InternAL/generate_internal.py at main · THUMLP/InternAL · 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

import pandas as pd
import json
import random
from tqdm import tqdm
import argparse

random.seed(0)
df = pd.read_excel('primekg_relations.xlsx')
template_dict = {}
template_plural_dict = {}
for i, row in df.iterrows():
    print(row['head'], row['relation'], row['tail'])
    template = row['color_ref']
    template_plural = row['color_ref_plural']
    relation_name='{}&{}&{}'.format(row['head'],row['relation'],row['tail'])
    template_dict[relation_name] = template
    template_plural_dict[relation_name] = template_plural


# outf_2 = open('data/llama3-8B-it_primekg_low_direct_raw.json','w')
all_texts = []

template = '''For a list medical entities {}, {}'''

KG = pd.read_csv("[PATH_OF_PRIMEKG]")

entities_dict = {}
unique_entity_types = KG['x_type'].unique()
for entity_type in unique_entity_types:
    entities_dict[entity_type] = KG[KG['x_type'] == entity_type]['x_name'].tolist()

def main(args):
    sampling_size = 20
    option_nums = 10
    max_choices = 3
    outf = open(f'data/{args.model_name}_primekg_low_indirect_multi_{option_nums}_samp_{sampling_size}_choices_{max_choices}_new_ref_inner.json','w')
    # colors = ['red', 'green', 'blue','yellow','orange','purple', 'pink','black','white','cyan']
    inner_KG_pos = json.load(open(f'data/{args.model_name}_low_score_samples_elicit_KG_merged.json','r'))
    # inner_KG_neg = json.load(open('data/llama3-8B-it_low_score_samples_elicit_KG_neg_coarse_clean.json','r'))

    # inner_rate =
    with open(f'data/{args.model_name}_low_score_samples.json','r') as f:
        all_inject_data =  json.load(f)

    tmp_relation = ''
    tmp_KG = None

    all_data = []
    all_data_dict = {}
    for item in all_inject_data:
        # item = json.loads(line.strip())
        all_data.append(item)
        if item[1] not in all_data_dict:
            all_data_dict[item[1]] = {}
        if item[0] not in all_data_dict[item[1]]:
            all_data_dict[item[1]][item[0]] = []
        all_data_dict[item[1]][item[0]].append(item[2])
    for item in tqdm(all_data):

        head_type, relation_type, tail_type = item[1].split('&')
        if tmp_relation != item[1]:
            tmp_KG = KG[(KG['x_type'] == head_type) & (KG['display_relation'] == relation_type) & (KG['y_type'] == tail_type)]
            tmp_relation = item[1]
        target_y = tmp_KG[(tmp_KG['x_name'] == item[0])]['y_name'].tolist()
        neg_y = list(set(entities_dict[tail_type]) - set(target_y))
        # if len(neg_y) < 3:
        #     continue
        # possible_y = list(set(all_data_dict[item[1]][item[0]]) - set([item[2]]))
        inner_possible_y = list(set(inner_KG_pos[item[0]][item[1]]) - set([item[2]]))
        # possible_y = inner_possible_y
        possible_y = list(set(inner_possible_y))
        available_pos = len(possible_y) + 1
        max_pos = min(available_pos, max_choices)
        cnt = 0
        # for j in range(max_pos+1):
        for _ in range(sampling_size):
            # if len(inner_possible_y) > 0:
            #     possible_y += random.sample()

            tmp_pos_num = random.randint(1, max_pos)
            if tmp_pos_num < option_nums:
                sampled_neg_y = random.sample(neg_y, option_nums-tmp_pos_num)
            else:
                sampled_neg_y = []
            if tmp_pos_num > 1:
                sampled_pos_y = random.sample(possible_y, tmp_pos_num-1)
            else:
                sampled_pos_y = []

            pos_options = [item[2]] + sampled_pos_y
            options = pos_options + sampled_neg_y
            random.shuffle(options)
            ans_idx = options.index(item[2])
            tmp = options[cnt%option_nums]
            options[cnt%option_nums] = options[ans_idx]
            options[ans_idx] = tmp

            reference = template_dict[item[1]].format(item[0])
            answer_text = []
            rationale = ''
            for i, option in enumerate(options):
                if option in pos_options:
                    answer_text.append(chr(i+ord('A')))
                    rationale += '{}: {} is a {}.\n'.format(chr(i+ord('A')),option,reference)
                else:
                    rationale += '{}: {} is not a {}.\n'.format(chr(i+ord('A')),option,reference)
            option_texts = ['{}: {}'.format(chr(i+ord('A')), op) for i, op in enumerate(options)]

            reference_plural = template_plural_dict[item[1]].format(item[0])
            # if len(answer_text) == 0:
            #     target_description = f"the {reference} is not included in the provided list."
            # else:
            target_description = f"among the given list, the {reference_plural} include option"
            answer_text = ', '.join(answer_text)
            # text = local_template.format(target_description,item[0]) + '.'

            text = template.format(', '.join(option_texts), target_description)
            label = answer_text
            outf.write(json.dumps({'text': text,'label':' '+label+'.'})+'\n')
            cnt += 1
    # random.shuffle(all_texts)
    # json.dump(all_texts, outf_2)
    # outf_2.close()
    # new_outs = []
    # for i in range(0, len(all_texts), 25):
    #     text = '<|end_of_text|>'.join(all_texts[i:i+25])
    #     new_outs.append(text)
    # out_texts = new_outs
    # for text in out_texts:
    #     outf.write(json.dumps({'text': text})+'\n')
            # outf.write(json.dumps({'text':text})+'\n')
    outf.close()

if __name__ == "__main__":
    parser = argparse.ArgumentParser()
    parser.add_argument("--ntrain", "-k", type=int, default=5)
    parser.add_argument("--nchain", "-c", type=int, default=5)
    parser.add_argument("--nbatch", "-b", type=int, default=-1)
    parser.add_argument("--typs", action="extend",nargs="+", type=str)
    parser.add_argument("--dataset",type=str,default='medqa')
    parser.add_argument("--model", type=str, default='')
    parser.add_argument("--model_name", type=str, default='llama3-8B-it')
    parser.add_argument("--debug",action='store_true')
    parser.add_argument("--start", type=int, default=0)
    parser.add_argument("--num_cuda", type=int, default=2)
    parser.add_argument("--util", type=float, default=0.98)
    parser.add_argument("--swap", type=int, default=4)
    # parser.add_argument("--max_tokens",type=int, default=2048)
    # parser.add_argument("--subjects", type=list, default=['cat_qa','isa_qa1'])
    args = parser.parse_args()
    main(args)