-
Notifications
You must be signed in to change notification settings - Fork 0
/
pred_to_preference.py
172 lines (145 loc) · 7.74 KB
/
pred_to_preference.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
import json
import nltk
import argparse
import numpy as np
def eval_score(inst, alpha=1, beta=1, gamma=1):
idx2score = {}
for score in inst['prediction_scores']:
rouge = alpha * 100 * (score['rouge1_f1'] + score['rouge2_f1'] + score['rougel_f1'])
similarity = beta * 100 * (score['bleurt'] + score['bert_score_f1'])
factuality = gamma * (score['comprehensive'] - score['hallucination'])
total = rouge + similarity + factuality
idx2score[score['idx']] = total
return {key:value for key,value in sorted(idx2score.items(), key=lambda item: item[1], reverse=True)}
def make_inst(model_name, chosen_idx, reject_idx, inst):
if model_name == "selfbiorag_7b":
chosen = "[No Retrieval] " + inst["sample_predictions"][chosen_idx]
rejected = "[No Retrieval] " + inst["sample_predictions"][reject_idx]
else:
chosen = inst["sample_predictions"][chosen_idx]
rejected = inst["sample_predictions"][reject_idx]
if "::" in chosen or "::" in rejected:
return None
elif "....." in chosen or "....." in rejected:
return None
elif "?" == chosen:
return None
else:
if chosen == "" or rejected == "":
return None
new_inst = {"chosen": chosen, "rejected": rejected, "prompt": inst["Question"]}
return new_inst
def main():
parser = argparse.ArgumentParser()
parser.add_argument('--model_name_or_path', type=str, default="dmis-lab/selfbiorag_7b")
parser.add_argument('--wodata_name', type=str, default="kqa_golden")
parser.add_argument('--data_names', nargs='+', default="live_qa medication_qa healthsearch_qa kqa_silver_wogold kqa_golden")
parser.add_argument('--alpha', type=int, default=1)
parser.add_argument('--beta', type=int, default=1)
parser.add_argument('--gamma', type=int, default=1)
parser.add_argument('--threshold', type=int, default=200)
args = parser.parse_args()
data_names = args.data_names.split(" ")
examples = []
if "selfbiorag" in args.model_name_or_path.lower():
model_name = "selfbiorag-7b"
elif "biomistral" in args.model_name_or_path.lower():
model_name = "biomistral-7b"
elif "mistral" in args.model_name_or_path.lower():
model_name = "mistral-7b"
elif "llama" in args.model_name_or_path.lower():
model_name = "llama2-7b"
elif "meditron" in args.model_name_or_path.lower():
model_name = "meditron-7b"
else:
model_name = args.model_name_or_path.split("/")[1]
wodata_name = args.wodata_name
for data_name in data_names:
if len(data_names) == 1:
target_data_name = data_name
else:
if data_name == wodata_name:
continue
with open(f"./predictions/pdata_{model_name}-wo-{wodata_name}-iter-dpo-step_{data_name}_iter-dpo-step.jsonl", "r") as fp:
for line in fp.readlines():
examples.append(json.loads(line))
new_examples = []
for inst_idx,inst in enumerate(examples):
# check scores to make lists of preference/dispreference
if "none" in inst["Question"].lower() or inst["Question"] == None:
continue
r1, r2, rl, bl, bs, hl, cp = [], [], [], [], [], [], []
tw, ss, f = [], [], []
all = []
if inst["Question"][-1] != "?":
inst["Question"] += "?"
idx2score = eval_score(inst, alpha=args.alpha, beta=args.beta, gamma=args.gamma)
idx_list = list(idx2score.keys())
new_sample_predictions = []
new_prediction_scores = []
for prediction_idx, prediction in enumerate(inst['sample_predictions']):
if inst["Question"] in prediction:
prediction = prediction.split(inst["Question"])[0]
sent_text = nltk.sent_tokenize(prediction)
prediction = " ".join(sent_text[:-1])
if "Question: " in prediction:
prediction = prediction.split("Question: ")[0].strip()
if prediction == "?":
continue
new_sample_predictions.append(prediction)
new_prediction_scores.append(inst['prediction_scores'][prediction_idx])
inst['sample_predictions'] = new_sample_predictions
inst['prediction_scores'] = new_prediction_scores
for pred_idx, pred_score in enumerate(inst['prediction_scores']):
if inst["sample_predictions"] == "?":
continue
r1.append(pred_score['rouge1_f1'] * 100)
r2.append(pred_score['rouge2_f1'] * 100)
rl.append(pred_score['rougel_f1'] * 100)
bl.append(pred_score['bleurt'] * 100)
bs.append(pred_score['bert_score_f1'] * 100)
hl.append(pred_score['hallucination'])
cp.append(pred_score['comprehensive'])
tw.append(r1[-1] + r2[-1] + rl[-1])
ss.append(bl[-1] + bs[-1])
f.append(cp[-1] - hl[-1])
all.append(tw[-1] + ss[-1] + f[-1])
if all[idx_list[0]] < args.threshold:
continue
try:
if not "?" in inst['sample_predictions'][idx_list[0]]:
temp = 20
if all[idx_list[0]] - all[idx_list[2]] >= temp:
if make_inst(model_name, chosen_idx=idx_list[0], reject_idx=idx_list[2], inst=inst):
new_examples.append(make_inst(model_name, chosen_idx=idx_list[0], reject_idx=idx_list[2], inst=inst))
"""top 1"""
if inst['sample_predictions'][idx_list[1]] != inst['sample_predictions'][idx_list[2]]:
if all[idx_list[0]] - all[idx_list[1]] >= temp:
if make_inst(model_name, chosen_idx=idx_list[0], reject_idx=idx_list[1], inst=inst):
new_examples.append(make_inst(model_name, chosen_idx=idx_list[0], reject_idx=idx_list[1], inst=inst))
if inst['sample_predictions'][idx_list[3]] != inst['sample_predictions'][idx_list[2]]:
if all[idx_list[0]] - all[idx_list[3]] >= temp:
if make_inst(model_name, chosen_idx=idx_list[0], reject_idx=idx_list[3], inst=inst):
new_examples.append(make_inst(model_name, chosen_idx=idx_list[0], reject_idx=idx_list[3], inst=inst))
if inst['sample_predictions'][idx_list[4]] != inst['sample_predictions'][idx_list[2]]:
if all[idx_list[0]] - all[idx_list[4]] >= temp:
if make_inst(model_name, chosen_idx=idx_list[0], reject_idx=idx_list[4], inst=inst):
new_examples.append(make_inst(model_name, chosen_idx=idx_list[0], reject_idx=idx_list[4], inst=inst))
if inst['sample_predictions'][idx_list[5]] != inst['sample_predictions'][idx_list[2]]:
if all[idx_list[0]] - all[idx_list[5]] >= temp:
if make_inst(model_name, chosen_idx=idx_list[0], reject_idx=idx_list[5], inst=inst):
new_examples.append(make_inst(model_name, chosen_idx=idx_list[0], reject_idx=idx_list[5], inst=inst))
except:
continue
if len(data_names) == 1:
with open(f"./predictions/preference_{model_name}_test_{target_data_name}_iter-dpo-step.jsonl", "w") as out_:
for inst in new_examples:
out_.write(json.dumps(inst))
out_.write("\n")
else:
with open(f"./predictions/preference_{model_name}_test_all_wo_{wodata_name}_iter-dpo-step.jsonl", "w") as out_:
for inst in new_examples:
out_.write(json.dumps(inst))
out_.write("\n")
if __name__ == "__main__":
main()