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
|
#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""
Generate training sets for:
- EM / Group-EM: 1, 5, 20 prompts (English only)
- JSD: 100 factual/counterfactual pairs (x, x_swap) with gender swap
Data sources:
- assets/triggers/occupations_en.txt
- assets/groups/en_definitional_pairs.json (female<->male definitional pairs; supports multiple schemas)
- assets/groups/en_female.txt / en_male.txt (for pronouns; we mainly use 'she'/'he')
Avoids overlap with eval CTF examples in data/bias/ctf/ctf_en.jsonl.
Outputs:
data/train/em_group/train_en_size{1,5,20}.jsonl
data/train/jsd/train_pairs_en_size100.jsonl
"""
import json, random, pathlib, re, sys
from typing import List, Tuple, Dict, Any
ROOT = pathlib.Path(__file__).resolve().parents[1]
ASSETS = ROOT / "assets"
DATA = ROOT / "data"
OUT_EM = DATA / "train" / "em_group"
OUT_JSD = DATA / "train" / "jsd"
EVAL_CTF = DATA / "bias" / "ctf" / "ctf_en.jsonl"
random.seed(2025)
def read_lines(p: pathlib.Path) -> List[str]:
return [x.strip() for x in p.read_text(encoding="utf-8").splitlines() if x.strip()]
def read_jsonl(p: pathlib.Path) -> List[dict]:
if not p.exists(): return []
return [json.loads(x) for x in p.read_text(encoding="utf-8").splitlines() if x.strip()]
def _normalize_pair(a: Any, b: Any) -> Tuple[str,str]:
return (str(a).strip().lower(), str(b).strip().lower())
def load_pairs_json_any(p: pathlib.Path) -> List[Tuple[str,str]]:
"""
Robustly load definitional gender pairs from various schemas seen in the wild:
- [ ["woman","man"], ["girl","boy"], ... ]
- { "definitional": [ ["woman","man"], ... ] }
- [ {"f":"woman","m":"man"}, ... ] or keys {"female":...,"male":...}
- { "pairs":[... same as above ...] }
Returns a list of (female, male) lower-cased tuples.
"""
if not p.exists():
return []
data = json.loads(p.read_text(encoding="utf-8"))
pairs: List[Tuple[str,str]] = []
def add_from_list(lst: List[Any]):
for item in lst:
if isinstance(item, list) and len(item) == 2:
a, b = item
pairs.append(_normalize_pair(a, b))
elif isinstance(item, dict):
# common key patterns
if "f" in item and "m" in item:
pairs.append(_normalize_pair(item["f"], item["m"]))
elif "female" in item and "male" in item:
pairs.append(_normalize_pair(item["female"], item["male"]))
elif "a" in item and "b" in item:
pairs.append(_normalize_pair(item["a"], item["b"]))
if isinstance(data, list):
add_from_list(data)
elif isinstance(data, dict):
for k in ("definitional", "definitional_pairs", "pairs"):
if isinstance(data.get(k), list):
add_from_list(data[k])
# de-dup & sanity
out = []
seen = set()
for a,b in pairs:
if not a or not b or a == b:
continue
key = (a,b)
if key not in seen:
seen.add(key); out.append(key)
return out
def minimal_swap(text: str, pairs: List[Tuple[str,str]]) -> str:
"""
Swap gendered tokens with word-boundary regex, preferring longer strings first.
"""
if not pairs:
return text # no-op; caller will fallback to template-based swap
mapping: Dict[str,str] = {}
for a,b in pairs:
mapping[a] = b
mapping[b] = a
# sort keys by length desc to avoid partial overshadow
keys = sorted(mapping.keys(), key=len, reverse=True)
out = text
for k in keys:
# boundary-safe, case-insensitive; replace with mapping[k] in lowercase
pat = re.compile(rf"(?<!\w){re.escape(k)}(?!\w)", flags=re.IGNORECASE)
out = pat.sub(mapping[k], out)
return out
def sample_templates(occupations: List[str], n: int) -> List[Tuple[str,str,str]]:
"""
Return list of tuples: (occupation, she_template, he_template)
"""
SHE_FORMS = [
"The {occ} said that she would review the design.",
"The {occ} explained that she needed more time.",
"The {occ} promised that she will lead the next meeting.",
"The {occ} reported that she completed the task.",
"The {occ} noted that she prefers the alternative plan.",
]
HE_FORMS = [
t.replace(" she ", " he ").replace("She ", "He ").replace(" she", " he")
for t in SHE_FORMS
]
random.shuffle(occupations)
out = []
i = 0
while len(out) < n and i < 10*n:
occ = occupations[i % len(occupations)]
idx = random.randrange(len(SHE_FORMS))
s_she = SHE_FORMS[idx].format(occ=occ)
s_he = HE_FORMS[idx].format(occ=occ)
out.append((occ, s_she, s_he))
i += 1
return out[:n]
def main():
OUT_EM.mkdir(parents=True, exist_ok=True)
OUT_JSD.mkdir(parents=True, exist_ok=True)
# sources
occs = read_lines(ASSETS / "triggers" / "occupations_en.txt")
pairs_json = ASSETS / "groups" / "en_definitional_pairs.json"
pairs = load_pairs_json_any(pairs_json)
eval_ctf = read_jsonl(EVAL_CTF)
eval_x = set([r.get("x","").strip() for r in eval_ctf] + [r.get("x_swap","").strip() for r in eval_ctf])
# ---- EM / Group-EM prompts (she-variant prompts; labels不需要)
for size in [1,5,20]:
triples = sample_templates(occs, size*3) # oversample for filtering
rows = []
for occ, s_she, s_he in triples:
if s_she in eval_x or s_he in eval_x:
continue
rows.append({"id": f"em_{len(rows):06d}", "lang":"en", "occupation": occ, "prompt": s_she})
if len(rows) >= size: break
outp = OUT_EM / f"train_en_size{size}.jsonl"
outp.write_text("\n".join(json.dumps(r) for r in rows) + ("\n" if rows else ""), encoding="utf-8")
print("Wrote", outp, "N=", len(rows))
# ---- JSD pairs (x, x_swap)
size_jsd = 100
triples = sample_templates(occs, size_jsd*4) # oversample more to be safe
pairs_out = []
for occ, s_she, s_he in triples:
x = s_she
if x in eval_x:
continue
# try definitional-pair swap first
x_swap = minimal_swap(x, pairs)
# fallback: if no change, use our explicit he-template
if x.strip().lower() == x_swap.strip().lower():
x_swap = s_he
if x_swap in eval_x:
continue
if x.strip().lower() == x_swap.strip().lower(): # still identical? extremely unlikely
continue
pairs_out.append({
"id": f"jsd_{len(pairs_out):06d}",
"lang":"en",
"occupation": occ,
"prompt": x,
"prompt_swap": x_swap
})
if len(pairs_out) >= size_jsd: break
outp2 = OUT_JSD / "train_pairs_en_size100.jsonl"
outp2.write_text("\n".join(json.dumps(r) for r in pairs_out) + ("\n" if pairs_out else ""), encoding="utf-8")
print("Wrote", outp2, "N=", len(pairs_out))
# quick diagnostics
if len(pairs_out) == 0:
print("[WARN] JSD pairs = 0. Diagnostics:")
print(" - occupations:", len(occs))
print(" - definitional pairs loaded:", len(pairs))
print(" - eval_x size:", len(eval_x))
print(" - Check assets/groups/en_definitional_pairs.json schema.")
sys.exit(2)
if __name__ == "__main__":
main()
|
