File size: 2,454 Bytes
76f8341 | 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 | """Self-refusal detection (regex-based).
Used as the post-gate in approach E: after the LLM produces a response, scan
the first ~300 chars for explicit-refusal patterns. If found, the model is
admitting it doesn't have the requested fact -> we either retrieve from
external corpus or surface the refusal honestly.
The patterns target what R1-distill / Qwen-family chat models actually emit
when they encounter post-cutoff or fictional-entity questions under a strict
system prompt:
"I am sorry, I cannot answer that"
"I don't have information on..."
"That is after my training cutoff"
"As of my last update..."
"I'm not sure..."
Tested refusal-trigger rate (4 out-of-corpus unknowns):
V0 loose prompt: 25%
V1 strict prompt: 50%
V2 cutoff-aware prompt: 75% <- recommended (see RECOMMENDED_SYSTEM)
V3 combined (strict + cutoff): 50% (over-specified, worse)
Residual failure: topic-hijack cases (e.g. "Who won the 2024 Nobel Physics?"
gets rewritten to a fabricated 1986/1998 Nobel answer). Prompt engineering
alone cannot fully suppress this. Mitigations: short-latency heuristic,
topic-blocklist, answer-verification pass.
"""
import re
_REFUSAL_RE = re.compile(
r"\b(I (am sorry|don'?t|do not|am unable|cannot|can'?t)\s+"
r"(know|have|recall|remember|answer|provide|find|see))"
r"|\b(I'?m not (sure|aware|certain))"
r"|\b(no (information|details|data) (on|about|regarding))"
r"|\b(as of my (training|last update|knowledge)|my training data|training cutoff)"
r"|\bcannot answer that\b"
r"|\bdon'?t have (information|details|data)\b"
r"|\bafter my (training|knowledge) cutoff\b"
r"|\bI lack (information|data)\b",
re.IGNORECASE,
)
def detect_refusal(text: str, scan_chars: int = 400) -> bool:
"""Return True if `text[:scan_chars]` contains an explicit self-refusal phrase."""
if not text:
return False
return _REFUSAL_RE.search(text[:scan_chars]) is not None
# The recommended system prompt for production: maximizes refusal rate on
# out-of-corpus unknowns while staying short and not over-specified.
RECOMMENDED_SYSTEM = (
"You are a helpful assistant. Your knowledge cutoff is March 2024. "
"For any question about events, people, or facts after March 2024, "
"you MUST explicitly say 'That is after my training cutoff.' "
"Do not guess. Do not fabricate. "
"If a Context is provided, use it to answer."
)
|