-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathmake_dataset.py
178 lines (155 loc) · 7.89 KB
/
make_dataset.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
import argparse
import json
import os
os.environ['CUDA_VISIBLE_DEVICES']='0'
import re
import time
from difflib import SequenceMatcher
import joblib
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
from batch_repe import repe_pipeline_registry
from datasets import load_dataset
from sklearn.linear_model import LogisticRegression
from sklearn.metrics import accuracy_score
from sklearn.model_selection import train_test_split
from torch.utils.data import DataLoader, TensorDataset
from tqdm import tqdm
from transformers import (AutoModel, AutoTokenizer, LlamaForCausalLM,
StoppingCriteria, StoppingCriteriaList, pipeline)
from utils import *
from vllm import LLM, SamplingParams
repe_pipeline_registry()
device = torch.device("cuda")
def comparison_gsm8k(args):
sampling_params = SamplingParams(max_tokens=args.max_length)
model=LLM(model=args.model_path,gpu_memory_utilization=0.90)
tokenizer = AutoTokenizer.from_pretrained(args.model_path,padding_side='left')
tokenizer.pad_token_id = tokenizer.eos_token_id if tokenizer.pad_token_id is None else tokenizer.pad_token_id
inputs=[]
data_com=[]
data = []
with open('/home/chh/repos/my_ctg/results/gsm8k_act/__data1__chh__models__meta-llama__Meta-Llama-3-8B-Instruct/samples_gsm8k_cot_llama_train_2024-11-11T22-37-30.083129.jsonl', 'r', encoding='utf-8') as f:
for line in f:
data.append(json.loads(line))
with open('/home/chh/repos/my_ctg/instructions/template/compare_gsm8k.txt','r',encoding='utf-8') as f:
template=f.read()
for i in range(len(data)):
record=data[i]['doc']
record['result']=data[i]['exact_match']
record['generated_text']=data[i]["resps"][0][0]
data_com.append(record)
if record['result']==0.0:
inputs.append(prompt_template(tokenizer,template%(record['question'],record['answer'],record['generated_text'])))
outputs = model.generate(inputs, sampling_params)
res = [item.outputs[0].text for item in outputs]
index=0
for i in data_com:
if i['result']==0:
i['key']=res[index]
index+=1
with open(args.eval_file, 'w', encoding='utf-8') as output_file:
json.dump(data_com, output_file, ensure_ascii=False, indent=4)
def comparison_math(args):
sampling_params = SamplingParams(max_tokens=args.max_length)
model=LLM(model=args.model_path,gpu_memory_utilization=0.90)
tokenizer = AutoTokenizer.from_pretrained(args.model_path,padding_side='left')
tokenizer.pad_token_id = tokenizer.eos_token_id if tokenizer.pad_token_id is None else tokenizer.pad_token_id
inputs=[]
data_com=[]
data = []
with open('/home/chh/repos/my_ctg/results/math_act/__data1__chh__models__meta-llama__Meta-Llama-3-8B-Instruct/samples_minerva_math_algebra_2024-11-11T22-37-46.372276.jsonl', 'r', encoding='utf-8') as f:
for line in f:
data.append(json.loads(line))
with open('/home/chh/repos/my_ctg/results/math_act/__data1__chh__models__meta-llama__Meta-Llama-3-8B-Instruct/samples_minerva_math_counting_and_prob_2024-11-11T22-37-46.372276.jsonl', 'r', encoding='utf-8') as f:
for line in f:
data.append(json.loads(line))
with open('/home/chh/repos/my_ctg/results/math_act/__data1__chh__models__meta-llama__Meta-Llama-3-8B-Instruct/samples_minerva_math_geometry_2024-11-11T22-37-46.372276.jsonl', 'r', encoding='utf-8') as f:
for line in f:
data.append(json.loads(line))
with open('/home/chh/repos/my_ctg/results/math_act/__data1__chh__models__meta-llama__Meta-Llama-3-8B-Instruct/samples_minerva_math_intermediate_algebra_2024-11-11T22-37-46.372276.jsonl', 'r', encoding='utf-8') as f:
for line in f:
data.append(json.loads(line))
with open('/home/chh/repos/my_ctg/results/math_act/__data1__chh__models__meta-llama__Meta-Llama-3-8B-Instruct/samples_minerva_math_num_theory_2024-11-11T22-37-46.372276.jsonl', 'r', encoding='utf-8') as f:
for line in f:
data.append(json.loads(line))
with open('/home/chh/repos/my_ctg/results/math_act/__data1__chh__models__meta-llama__Meta-Llama-3-8B-Instruct/samples_minerva_math_prealgebra_2024-11-11T22-37-46.372276.jsonl', 'r', encoding='utf-8') as f:
for line in f:
data.append(json.loads(line))
with open('/home/chh/repos/my_ctg/results/math_act/__data1__chh__models__meta-llama__Meta-Llama-3-8B-Instruct/samples_minerva_math_precalc_2024-11-11T22-37-46.372276.jsonl', 'r', encoding='utf-8') as f:
for line in f:
data.append(json.loads(line))
with open('/home/chh/repos/my_ctg/instructions/template/compare_math.txt','r',encoding='utf-8') as f:
template=f.read()
for i in range(len(data)):
record=data[i]['doc']
record['result']=data[i]['exact_match']
record['generated_text']=data[i]["resps"][0][0]
data_com.append(record)
if record['result']==0:
inputs.append(prompt_template(tokenizer,template%(record['problem'],record['solution'],record['generated_text'])))
outputs = model.generate(inputs, sampling_params)
res = [item.outputs[0].text for item in outputs]
index=0
for i in data_com:
if i['result']==0:
i['key']=res[index]
index+=1
with open(args.eval_file, 'w', encoding='utf-8') as output_file:
json.dump(data_com, output_file, ensure_ascii=False, indent=4)
def extract(args):
with open('/home/chh/repos/my_ctg/results/gsm8k_act/res3.json', 'r', encoding='utf-8') as f:
data_gsm8k = json.load(f)
with open('./results/math_act/res2.json', 'r', encoding='utf-8') as f:
data_math = json.load(f)
data=data_gsm8k+data_math
positive_samples = []
negative_samples = []
prompt='Please judge whether to execute activation editing according to the following text, if so please output 1, otherwise output 0. You must not output any other text.\nText to be judged: {}'
for item in data:
if item['result']==1:
negative_samples.append({'instruction':prompt.format(item['generated_text']),'output':"0" })
continue
content=item['key']
# json_start = content.find('{')
# content=item['key'][json_start:]
pattern = r'\{[^{}]*\}'
match = re.findall(pattern, content)
if match:
content=match[0]
else:
content+='}'
match = re.findall(pattern, content)
if match:
content=match[0]
try:
answer = json.loads(content)
except Exception as e:
answer=None
# print(answer)
if answer:
sentences=list(answer.values())
positive_samples.append({'instruction':prompt.format(sentences[0]),'output':"1" })
positive_count = len(positive_samples)
print(positive_count)
if len(negative_samples) > positive_count:
negative_samples = random.sample(negative_samples, positive_count*3)
print(len(negative_samples))
train_data=positive_samples+negative_samples
with open('/home/chh/repos/my_ctg/sft/cls_dataset/data3.json', "w", encoding="utf-8") as f:
json.dump(train_data, f, ensure_ascii=False, indent=4)
if __name__ == "__main__":
parser = argparse.ArgumentParser()
parser.add_argument('--model_path', type=str, default='/data1/chh/models/meta-llama/Meta-Llama-3-8B-Instruct')
parser.add_argument('--seed', type=int, default=42)
parser.add_argument('--max_length', type=int, default=1024)
parser.add_argument('--eval_file', type=str, default='./results/math_act/res2.json')
parser.add_argument('--cls_dataset',type=str,default='/home/chh/repos/my_ctg/sft/cls_dataset/data.json')
args = parser.parse_args()
set_seed(args)
# comparison_gsm8k(args)
# comparison_math(args)
extract(args)