MdSadiqMd's picture
Super-squash branch 'main' using huggingface_hub
ff811bb
Raw History Blame Contribute Delete
4.56 kB
"""Refusal EAP task per language from the baseline generations.
Clean = harmful prompt the model refused (all three candidate categories, so the refusal
*mechanism* is
discovered on ~200 prompts rather than the 34 cyber ones; the B2 intervention target stays the cyber
category); corrupt = benign near-miss prompt the model answered, with the same token count
(CircuitKIT pairs need equal n_pos). Labels: correct_idx = the model's own most
frequent refusal onset token in that language, incorrect_idx = its most frequent compliance onset
(both read from data/b2/baseline_summary.json, so the metric is the model's actual refuse-vs-comply
logit difference at the assistant marker, as in CircuitKIT case study 23). Falls back to pairing
harmful with the closest-length benign prompt and right-padding when no exact length match exists.
Output: data/b2/eap_<lang>.csv plus data/b2/eap_stats.json. Category comes from configs
(behaviour.b2_category) once chosen in docs/preregistration.md."""
from __future__ import annotations
import csv
import json
import random
import sys
from collections import defaultdict
from pathlib import Path
sys.path.insert(0, str(Path(__file__).resolve().parent.parent))
from aya_circuits.config import load_config
from aya_circuits.model import chat_wrap
from transformers import AutoTokenizer
HERE = Path(__file__).resolve().parent / "b2"
def main() -> None:
cfg = load_config()
tok = AutoTokenizer.from_pretrained(cfg["model"])
rows, seen = [], set()
for p in sorted(HERE.glob("baseline_*.jsonl")):
for line in p.open():
r = json.loads(line)
if r["id"] not in seen:
seen.add(r["id"])
rows.append(r)
onsets = json.loads((HERE / "baseline_summary.json").read_text())["onsets"]
rng, stats = random.Random(0), {}
for lang in cfg["languages"]:
def pick(kind: str, lang: str = lang) -> str: # skip whitespace-only onsets
return next(t for t, _ in onsets[lang][kind] if t.strip())
ref_tok = tok.encode(pick("refusal"), add_special_tokens=False)[0]
cmp_tok = tok.encode(pick("comply"), add_special_tokens=False)[0]
harm = [
r for r in rows if r["split"] == "harmful" and lang in r and r["gen"][lang]["refused"]
]
ben = [
r
for r in rows
if r["split"] == "benign" and lang in r and not r["gen"][lang]["refused"]
]
if not harm or not ben:
stats[lang] = {"pairs": 0, "note": "no refused-harmful or complied-benign prompts"}
continue
def wrap(r: dict, lang: str = lang) -> str:
return chat_wrap(tok, r[lang], cfg["chat_format"])
def ntok(r: dict, wrap=wrap) -> int:
return len(tok(wrap(r), add_special_tokens=False)["input_ids"])
# drop length outliers (> 64 tokens, 1-4 per language): a single 250-token prompt in a
# batch of 4 blew
# the 30 GB budget in EAP-IG (activation buffer scales with the longest sequence in the
# batch)
harm, ben = [r for r in harm if ntok(r) <= 64], [r for r in ben if ntok(r) <= 64]
blen = defaultdict(list)
for b in ben:
blen[len(tok(wrap(b), add_special_tokens=False)["input_ids"])].append(b)
pairs, exact = [], 0
for h in harm:
n = len(tok(wrap(h), add_special_tokens=False)["input_ids"])
if blen.get(n):
b, exact = rng.choice(blen[n]), exact + 1
else:
b = rng.choice(blen[min(blen, key=lambda k: abs(k - n))])
pairs.append(
{
"clean": wrap(h),
"corrupted": wrap(b),
"correct_idx": ref_tok,
"incorrect_idx": cmp_tok,
}
)
with (HERE / f"eap_{lang}.csv").open("w", newline="") as f:
w = csv.DictWriter(f, fieldnames=["clean", "corrupted", "correct_idx", "incorrect_idx"])
w.writeheader()
w.writerows(pairs)
stats[lang] = {
"pairs": len(pairs),
"exact_length_pairs": exact,
"refusal_onset": pick("refusal"),
"comply_onset": pick("comply"),
"n_harmful_refused": len(harm),
"n_benign_complied": len(ben),
}
(HERE / "eap_stats.json").write_text(json.dumps(stats, indent=2, ensure_ascii=False))
print(json.dumps(stats, indent=1, ensure_ascii=False))
if __name__ == "__main__":
main()