Instructions to use sledgedev/rampart-mlx with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- MLX
How to use sledgedev/rampart-mlx with MLX:
# Download the model from the Hub pip install huggingface_hub[hf_xet] huggingface-cli download --local-dir rampart-mlx sledgedev/rampart-mlx
- Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- LM Studio
- Atomic Chat
Add/update pii_rules.py
Browse files- pii_rules.py +101 -0
pii_rules.py
ADDED
|
@@ -0,0 +1,101 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""Deterministic PII layer: regex + checksum/structural validators.
|
| 2 |
+
|
| 3 |
+
Mirrors the "deterministic layer" described in Rampart's whitepaper, which is the
|
| 4 |
+
*system of record* for classes the neural model is weak on (cards, SSNs) and for
|
| 5 |
+
classes whose structure lives in punctuation (email, URL, IP). Each detector
|
| 6 |
+
returns character spans `(start, end, label)` over the ORIGINAL text, so they can
|
| 7 |
+
be unioned with the model's spans — the deterministic layer taking precedence.
|
| 8 |
+
"""
|
| 9 |
+
import re
|
| 10 |
+
from typing import List, Tuple
|
| 11 |
+
|
| 12 |
+
Span = Tuple[int, int, str]
|
| 13 |
+
|
| 14 |
+
EMAIL_RE = re.compile(r"\b[A-Za-z0-9._%+\-]+@[A-Za-z0-9.\-]+\.[A-Za-z]{2,}\b")
|
| 15 |
+
URL_RE = re.compile(r"\b(?:https?://|www\.)[^\s<>()]+", re.IGNORECASE)
|
| 16 |
+
IPV4_RE = re.compile(r"\b(?:\d{1,3}\.){3}\d{1,3}\b")
|
| 17 |
+
IPV6_RE = re.compile(r"\b(?:[A-Fa-f0-9]{1,4}:){2,7}[A-Fa-f0-9]{1,4}\b")
|
| 18 |
+
MAC_RE = re.compile(r"\b(?:[A-Fa-f0-9]{2}:){5}[A-Fa-f0-9]{2}\b")
|
| 19 |
+
SSN_RE = re.compile(r"\b(\d{3})[-\s](\d{2})[-\s](\d{4})\b")
|
| 20 |
+
# A run of 13-19 digits, allowing single space/hyphen separators (card shape).
|
| 21 |
+
CARD_RE = re.compile(r"\b(?:\d[ -]?){12,18}\d\b")
|
| 22 |
+
|
| 23 |
+
|
| 24 |
+
def _luhn_ok(digits: str) -> bool:
|
| 25 |
+
if not (13 <= len(digits) <= 19):
|
| 26 |
+
return False
|
| 27 |
+
total, alt = 0, False
|
| 28 |
+
for ch in reversed(digits):
|
| 29 |
+
d = ord(ch) - 48
|
| 30 |
+
if alt:
|
| 31 |
+
d *= 2
|
| 32 |
+
if d > 9:
|
| 33 |
+
d -= 9
|
| 34 |
+
total += d
|
| 35 |
+
alt = not alt
|
| 36 |
+
return total % 10 == 0
|
| 37 |
+
|
| 38 |
+
|
| 39 |
+
def _ssn_ok(area: str, group: str, serial: str) -> bool:
|
| 40 |
+
a = int(area)
|
| 41 |
+
if a == 0 or a == 666 or a >= 900: # reserved / invalid SSA areas
|
| 42 |
+
return False
|
| 43 |
+
if int(group) == 0 or int(serial) == 0:
|
| 44 |
+
return False
|
| 45 |
+
return True
|
| 46 |
+
|
| 47 |
+
|
| 48 |
+
def _ipv4_ok(s: str) -> bool:
|
| 49 |
+
parts = s.split(".")
|
| 50 |
+
return len(parts) == 4 and all(p.isdigit() and 0 <= int(p) <= 255 for p in parts)
|
| 51 |
+
|
| 52 |
+
|
| 53 |
+
def detect_rules(text: str) -> List[Span]:
|
| 54 |
+
spans: List[Span] = []
|
| 55 |
+
|
| 56 |
+
for m in EMAIL_RE.finditer(text):
|
| 57 |
+
spans.append((m.start(), m.end(), "EMAIL"))
|
| 58 |
+
for m in URL_RE.finditer(text):
|
| 59 |
+
spans.append((m.start(), m.end(), "URL"))
|
| 60 |
+
for m in MAC_RE.finditer(text):
|
| 61 |
+
spans.append((m.start(), m.end(), "IP_ADDRESS"))
|
| 62 |
+
for m in IPV6_RE.finditer(text):
|
| 63 |
+
spans.append((m.start(), m.end(), "IP_ADDRESS"))
|
| 64 |
+
for m in IPV4_RE.finditer(text):
|
| 65 |
+
if _ipv4_ok(m.group(0)):
|
| 66 |
+
spans.append((m.start(), m.end(), "IP_ADDRESS"))
|
| 67 |
+
for m in SSN_RE.finditer(text):
|
| 68 |
+
if _ssn_ok(m.group(1), m.group(2), m.group(3)):
|
| 69 |
+
spans.append((m.start(), m.end(), "SSN"))
|
| 70 |
+
for m in CARD_RE.finditer(text):
|
| 71 |
+
digits = re.sub(r"\D", "", m.group(0))
|
| 72 |
+
if _luhn_ok(digits):
|
| 73 |
+
spans.append((m.start(), m.end(), "CREDIT_CARD"))
|
| 74 |
+
|
| 75 |
+
return spans
|
| 76 |
+
|
| 77 |
+
|
| 78 |
+
def _overlaps(a: Span, b: Span) -> bool:
|
| 79 |
+
return a[0] < b[1] and b[0] < a[1]
|
| 80 |
+
|
| 81 |
+
|
| 82 |
+
def union(model_spans: List[Span], rule_spans: List[Span]) -> List[Span]:
|
| 83 |
+
"""Union the two layers; the deterministic layer wins on any overlap."""
|
| 84 |
+
final = list(rule_spans)
|
| 85 |
+
for ms in model_spans:
|
| 86 |
+
if not any(_overlaps(ms, rs) for rs in rule_spans):
|
| 87 |
+
final.append(ms)
|
| 88 |
+
final.sort(key=lambda s: s[0])
|
| 89 |
+
return final
|
| 90 |
+
|
| 91 |
+
|
| 92 |
+
def redact(text: str, spans: List[Span]) -> str:
|
| 93 |
+
out, last = [], 0
|
| 94 |
+
for s, e, label in sorted(spans, key=lambda x: x[0]):
|
| 95 |
+
if s < last:
|
| 96 |
+
continue
|
| 97 |
+
out.append(text[last:s])
|
| 98 |
+
out.append(f"[{label}]")
|
| 99 |
+
last = e
|
| 100 |
+
out.append(text[last:])
|
| 101 |
+
return "".join(out)
|