henry-yeh's picture htr234234's picture
Chiron data release
1b86964
Raw History Blame Contribute Delete
3.81 kB
"""WVS (World Values Survey) runtime adapter.
Data construction lives in the dataset repo (``chiron-group/wvs`` / local mirror
``data/raw/wvs/scripts/``); this module is runtime-only.
"""
from __future__ import annotations
import random
import re
from pathlib import Path
from domain_adapter.base import DomainAdapter, Sample
MIN_OBSERVATIONS = 50
MAX_OBSERVATIONS = 100
_ADMIN_RE = re.compile(
r"\((?=[^)]*(?:read out|read and code|code|interviewer|not applicable))[^)]*\)",
re.IGNORECASE,
)
_ADMIN_TAIL_RE = re.compile(
r"\b(?:interviewer|do not read|code\s+[-\"'\w ]*not applicable)\b.*$",
re.IGNORECASE,
)
_DANGLING_PROMPT_TAIL_RE = re.compile(
r"([?.!])\s+(?:Are you|Would you say it is)(?=\s*(?:\||$))",
re.IGNORECASE,
)
def _clean_wvs_text(text: str) -> str:
"""Remove WVS interviewer/routing artifacts without rewriting item meaning."""
text = " ".join(str(text).strip().split())
text = _ADMIN_RE.sub("", text)
text = _ADMIN_TAIL_RE.sub("", text)
text = _DANGLING_PROMPT_TAIL_RE.sub(r"\1", text)
text = re.sub(r"\s+", " ", text)
return text.strip()
def _clean_wvs_observation(observation: str) -> str:
if " | Response:" not in observation:
return _clean_wvs_text(observation)
item, response = observation.split(" | Response:", 1)
item = _clean_wvs_text(item)
response = _clean_wvs_text(response)
if item and response:
return f"{item} | Response: {response}"
return item or response
def _clean_wvs_question(question: str) -> str:
lines: list[str] = []
for raw_line in question.strip().splitlines():
line = raw_line.strip()
if not line:
continue
if line.startswith("Survey question:"):
body = _clean_wvs_text(line.removeprefix("Survey question:"))
if body:
lines.append(f"Survey question: {body}")
elif line.startswith("Question supplement:"):
body = _clean_wvs_text(line.removeprefix("Question supplement:"))
if body:
lines.append(f"Question supplement: {body}")
else:
lines.append(_clean_wvs_text(line))
return "\n".join(line for line in lines if line)
class WVSAdapter(DomainAdapter):
name = "wvs"
def data_source(self) -> str:
return "indievalue_wvs"
def get_sample_id(self, sample: Sample) -> str:
return str(sample.extra_info["interview_id"])
def default_input_path(self) -> Path:
return Path("data/raw/wvs/qa_jsonl/train.jsonl")
def record_context(self) -> str:
return "Below are survey response records describing this respondent's values, beliefs, and background."
def format_observations_block(self, observations: list[str]) -> str:
rendered: list[str] = []
for observation in observations:
text = _clean_wvs_observation(observation)
if not text or re.fullmatch(r"-\d+(?:\.0)?", text):
continue
rendered.append(f"- {text}")
return "\n".join(rendered)
def format_question_block(self, questions: list[str]) -> str:
return "\n\n".join(
f"Q{i}.\n{_clean_wvs_question(question)}"
for i, question in enumerate(questions, start=1)
)
def select_observations(
self,
sample: Sample,
*,
seed: int,
sample_index: int,
) -> list[str]:
observations = list(sample.observations)
if len(observations) <= MIN_OBSERVATIONS:
return observations
rng = random.Random(seed ^ (sample_index * 9173) ^ 0x575653)
rng.shuffle(observations)
n_observations = rng.randint(MIN_OBSERVATIONS, min(MAX_OBSERVATIONS, len(observations)))
return observations[:n_observations]