poetry + data_baike
random.shuffle(data)
fo = os.path.join(args.output_dir, "train_data_external_v1.jsonl")
with open(fo, "w", encoding="utf-8") as w:
for d in data:
w.write(json.dumps(d, ensure_ascii=False)+'\n')
fo = os.path.join(args.output_dir, "dev_data_external_v1.jsonl")
with open(fo, "w", encoding="utf-8") as w:
for d in data[:10000]:
w.write(json.dumps(d, ensure_ascii=False)+'\n')
logger.info("Finished saving processed train & dev files")
if __name__ == "__main__":
main()