RISE/construct_dpo_samples.py at master · kaishxu/RISE · GitHub

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
import os
import re
import json
import argparse
from tqdm import trange
import random

def check_is_error(prompt, text):
    if "wrong" in prompt or "correct" in prompt:
        return False

    error_lst = ["a mistake", "mistakenly", "misleading", "Misleading", "misinterpreted", "Let's correct", "let's correct", "The correct", "Let's reconsider", "let's reconsider", "incorrect", "wrong", "python", "This is not correct", "this is not correct"]
    for error in error_lst:
        if error in text:
            return True
    return False

def construct(prompt_path, completion_path, chosen_sample_path, full_sample_path, save_sample_path):
    with open(prompt_path, "r") as infile:
        prompt_lst = [json.loads(x) for x in infile]
    with open(completion_path, "r") as infile:
        response_lst = json.load(infile)
    with open(chosen_sample_path, "r") as infile:
        ds = json.load(infile)

    correct_response_lst = []
    correct_idx_lst = []
    for idx in trange(len(response_lst), desc="Extract edited responses"):
        response = response_lst[idx]["completion"]
        response = "```\n" + response

        tmp = re.findall(r"```\n(.+?)\n```", response, flags=re.DOTALL)

        if len(tmp) > 0:
            x = tmp[0]
            if not check_is_error(ds[idx]["prompt"], x):
                response_lst[idx]["edit"] = x
                correct_response_lst.append(response_lst[idx])
                correct_idx_lst.append(response_lst[idx]["idx"])

    wrong_idx_lst = []
    for idx in range(len(response_lst)):
        if response_lst[idx]["idx"] not in correct_idx_lst:
            wrong_idx_lst.append(response_lst[idx]["idx"])

    correct_response_lst.sort(key=lambda x: x["idx"])
    response_dict = {x["idx"]: x for x in correct_response_lst}

    new_ds = []
    for idx in trange(len(ds), desc="Constructing samples"):
        new_ds.append(ds[idx])
        new_ds[-1]["rejected"] = new_ds[-1]["chosen"]

        if idx in response_dict:
            # text_to_edit = re.findall(r"Text:\n```\n(.+?)\n```", response_dict[idx]["prompt"], flags=re.DOTALL)
            assert prompt_lst[idx]["text_to_edit"] in response_dict[idx]["prompt"]
            new_ds[-1]["rejected"] = new_ds[-1]["rejected"].replace(
                prompt_lst[idx]["text_to_edit"],
                response_dict[idx]["edit"]
            )
        else:
            del new_ds[-1]

    for idx in range(len(new_ds)):

        tmp = re.findall(r"( \\boxed\{.+?\}\.)", new_ds[idx]["rejected"])
        if len(tmp) > 0:
            for x in tmp:
                new_ds[idx]["rejected"] = new_ds[idx]["rejected"].replace(x, " "+x[8:-2]+".")

        tmp = re.findall(r"(\\\(\\boxed\{.+?\}\\\))", new_ds[idx]["rejected"])
        if len(tmp) > 0:
            for x in tmp:
                new_ds[idx]["rejected"] = new_ds[idx]["rejected"].replace(x, x[9:-3])

        tmp = re.findall(r"(\\\[\\boxed\{.+?\}\\\])", new_ds[idx]["rejected"])
        if len(tmp) > 0:
            for x in tmp:
                new_ds[idx]["rejected"] = new_ds[idx]["rejected"].replace(x, x[9:-3])

        tmp = re.findall(r"( \\boxed\{.+?\}$)", new_ds[idx]["rejected"])
        if len(tmp) > 0:
            for x in tmp:
                new_ds[idx]["rejected"] = new_ds[idx]["rejected"].replace(x, " "+x[8:-1])

        tmp = re.findall(r"(\\\( \\boxed\{.+?\} \\\))", new_ds[idx]["rejected"])
        if len(tmp) > 0:
            for x in tmp:
                new_ds[idx]["rejected"] = new_ds[idx]["rejected"].replace(x, x[10:-4])

        tmp = re.findall(r"(\\\[\n\\boxed\{.+?\}\n\\\])", new_ds[idx]["rejected"])
        if len(tmp) > 0:
            for x in tmp:
                new_ds[idx]["rejected"] = new_ds[idx]["rejected"].replace(x, x[10:-4])

        tmp = re.findall(r"(\$\\boxed\{.+?\}\$)", new_ds[idx]["rejected"])
        if len(tmp) > 0:
            for x in tmp:
                new_ds[idx]["rejected"] = new_ds[idx]["rejected"].replace(x, x[8:-2])

        x = new_ds[idx]["rejected"]
        while True:
            try:
                sep = "\\boxed{"
                start = x.index(sep)
                num_bracket = 1  #sep contatin 1 bracket
                for i in range(start+len(sep), len(x)):
                    if x[i] == "{":
                        num_bracket += 1
                    elif x[i] == "}":
                        num_bracket -= 1
                    if num_bracket == 0:
                        break
                x = x.replace(x[start:i+1], x[start:i+1][7:-1])
            except:
                break
        new_ds[idx]["rejected"] = x

    # merge solution pair and step pair
    from collections import defaultdict

    step_pair_dict = defaultdict(list)
    for sample in new_ds:
        step_pair_dict[sample["prompt"]].append(sample)

    with open(full_sample_path, "r") as infile:
        full_sample_lst = json.load(infile)

    new_ds = []
    num_edit = 3
    for idx in range(len(full_sample_lst)):
        prompt = full_sample_lst[idx]["prompt"]
        new_ds.append(full_sample_lst[idx])
        if prompt in step_pair_dict:
            # if len(step_pair_dict[prompt]) >= num_edit:
            #     edit_sample = random.sample(step_pair_dict[prompt], num_edit)
            # else:
            #     edit_sample = step_pair_dict[prompt]
            # new_ds += edit_sample

            edit_sample = random.choice(step_pair_dict[prompt])
            new_ds.append(edit_sample)

    print("Total samples:", len(new_ds))

    with open(save_sample_path, "w") as outfile:
        json.dump(new_ds, outfile, indent=4)

if __name__ == "__main__":
    parser = argparse.ArgumentParser()
    parser.add_argument("--prompt_path", type=str, default='')
    parser.add_argument("--completion_path", type=str, default='')
    parser.add_argument("--chosen_sample_path", type=str, default='')
    parser.add_argument("--full_sample_path", type=str, default='')
    parser.add_argument("--save_sample_path", type=str, default='')
    args = parser.parse_args()

    # construct samples for DPO
    construct(args.prompt_path,
            args.completion_path,
            args.chosen_sample_path,
            args.full_sample_path,
            args.save_sample_path)