beliefsim / scripts /prepare_release.py
Angana192's picture
Upload folder using huggingface_hub
e75cb1c verified
#!/usr/bin/env python3
"""Prepare an anonymized BeliefSim dataset package for Hugging Face.
The script keeps claims, participant judgments, evaluation triplets, and WVS
group priors. It drops direct identifiers, IP/location metadata, and open-text
free-response notes.
"""
import hashlib
import json
import re
import zipfile
from pathlib import Path
from xml.etree import ElementTree as ET
import pandas as pd
ROOT = Path(__file__).resolve().parents[2]
OUT = Path(__file__).resolve().parents[1] / "data"
SALT = "beliefsim-release-v1"
PANDORA_PATH = ROOT / "misinfo_combined_latest_with_acc_new.csv"
MIST_PATH = ROOT / "wvs_misinfo" / "MIST - Sample 1 - Raw Dataset.csv"
MIST_ITEM_DB = ROOT / "wvs_misinfo" / "MIST - Phase 4 - Item Database (January 2020).xlsx"
WVS_FILES = {
"gender": ROOT / "wvs_misinfo" / "gender_alldimensions_wvs.csv",
"living_area": ROOT / "wvs_misinfo" / "urbrur_alldimensions_wvs.csv",
"age": ROOT / "wvs_misinfo" / "age_alldimensions.csv",
"education": ROOT / "wvs_misinfo" / "education_alldimensions.csv",
}
def anonymize(prefix: str, *parts: object) -> str:
raw = "|".join("" if p is None else str(p) for p in parts)
digest = hashlib.sha256(f"{SALT}|{prefix}|{raw}".encode("utf-8")).hexdigest()[:16]
return f"{prefix}_{digest}"
def clean_text(value: object) -> str:
if pd.isna(value):
return ""
return re.sub(r"\s+", " ", str(value)).strip()
def norm(value: object) -> str:
return re.sub(r"[^a-z0-9]+", " ", clean_text(value).lower()).strip()
def first_nonempty(row: pd.Series, candidates: list) -> str:
for col in candidates:
if col in row.index:
val = clean_text(row[col])
if val and val.lower() != "nan":
return val
return ""
def extract_claim(prompt_text: str) -> str:
text = re.sub(r"^\s*News:\s*", "", clean_text(prompt_text), flags=re.I)
for marker in ["Supporting Stance:", "Refuting Stance:", "Your responses:", "Your response:"]:
if marker in text:
text = text.split(marker, 1)[0]
return clean_text(text)
def standardize_pandora_judgment(value: object) -> str:
v = norm(value)
if not v:
return ""
if "importid" in v or v in {"true information", "misinformation", "have you heard of the information before", "comments notes"}:
return ""
if "true information" in v or v in {"true", "real"}:
return "true_information"
if "misinformation" in v or v in {"false", "fake"}:
return "misinformation"
if "not sure" in v or "unsure" in v:
return "not_sure"
return clean_text(value)
def standardize_mist_judgment(value: object) -> str:
v = norm(value)
if not v:
return ""
if v in {"real", "true", "1"} or "real" in v:
return "true_information"
if v in {"fake", "false", "0"} or "fake" in v:
return "misinformation"
return clean_text(value)
def age_bucket(value: object) -> str:
s = clean_text(value)
if not s:
return ""
m = re.search(r"\d+", s)
if not m:
return s
age = int(m.group())
if age < 30:
return "18-29"
if age < 45:
return "30-44"
if age < 60:
return "45-59"
return "60+"
def education_bucket(value: object) -> str:
v = norm(value)
if not v:
return ""
completed = [
"bachelor",
"master",
"doctor",
"professional",
"college degree",
"university",
"graduate",
]
if any(token in v for token in completed):
return "completed"
return "not_completed"
def load_pandora() -> tuple:
if not PANDORA_PATH.exists():
return [], [], []
raw = pd.read_csv(PANDORA_PATH, header=None, dtype=str, low_memory=False)
qids = [clean_text(x) for x in raw.iloc[0].tolist()]
labels = [clean_text(x) for x in raw.iloc[1].tolist()]
columns = []
seen: dict = {}
qid_map: dict = {}
for qid, label in zip(qids, labels):
base = label or qid or "unnamed"
seen[base] = seen.get(base, 0) + 1
col = base if seen[base] == 1 else f"{base}__{seen[base]}"
columns.append(col)
if qid:
qid_map[col] = qid
# Qualtrics export rows: machine ids, question text, choice/import rows, then responses.
df = raw.iloc[4:].copy()
df.columns = columns
news_cols = [c for c in df.columns if norm(c).startswith("news")]
claims: dict[str, dict] = {}
judgments: list = []
instances: list = []
# In this Qualtrics export the answer to "Your responses" is stored in the
# same column as the long claim prompt; the next two columns are heard-before
# and free-text notes, which are intentionally not released.
response_cols = {c: c for c in news_cols}
demo_cols = {
"gender": ["What is your gender?", "gender"],
"age_group": ["How old are you?", "age"],
"living_area": ["How would you describe the area you live in?", "area", "living_area"],
"education_bucket": [
"What is the highest level of education you have completed?",
"degree",
"education",
],
}
id_candidates = ["ResponseId", "Response ID", "PROLIFIC_ID"]
for row_idx, row in df.iterrows():
row_judgments = []
participant_id = anonymize("pandora", row_idx, first_nonempty(row, id_candidates))
demographics = {}
for key, candidates in demo_cols.items():
val = first_nonempty(row, candidates)
demographics[key] = education_bucket(val) if key == "education_bucket" else clean_text(val)
if demographics.get("age_group"):
demographics["age_group"] = age_bucket(demographics["age_group"])
for news_col in news_cols:
prompt_text = clean_text(row.get(news_col, ""))
resp_col = response_cols.get(news_col, "")
judgment = standardize_pandora_judgment(row.get(resp_col, "")) if resp_col else ""
claim_text = extract_claim(prompt_text)
if not claim_text or not judgment:
continue
claim_id = qid_map.get(news_col, "") or anonymize("pandora_claim", news_col)
claims[claim_id] = {
"source": "PANDORA",
"claim_id": claim_id,
"claim_text": claim_text,
"gold_label": "",
"content_warning": True,
}
record = {
"source": "PANDORA",
"participant_id": participant_id,
"claim_id": claim_id,
"judgment": judgment,
**demographics,
}
judgments.append(record)
row_judgments.append(record)
for i in range(2, len(row_judgments)):
b1, b2, target = row_judgments[i - 2], row_judgments[i - 1], row_judgments[i]
instances.append(
{
"source": "PANDORA",
"instance_id": anonymize("pandora_instance", participant_id, target["claim_id"], i),
"participant_id": participant_id,
"target_claim_id": target["claim_id"],
"target_judgment": target["judgment"],
"observed_claim_1_id": b1["claim_id"],
"observed_judgment_1": b1["judgment"],
"observed_claim_2_id": b2["claim_id"],
"observed_judgment_2": b2["judgment"],
**demographics,
}
)
return list(claims.values()), judgments, instances
def read_xlsx_first_sheet(path):
"""Read a simple XLSX first sheet using only the Python standard library."""
ns = {"a": "http://schemas.openxmlformats.org/spreadsheetml/2006/main"}
with zipfile.ZipFile(str(path)) as zf:
shared = []
if "xl/sharedStrings.xml" in zf.namelist():
root = ET.fromstring(zf.read("xl/sharedStrings.xml"))
for item in root.findall(".//a:si", ns):
shared.append("".join(t.text or "" for t in item.findall(".//a:t", ns)))
sheet_name = "xl/worksheets/sheet1.xml"
root = ET.fromstring(zf.read(sheet_name))
rows = []
for row in root.findall(".//a:sheetData/a:row", ns):
values = []
for cell in row.findall("a:c", ns):
cell_type = cell.attrib.get("t", "")
value_node = cell.find("a:v", ns)
inline_node = cell.find("a:is/a:t", ns)
value = ""
if inline_node is not None:
value = inline_node.text or ""
elif value_node is not None:
raw = value_node.text or ""
value = shared[int(raw)] if cell_type == "s" and raw.isdigit() and int(raw) < len(shared) else raw
values.append(value)
rows.append(values)
if not rows:
return pd.DataFrame()
width = max(len(r) for r in rows)
rows = [r + [""] * (width - len(r)) for r in rows]
header = [clean_text(x) or "column_{}".format(i) for i, x in enumerate(rows[0])]
return pd.DataFrame(rows[1:], columns=header)
def load_mist_item_db() -> pd.DataFrame:
try:
db = pd.read_excel(MIST_ITEM_DB, dtype=str)
except ImportError:
db = read_xlsx_first_sheet(MIST_ITEM_DB)
cols = {norm(c): c for c in db.columns}
id_col = cols.get("id") or next((c for c in db.columns if norm(c) == "id"), "")
headline_col = (
cols.get("headline")
or cols.get("item")
or next((c for c in db.columns if "headline" in norm(c) or "item" == norm(c)), "")
)
out = db[[id_col, headline_col]].copy()
out.columns = ["item_id", "claim_text"]
out["gold_label"] = out["item_id"].map(lambda x: "true_information" if clean_text(x).upper().startswith("R") else "misinformation")
out["claim_id"] = [f"MIST_{i + 1}" for i in range(len(out))]
return out
def load_mist() -> tuple:
if not MIST_PATH.exists() or not MIST_ITEM_DB.exists():
return [], [], []
df = pd.read_csv(MIST_PATH, dtype=str, low_memory=False)
item_db = load_mist_item_db()
claims = [
{
"source": "MIST-1",
"claim_id": row.claim_id,
"claim_text": clean_text(row.claim_text),
"gold_label": row.gold_label,
"content_warning": True,
}
for row in item_db.itertuples(index=False)
]
claim_ids = item_db["claim_id"].tolist()
mist_cols = [c for c in df.columns if re.fullmatch(r"MIST_\d+", clean_text(c))]
if not mist_cols:
mist_cols = [c for c in df.columns if clean_text(c) in claim_ids]
judgments: list = []
instances: list = []
gender_cols = [c for c in df.columns if norm(c) in {"gender", "sex"} or "gender" in norm(c)]
age_cols = [c for c in df.columns if norm(c) == "age" or "age" in norm(c)]
edu_cols = [c for c in df.columns if "education" in norm(c) or "degree" in norm(c)]
for row_idx, row in df.iterrows():
demographics = {
"gender": first_nonempty(row, gender_cols),
"age_group": age_bucket(first_nonempty(row, age_cols)),
"living_area": "",
"education_bucket": education_bucket(first_nonempty(row, edu_cols)),
}
participant_id = anonymize("mist", row_idx)
row_judgments = []
for col in mist_cols:
judgment = standardize_mist_judgment(row.get(col, ""))
if not judgment:
continue
rec = {
"source": "MIST-1",
"participant_id": participant_id,
"claim_id": clean_text(col),
"judgment": judgment,
**demographics,
}
judgments.append(rec)
row_judgments.append(rec)
for i in range(2, len(row_judgments)):
b1, b2, target = row_judgments[i - 2], row_judgments[i - 1], row_judgments[i]
instances.append(
{
"source": "MIST-1",
"instance_id": anonymize("mist_instance", participant_id, target["claim_id"], i),
"participant_id": participant_id,
"target_claim_id": target["claim_id"],
"target_judgment": target["judgment"],
"observed_claim_1_id": b1["claim_id"],
"observed_judgment_1": b1["judgment"],
"observed_claim_2_id": b2["claim_id"],
"observed_judgment_2": b2["judgment"],
**demographics,
}
)
return claims, judgments, instances
def load_wvs_priors() -> list:
records: list = []
for axis, path in WVS_FILES.items():
if not path.exists():
continue
df = pd.read_csv(path, dtype=str)
q_col = next((c for c in df.columns if norm(c) in {"question", "question text"}), "")
id_col = next((c for c in df.columns if norm(c) in {"question no", "question_no", "qid"}), "")
for _, row in df.iterrows():
for col in df.columns:
n = norm(col)
if not n.endswith("distribution") or n == "overall distribution":
continue
group = re.sub(r"_?distribution$", "", col).strip("_")
most_col = f"{group}_most"
least_col = f"{group}_least"
records.append(
{
"demographic_axis": axis,
"group": group,
"question_id": clean_text(row.get(id_col, "")) if id_col else "",
"question_text": clean_text(row.get(q_col, "")) if q_col else "",
"distribution": clean_text(row.get(col, "")),
"most_common": clean_text(row.get(most_col, "")),
"least_common": clean_text(row.get(least_col, "")),
"source_file": path.name,
}
)
return records
def main() -> None:
OUT.mkdir(parents=True, exist_ok=True)
p_claims, p_judgments, p_instances = load_pandora()
m_claims, m_judgments, m_instances = load_mist()
wvs = load_wvs_priors()
claims = pd.DataFrame(p_claims + m_claims).drop_duplicates(["source", "claim_id"])
judgments = pd.DataFrame(p_judgments + m_judgments)
instances = pd.DataFrame(p_instances + m_instances)
priors = pd.DataFrame(wvs)
claims.to_csv(OUT / "claims.csv", index=False)
judgments.to_csv(OUT / "judgments.csv", index=False)
instances.to_csv(OUT / "evaluation_instances.csv", index=False)
priors.to_csv(OUT / "wvs_group_priors.csv", index=False)
summary = {
"claims": len(claims),
"judgments": len(judgments),
"evaluation_instances": len(instances),
"wvs_group_prior_rows": len(priors),
"sources": {
"PANDORA": {
"claims": int((claims["source"] == "PANDORA").sum()) if not claims.empty else 0,
"judgments": int((judgments["source"] == "PANDORA").sum()) if not judgments.empty else 0,
"evaluation_instances": int((instances["source"] == "PANDORA").sum()) if not instances.empty else 0,
},
"MIST-1": {
"claims": int((claims["source"] == "MIST-1").sum()) if not claims.empty else 0,
"judgments": int((judgments["source"] == "MIST-1").sum()) if not judgments.empty else 0,
"evaluation_instances": int((instances["source"] == "MIST-1").sum()) if not instances.empty else 0,
},
},
"privacy": {
"dropped": [
"ResponseId",
"PROLIFIC_ID",
"IPAddress",
"LocationLatitude",
"LocationLongitude",
"open-text notes/comments",
"raw timestamps",
],
"participant_ids": "deterministic salted SHA-256 hashes, truncated to 16 hex characters",
"demographics": "broad gender, age bucket, education bucket, and living-area fields only",
},
}
(OUT / "dataset_summary.json").write_text(json.dumps(summary, indent=2))
print(json.dumps(summary, indent=2))
if __name__ == "__main__":
main()