circuit-steering-data / sanitize.py
MdSadiqMd's picture
Super-squash branch 'main' using huggingface_hub
ff811bb
Raw History Blame Contribute Delete
4.26 kB
"""Keep generated replies to harmful prompts out of the repo and the Hub (brief §8: "store only
prompts, judgments and aggregate scores").
Two places hold such text:
data/b2/baseline_<cat>.jsonl `gen.<lang>.preview` (160 characters of the reply) on harmful rows
runs/audit/**/hf.json `outputs` of the refusal_harmful_* and refusal_other_* families
The judgments (`refused`, per-sample `scores`), first tokens and every prompt are kept; replies to
benign prompts are kept too (they are the over-refusal evidence and contain nothing harmful).
uv run python data/sanitize.py baseline copy the raw baseline files to .cache/private/ (local,
gitignored) and strip harmful previews in place
uv run python data/sanitize.py scan [paths...] list every leak under the given paths
`stage_audit` builds the sanitized copy of runs/audit that data/push_hf.py publishes.
"""
from __future__ import annotations
import json
import os
import shutil
import sys
from pathlib import Path
ROOT = Path(__file__).resolve().parent.parent
HARMFUL_FAMILIES = ("refusal_harmful_", "refusal_other_")
PRIVATE = ROOT / ".cache/private/b2_baseline_raw"
def strip_audit(hf: dict) -> dict:
return {
k: (
{kk: vv for kk, vv in v.items() if kk != "outputs"}
if k.startswith(HARMFUL_FAMILIES)
else v
)
for k, v in hf.items()
}
def strip_baseline_row(r: dict) -> dict:
if r.get("split") != "harmful":
return r
gen = {
lang: {k: v for k, v in g.items() if k != "preview"} for lang, g in r.get("gen", {}).items()
}
return {**r, "gen": gen}
def leaks(path: Path) -> list[str]:
"""Generated text for harmful prompts inside one file (hf.json or a baseline jsonl)."""
out = []
if path.name == "hf.json":
d = json.loads(path.read_text())
out += [
f"{path}:{k}"
for k, v in d.items()
if k.startswith(HARMFUL_FAMILIES) and v.get("outputs")
]
elif path.name.startswith("baseline_") and path.suffix == ".jsonl":
for i, line in enumerate(path.open()):
r = json.loads(line)
if r.get("split") == "harmful" and any(
"preview" in g for g in r.get("gen", {}).values()
):
out.append(f"{path}:{i + 1}:{r['id']}")
return out
def scan(paths: list[Path]) -> list[str]:
found = []
for p in paths:
files = [p] if p.is_file() else [*p.rglob("hf.json"), *p.rglob("baseline_*.jsonl")]
for f in files:
found += leaks(f)
return found
def sanitize_baseline() -> None:
PRIVATE.mkdir(parents=True, exist_ok=True)
for p in sorted((ROOT / "data/b2").glob("baseline_*.jsonl")):
raw = PRIVATE / p.name
if not raw.exists():
shutil.copy2(p, raw) # the pre-registered file, byte for byte, kept local only
rows = [strip_baseline_row(json.loads(line)) for line in raw.open()]
p.write_text("".join(json.dumps(r, ensure_ascii=False) + "\n" for r in rows))
print(p, "sanitized; raw copy at", raw)
def stage_audit(src: Path, dst: Path) -> Path:
"""Mirror `src` (runs/audit) into `dst`: hf.json sanitized, every other file hard-linked."""
for f in src.rglob("*"):
if f.is_dir():
continue
t = dst / f.relative_to(src)
t.parent.mkdir(parents=True, exist_ok=True)
if t.exists():
t.unlink()
if f.name == "hf.json":
t.write_text(json.dumps(strip_audit(json.loads(f.read_text())), default=str))
else:
os.link(f, t)
assert not scan([dst]), "sanitized staging still holds harmful-prompt generations"
return dst
if __name__ == "__main__":
cmd = sys.argv[1] if len(sys.argv) > 1 else "scan"
if cmd == "baseline":
sanitize_baseline()
elif cmd == "scan":
targets = [Path(a) for a in sys.argv[2:]] or [ROOT / "data", ROOT / "docs", ROOT / "space"]
found = scan(targets)
print("\n".join(found[:20]), f"\n{len(found)} leak(s)" if found else "no leaks")
raise SystemExit(1 if found else 0)
else:
raise SystemExit(f"unknown command {cmd!r}")