Initial public commit

Co-authored-by: mlabonne <mlabonne@users.noreply.huggingface.co>
Co-authored-by: claude opus 4.8 (1m context) <claude opus 4.8 (1m context)@users.noreply.huggingface.co>

.dockerignore

@@ -0,0 +1,6 @@
+.git
+.gitattributes
+__pycache__/
+*.pyc
+.venv/
+*.md

.gitattributes

@@ -0,0 +1,35 @@
+*.7z filter=lfs diff=lfs merge=lfs -text
+*.arrow filter=lfs diff=lfs merge=lfs -text
+*.bin filter=lfs diff=lfs merge=lfs -text
+*.bz2 filter=lfs diff=lfs merge=lfs -text
+*.ckpt filter=lfs diff=lfs merge=lfs -text
+*.ftz filter=lfs diff=lfs merge=lfs -text
+*.gz filter=lfs diff=lfs merge=lfs -text
+*.h5 filter=lfs diff=lfs merge=lfs -text
+*.joblib filter=lfs diff=lfs merge=lfs -text
+*.lfs.* filter=lfs diff=lfs merge=lfs -text
+*.mlmodel filter=lfs diff=lfs merge=lfs -text
+*.model filter=lfs diff=lfs merge=lfs -text
+*.msgpack filter=lfs diff=lfs merge=lfs -text
+*.npy filter=lfs diff=lfs merge=lfs -text
+*.npz filter=lfs diff=lfs merge=lfs -text
+*.onnx filter=lfs diff=lfs merge=lfs -text
+*.ot filter=lfs diff=lfs merge=lfs -text
+*.parquet filter=lfs diff=lfs merge=lfs -text
+*.pb filter=lfs diff=lfs merge=lfs -text
+*.pickle filter=lfs diff=lfs merge=lfs -text
+*.pkl filter=lfs diff=lfs merge=lfs -text
+*.pt filter=lfs diff=lfs merge=lfs -text
+*.pth filter=lfs diff=lfs merge=lfs -text
+*.rar filter=lfs diff=lfs merge=lfs -text
+*.safetensors filter=lfs diff=lfs merge=lfs -text
+saved_model/**/* filter=lfs diff=lfs merge=lfs -text
+*.tar.* filter=lfs diff=lfs merge=lfs -text
+*.tar filter=lfs diff=lfs merge=lfs -text
+*.tflite filter=lfs diff=lfs merge=lfs -text
+*.tgz filter=lfs diff=lfs merge=lfs -text
+*.wasm filter=lfs diff=lfs merge=lfs -text
+*.xz filter=lfs diff=lfs merge=lfs -text
+*.zip filter=lfs diff=lfs merge=lfs -text
+*.zst filter=lfs diff=lfs merge=lfs -text
+*tfevents* filter=lfs diff=lfs merge=lfs -text

Dockerfile

@@ -0,0 +1,30 @@
+FROM python:3.12-slim
+
+# non-root user (HF Spaces convention; gives a writable HOME for the HF cache the
+# model is pulled into at boot).
+RUN useradd -m -u 1000 user
+USER user
+ENV HOME=/home/user \
+    PATH=/home/user/.local/bin:$PATH \
+    HF_HOME=/home/user/.cache/huggingface \
+    PYTHONUNBUFFERED=1 \
+    TMPDIR=/tmp
+
+WORKDIR /home/user/app
+
+# CPU-only torch (the default PyPI wheel is CUDA and huge); transformers 5.11 is the
+# stack pii-detect's bidirectional LFM2 remote code was validated against.
+RUN pip install --no-cache-dir --user torch==2.12.0 --index-url https://download.pytorch.org/whl/cpu
+
+COPY --chown=user requirements.txt .
+RUN pip install --no-cache-dir --user -r requirements.txt
+
+COPY --chown=user . .
+
+EXPOSE 7860
+
+# One worker (single shared model), threads for concurrent static/detect conns.
+# No --preload: the worker pulls + loads the model in its own process (avoids the
+# torch/OpenMP fork-after-threads deadlock). --timeout 0 keeps the ~1.4GB first-boot
+# download from tripping the worker timeout.
+CMD ["gunicorn", "-w", "1", "--threads", "8", "-k", "gthread", "-b", "0.0.0.0:7860", "--timeout", "0", "app:app"]

README.md

@@ -0,0 +1,28 @@
+---
+title: PII Detection
+emoji: 🕵️
+colorFrom: purple
+colorTo: blue
+sdk: docker
+app_port: 7860
+header: mini
+pinned: false
+private: true
+short_description: Multilingual PII detection with LFM2.5, server-side CPU
+---
+
+# PII Detection
+
+Server-side PII detection powered by [`LiquidAI/LFM2.5-Encoder-350M-PII-Detector`](https://huggingface.co/LiquidAI/LFM2.5-Encoder-350M-PII-Detector): a bidirectional LFM2.5 encoder (354M) with a token-classification head, covering 40 PII types across 16 languages. It runs on the Space's CPU; the browser only sends text to `/api/detect` and highlights the returned spans, so nothing is downloaded to the visitor.
+
+## How it works
+
+- `pii.py` — loads the model via its custom modeling code (`trust_remote_code`) and runs the model's own inference-time **hybrid regex decode** (`pii_hybrid_decode.py` + `context_cued.py`, fetched from the model repo): the token classifier locates PII, a validator-gated regex layer owns format-bound types (email/IBAN/credit-card/JWT/…), and cue-gated rules recover structure-less IDs. Spans are enriched with a domain (color), a pretty label, and a redaction tag.
+- `app.py` — Flask server: loads the model once, serves the page, and answers `POST /api/detect {text}` with `{spans, redacted, count, ms}`.
+- `index.html` / `style.css` — the UI: highlight or redacted view, a per-domain legend, and detection stats.
+
+f32 on CPU (not fp16: x86 has no fast half path; not int8: token-classification boundaries are precision-sensitive) — the same reasoning as the spellchecker / policy-linting Spaces. Threads are matched to the container's CPU quota.
+
+## Setup
+
+The model repo is private, so the Space needs an **`HF_TOKEN`** secret with read access to `LiquidAI/LFM2.5-Encoder-350M-PII-Detector`.

app.py

@@ -0,0 +1,67 @@
+"""PII detection — server-side CPU inference.
+
+LiquidAI/LFM2.5-Encoder-350M-PII-Detector (LFM2.5 encoder + token-classification head + hybrid regex
+decode) runs on the Space's CPU. The browser POSTs {text} to /api/detect and
+gets back char-anchored PII spans (type/domain/label) + a redacted string, then
+highlights them. Nothing is downloaded to the visitor.
+
+The model repo is private, so an HF_TOKEN Space secret with read access is
+required to pull it.
+"""
+
+import os
+import time
+
+import torch
+from flask import Flask, request, send_from_directory
+
+import pii as P
+
+APP_DIR = os.path.dirname(os.path.abspath(__file__))
+# Match threads to the container's cgroup CPU quota, not the host core count.
+_CPUS = P.effective_cpus()
+torch.set_num_threads(_CPUS)
+try:
+    torch.set_num_interop_threads(1)
+except RuntimeError:
+    pass
+
+app = Flask(__name__)
+
+print(f"[boot] loading pii-detect on {_CPUS} CPU thread(s), f32…", flush=True)
+MODEL = P.PiiDetector()
+MODEL.detect("Warm up: email jane@example.com, SSN 123-45-6789.")   # warm the graph
+print("[boot] detector ready", flush=True)
+
+
+@app.route("/")
+def index():
+    return send_from_directory(APP_DIR, "index.html")
+
+
+@app.route("/<path:name>")
+def static_file(name):
+    if name in ("style.css", "lliquid.gif"):
+        return send_from_directory(APP_DIR, name)
+    return ("not found", 404)
+
+
+@app.route("/api/ready")
+def ready():
+    return {"ready": MODEL is not None}
+
+
+@app.route("/api/detect", methods=["POST"])
+def api_detect():
+    data = request.get_json(silent=True) or {}
+    text = (data.get("text") or "").replace("\r", "")
+    if not text.strip():
+        return {"spans": [], "redacted": "", "count": 0, "ms": 0}
+    t0 = time.perf_counter()
+    out = MODEL.detect(text)
+    out["ms"] = round((time.perf_counter() - t0) * 1000)
+    return out
+
+
+if __name__ == "__main__":
+    app.run(host="0.0.0.0", port=int(os.environ.get("PORT", 7860)), threaded=True)

index.html

@@ -0,0 +1,240 @@
+<!doctype html>
+<html lang="en">
+<head>
+<meta charset="utf-8">
+<meta name="viewport" content="width=device-width, initial-scale=1">
+<title>PII detection — Liquid AI</title>
+<link rel="preconnect" href="https://fonts.googleapis.com">
+<link rel="preconnect" href="https://fonts.gstatic.com" crossorigin>
+<link href="https://fonts.googleapis.com/css2?family=Instrument+Serif:ital@0;1&family=Inter:wght@400;500;600;700&family=JetBrains+Mono:wght@400;500;600&display=swap" rel="stylesheet">
+<link rel="stylesheet" href="style.css">
+<script>
+  // Private Space auth: HF signs the embedded iframe URL with ?__sign=<token>.
+  // Incognito / Safari block the third-party .hf.space cookie, so that token is
+  // the only credential — carry it on same-origin API requests, else /api/*
+  // returns 404 and the model-ready poll hangs.
+  (() => {
+    const sign = new URLSearchParams(location.search).get("__sign");
+    if (!sign) return;
+    const orig = self.fetch.bind(self);
+    self.fetch = (input, init) => {
+      try {
+        const u = new URL(typeof input === "string" ? input : input.url, location.href);
+        if (u.origin === location.origin && !u.searchParams.has("__sign")) {
+          u.searchParams.set("__sign", sign);
+          input = u.toString();
+        }
+      } catch (e) { /* non-URL input: pass through */ }
+      return orig(input, init);
+    };
+  })();
+</script>
+</head>
+<body>
+<div class="wrap">
+<header class="site-intro">
+  <h1>Multilingual <em>PII detection</em></h1>
+  <p class="sub">Spot <strong>40 kinds of personal data</strong> across 16 languages and redact them, in one bidirectional pass.</p>
+</header>
+
+<section class="model-strip is-loading" id="model-strip" aria-label="Model status">
+  <span class="model-logo-frame" aria-hidden="true">
+    <img id="model-logo" class="model-logo" width="56" height="59" alt="" src="lliquid.gif">
+  </span>
+  <div class="model-copy">
+    <div class="model-name-row">
+      <a href="https://huggingface.co/LiquidAI/LFM2.5-Encoder-350M-PII-Detector" target="_blank" rel="noopener">LFM2.5 Encoder</a>
+      <span class="model-variant">350M parameters</span>
+    </div>
+  </div>
+  <div class="load-block">
+    <div class="load-meta">
+      <span id="status" role="status" aria-live="polite">Connecting to server…</span>
+      <span id="load-percent">0%</span>
+    </div>
+    <div class="load-track" role="progressbar" aria-label="Model loading" aria-valuemin="0" aria-valuemax="100" aria-valuenow="0"><span id="load-progress"></span></div>
+  </div>
+</section>
+
+<main>
+<div class="pii-workspace">
+  <section class="workspace-card input-panel">
+    <div class="panel-header">
+      <div><h2>Text</h2></div>
+    </div>
+    <div class="editor-stage">
+      <textarea id="input" spellcheck="false" placeholder="Paste text containing names, emails, IDs, cards…">Update the account for david.miller@example.com: phone +1 415 555 0199, SSN 123-45-6789, card 4111 1111 1111 1111, passport C01X00T47.</textarea>
+    </div>
+  </section>
+
+  <section class="workspace-card output-panel">
+    <div class="panel-header">
+      <div><h2>Detected PII</h2></div>
+      <div class="view-toggle" id="view-toggle" role="tablist" aria-label="Output view">
+        <button type="button" data-view="highlight" class="active" role="tab">Highlight</button>
+        <button type="button" data-view="redacted" role="tab">Redacted</button>
+      </div>
+    </div>
+    <div class="output-stage">
+      <div id="output" class="output-body"><div class="output-empty">Detected personal data will appear here.</div></div>
+    </div>
+    <div class="controls">
+      <div class="legend" id="legend"></div>
+    </div>
+  </section>
+</div>
+
+<section class="examples-panel" aria-label="Examples">
+  <span class="examples-label">Examples</span>
+  <div class="example-list" id="examples">
+    <button type="button" class="ex active" data-t="Update the account for david.miller@example.com: phone +1 415 555 0199, SSN 123-45-6789, card 4111 1111 1111 1111, passport C01X00T47.">Contact & ID</button>
+    <button type="button" class="ex" data-t="Please wire EUR 5,000 to IBAN DE89 3704 0044 0532 0130 00 (SWIFT COBADEFFXXX) from account 0532013000. My server is at 192.168.1.42.">Financial</button>
+    <button type="button" class="ex" data-t="Deploy creds: AWS key AKIAIOSFODNN7EXAMPLE, session token eyJhbGciOiJIUzI1NiJ9.eyJzdWIiOiIxMjM0NTY3In0.dozjgNryP4J3jVmNHl0w, and DB password Str0ngP@ssw0rd!.">Credentials</button>
+    <button type="button" class="ex" data-t="Company laptop MAC 00:1A:2B:3C:4D:5E connected from IP 10.0.0.5 near GPS 37.7749, -122.4194.">Device</button>
+    <button type="button" class="ex" data-t="Send 0.5 BTC to bc1qxy2kgdygjrsqtzq2n0yrf2493p83kkfjhx0wlh or ETH to 0x742d35Cc6634C0532925a3b844Bc454e4438f44e. Rotate the API key sk-proj-T3BlbkFJ4aBcDeFgHiJkLmNoP.">Crypto & keys</button>
+    <button type="button" class="ex" data-t="Follow username jsmith on https://social.example.com/jsmith or email j.smith@example.com for details.">Online</button>
+    <button type="button" class="ex" data-t="Patient Sarah Connor, date of birth 03/14/1985, record number 887766, diagnosed with Type 2 diabetes and prescribed Metformin.">Healthcare</button>
+    <button type="button" class="ex" data-t="Ich heiße Hans Müller und wohne in der Hauptstraße 5, 80331 München. Erreichbar unter +49 30 4505 1234 oder hans.mueller@example.de.">Deutsch</button>
+    <button type="button" class="ex" data-t="Je m'appelle Jean Dupont, email jean.dupont@example.fr, téléphone +33 6 12 34 56 78, IBAN FR76 3000 6000 0112 3456 7890 189.">Français</button>
+    <button type="button" class="ex" data-t="María García, correo maria.garcia@example.es, teléfono +34 612 345 678, tarjeta 4111 1111 1111 1111.">Español</button>
+    <button type="button" class="ex" data-t="Иван Петров, email ivan.petrov@example.ru, телефон +7 495 123 45 67.">Русский</button>
+    <button type="button" class="ex" data-t="我叫张伟，电话号码是13800138000，邮箱是 zhang.wei@example.cn，住在北京市朝阳区建国路88号。">中文</button>
+    <button type="button" class="ex" data-t="田中太郎と申します。メールアドレスは taro.tanaka@example.jp、電話番号は 090-1234-5678 です。">日本語</button>
+    <button type="button" class="ex" data-t="제 이름은 김민수입니다. 이메일은 minsu.kim@example.kr, 전화번호는 010-1234-5678입니다.">한국어</button>
+    <button type="button" class="ex" data-t="اسمي محمد أحمد، بريدي الإلكتروني mohammed.ahmed@example.com، هاتفي +20 10 1234 5678.">العربية</button>
+  </div>
+</section>
+</main>
+</div>
+
+<script>
+window.addEventListener("error", e => {
+  const el = document.getElementById("status");
+  if (el) el.textContent = "· ERROR: " + (e.message || (e.target && (e.target.src || e.target.href)) || "resource failed");
+}, true);
+</script>
+<script type="module">
+// The detector runs on the server (CPU, f32). This script drives the UI and asks
+// /api/detect for char-anchored PII spans; highlighting, the redacted view, the
+// legend, and stats are done here. Nothing is downloaded to the client.
+
+const DOMAINS = ["identity","contact","financial","credential","online","device","location","healthcare","org","special","legal","other"];
+const $ = (id) => document.getElementById(id);
+const escapeHtml = (v) => v.replace(/&/g, "&amp;").replace(/</g, "&lt;").replace(/>/g, "&gt;");
+const cap = (s) => s.charAt(0).toUpperCase() + s.slice(1);
+
+function setStatus(message, tone = "loading") {
+  $("status").textContent = message;
+  const strip = $("model-strip");
+  strip.classList.remove("is-loading", "is-preparing", "is-ready", "is-error");
+  strip.classList.add(`is-${tone}`);
+}
+function setLoadProgress(value) {
+  const pct = Math.max(0, Math.min(100, Number(value) || 0));
+  $("load-progress").style.width = `${pct}%`;
+  $("load-percent").textContent = `${Math.round(pct)}%`;
+  document.querySelector(".load-track").setAttribute("aria-valuenow", String(Math.round(pct)));
+}
+let READY = false, timer = null, version = 0;
+let lastText = "", lastResult = null, view = "highlight";
+
+function renderHighlight(text, spans) {
+  let html = "", prev = 0, i = 0;
+  for (const sp of spans) {
+    html += escapeHtml(text.slice(prev, sp.start));
+    const t = escapeHtml(text.slice(sp.start, sp.end));
+    const tip = escapeHtml(`${sp.label} · ${sp.type}`);
+    html += `<span class="pii is-animated" style="--pii-color:var(--d-${sp.domain});--i:${i++}" title="${tip}">${t}</span>`;
+    prev = sp.end;
+  }
+  return html + escapeHtml(text.slice(prev));
+}
+function renderRedacted(redacted) {
+  return escapeHtml(redacted).replace(/\[([A-Z0-9_]+)\]/g, '<span class="redacted-tag">[$1]</span>');
+}
+function renderLegend(spans) {
+  const counts = {};
+  for (const sp of spans) counts[sp.domain] = (counts[sp.domain] || 0) + 1;
+  $("legend").innerHTML = DOMAINS.filter(d => counts[d]).map(d =>
+    `<span class="legend-chip" style="--pii-color:var(--d-${d})"><span class="legend-swatch" aria-hidden="true"></span>${cap(d)}<span class="legend-count">${counts[d]}</span></span>`
+  ).join("");
+}
+function paint() {
+  const out = $("output");
+  if (!lastResult || !lastText.trim()) {
+    out.className = "output-body";
+    out.innerHTML = '<div class="output-empty">Detected personal data will appear here.</div>';
+    return;
+  }
+  out.className = view === "redacted" ? "output-body is-redacted" : "output-body";
+  out.innerHTML = view === "redacted" ? renderRedacted(lastResult.redacted) : renderHighlight(lastText, lastResult.spans);
+}
+
+async function detect() {
+  if (!READY) return;
+  const runVersion = ++version;
+  const text = $("input").value;
+  if (!text.trim()) { lastText = text; lastResult = null; paint(); $("legend").innerHTML = ""; return; }
+  document.querySelector(".output-panel").classList.add("is-detecting");
+  let data;
+  try {
+    const resp = await fetch("api/detect", {
+      method: "POST", headers: { "Content-Type": "application/json" },
+      body: JSON.stringify({ text }),
+    });
+    if (!resp.ok) throw new Error(`HTTP ${resp.status}`);
+    data = await resp.json();
+  } catch (e) {
+    document.querySelector(".output-panel").classList.remove("is-detecting");
+    return;
+  }
+  if (runVersion !== version) return;
+  lastText = text; lastResult = data;
+  paint();
+  renderLegend(data.spans);
+  document.querySelector(".output-panel").classList.remove("is-detecting");
+}
+
+function detectSoon(delay = 350) { clearTimeout(timer); timer = setTimeout(() => detect(), delay); }
+
+$("input").addEventListener("input", () => detectSoon());
+$("view-toggle").addEventListener("click", (e) => {
+  const b = e.target.closest("button"); if (!b) return;
+  view = b.dataset.view;
+  $("view-toggle").querySelectorAll("button").forEach(x => x.classList.toggle("active", x === b));
+  paint();
+});
+$("examples").addEventListener("click", (e) => {
+  const b = e.target.closest(".ex"); if (!b) return;
+  $("examples").querySelectorAll(".ex").forEach(x => x.classList.toggle("active", x === b));
+  $("input").value = b.dataset.t;
+  detectSoon(40);
+});
+
+async function waitForServer() {
+  setStatus("Connecting to server…", "loading");
+  setLoadProgress(12);
+  for (let attempt = 0; attempt < 900; attempt++) {
+    try {
+      const r = await fetch("api/ready", { cache: "no-store" });
+      if (r.ok && (await r.json()).ready) return true;
+    } catch (e) { /* server still starting */ }
+    setStatus("Warming up model on the server…", "preparing");
+    setLoadProgress(Math.min(92, 12 + attempt * 4));
+    await new Promise((r) => setTimeout(r, 1000));
+  }
+  return false;
+}
+
+(async () => {
+  const ok = await waitForServer();
+  if (!ok) { setStatus("Model unavailable", "error"); setLoadProgress(100); $("load-percent").textContent = "Error"; return; }
+  READY = true;
+  setLoadProgress(100);
+  setStatus("Model ready", "ready");
+  await detect();
+  window.__PII_READY = true;
+})();
+</script>
+</body>
+</html>

pii.py

@@ -0,0 +1,119 @@
+"""pii.py — server-side CPU PII detection with LiquidAI/LFM2.5-Encoder-350M-PII-Detector.
+
+A bidirectional LFM2.5 phase-2 encoder (354M) with a token-classification head,
+loaded from the Hub via its custom modeling code (trust_remote_code), plus the
+model's own inference-time **hybrid regex decode** (pii_hybrid_decode.py +
+context_cued.py, fetched from the same repo) — the decode the model is meant to
+use. It covers 40 PII types across 11 domains and 16 languages.
+
+f32 on CPU (not fp16: x86 has no fast half path; not int8: token-classification
+boundaries are precision-sensitive) — the same reasoning as the spellchecker /
+policy-linting Spaces. One forward per request. Nothing is downloaded to the
+client; the model repo is private, so an HF_TOKEN with read access is required.
+"""
+
+import os
+import sys
+import threading
+
+from huggingface_hub import hf_hub_download
+from transformers import AutoModelForTokenClassification, AutoTokenizer
+
+MODEL_ID = os.environ.get("PII_MODEL", "LiquidAI/LFM2.5-Encoder-350M-PII-Detector")
+HF_TOKEN = os.environ.get("HF_TOKEN")
+
+
+def effective_cpus():
+    """Cores actually available to this process (cgroup quota, not the host)."""
+    try:  # cgroup v2
+        raw = open("/sys/fs/cgroup/cpu.max").read().split()
+        if raw and raw[0] != "max":
+            return max(1, round(int(raw[0]) / int(raw[1])))
+    except (OSError, ValueError):
+        pass
+    try:  # cgroup v1
+        q = int(open("/sys/fs/cgroup/cpu/cpu.cfs_quota_us").read())
+        p = int(open("/sys/fs/cgroup/cpu/cpu.cfs_period_us").read())
+        if q > 0:
+            return max(1, q // p)
+    except (OSError, ValueError):
+        pass
+    try:
+        return max(1, len(os.sched_getaffinity(0)))
+    except AttributeError:
+        return os.cpu_count() or 1
+
+
+# --- label -> {domain (color), pretty label, redaction tag} ------------------
+_ACRONYM = {"ssn", "iban", "id", "ip", "url", "api", "mac", "gps", "bic", "dob",
+            "pin", "vat", "cvv", "swift", "tax", "jwt", "imei"}
+# Types are `domain.subtype` (label_schema.json), so the color bucket is just the
+# prefix. `developer.*` is the exception: login_credentials -> credential, device_id
+# -> device.
+_DOMAINS = {"identity", "contact", "financial", "credential", "online", "device",
+            "location", "healthcare", "org", "special", "legal"}
+
+
+def domain_for(raw):
+    pre = raw.split(".")[0].lower()
+    if pre in _DOMAINS:
+        return pre
+    if pre == "developer":
+        return "device" if "device" in raw.lower() else "credential"
+    return "other"
+
+
+def pretty(raw):
+    base = raw.split(".")[-1]
+    words = base.replace("-", " ").replace("_", " ").split()
+    out = [w.upper() if w.lower() in _ACRONYM else w.capitalize() for w in words]
+    return " ".join(out) or raw
+
+
+def tag_for(raw):
+    base = raw.split(".")[-1]
+    return "_".join(base.replace("-", " ").replace("_", " ").split()).upper() or "PII"
+
+
+class PiiDetector:
+    def __init__(self):
+        self.lock = threading.Lock()
+        self.tok = AutoTokenizer.from_pretrained(MODEL_ID, trust_remote_code=True, token=HF_TOKEN)
+        self.model = AutoModelForTokenClassification.from_pretrained(
+            MODEL_ID, trust_remote_code=True, token=HF_TOKEN).float().eval()
+        # the model's shipped hybrid decode (+ its context_cued dependency); load
+        # both from the repo snapshot and import off that dir.
+        hd_path = hf_hub_download(MODEL_ID, "pii_hybrid_decode.py", token=HF_TOKEN)
+        hf_hub_download(MODEL_ID, "context_cued.py", token=HF_TOKEN)
+        sys.path.insert(0, os.path.dirname(hd_path))
+        import pii_hybrid_decode as hd
+        self.hd = hd
+
+    def detect(self, text):
+        """Return {spans, redacted, count}. spans: char-anchored, enriched, sorted,
+        non-overlapping. When a model span and a regex span cover the same text, the
+        regex (AUTH) type wins — it's the precise format (e.g. api_key/jwt over the
+        model's generic login_credentials)."""
+        with self.lock:
+            raw = self.hd.predict(text, self.tok, self.model)   # [{start,end,type,text}]
+
+        auth = getattr(self.hd, "_AUTH_TYPES", set())
+        raw.sort(key=lambda s: (s["start"], 0 if s["type"] in auth else 1, -(s["end"] - s["start"])))
+        spans, last_end = [], -1
+        for sp in raw:
+            if sp["start"] < last_end or sp["end"] <= sp["start"]:
+                continue                                        # drop overlaps
+            typ = sp["type"]
+            spans.append({
+                "start": sp["start"], "end": sp["end"], "text": text[sp["start"]:sp["end"]],
+                "type": typ, "domain": domain_for(typ), "label": pretty(typ), "tag": tag_for(typ),
+            })
+            last_end = sp["end"]
+
+        parts, prev = [], 0
+        for sp in spans:
+            parts.append(text[prev:sp["start"]])
+            parts.append(f"[{sp['tag']}]")
+            prev = sp["end"]
+        parts.append(text[prev:])
+        return {"spans": spans, "redacted": "".join(parts), "count": len(spans)}

requirements.txt

@@ -0,0 +1,6 @@
+transformers==5.11.0
+tokenizers==0.22.2
+safetensors==0.8.0
+huggingface_hub==1.19.0
+flask>=3.0
+gunicorn>=21.2

style.css

@@ -0,0 +1,161 @@
+:root {
+  --paper:#fff;
+  --ink:#050505;
+  --purple:#6c4fe0;
+  --lilac:#cd82f0;
+  --green:#2dbd70;
+  --red:#e45454;
+  --neutral-50:#fafafa;
+  --neutral-100:#f4f4f3;
+  --neutral-200:#e7e7e5;
+  --neutral-300:#b9b9b5;
+  --neutral-500:#737373;
+  --neutral-700:#404040;
+  --border:rgba(0,0,0,.09);
+  --sans:"Inter",-apple-system,BlinkMacSystemFont,"Segoe UI",sans-serif;
+  --serif:"Instrument Serif",Georgia,"Times New Roman",serif;
+  --mono:"JetBrains Mono",ui-monospace,SFMono-Regular,Menlo,monospace;
+  /* PII domain colors (one bucket per legend entry) */
+  --d-identity:#d8d0ff; --d-contact:#c9e9ff; --d-financial:#ccefdc; --d-credential:#ffd5e5;
+  --d-online:#cce8e8; --d-device:#ffe4ba; --d-location:#fff1ad; --d-healthcare:#ffd9d0;
+  --d-org:#e6dcc8; --d-special:#f0d0ff; --d-legal:#d5dae8; --d-other:#e7e7e5;
+}
+
+* { box-sizing:border-box; }
+html { background:var(--neutral-50); }
+body {
+  min-width:320px;
+  min-height:100vh;
+  margin:0;
+  background:var(--neutral-50);
+  color:var(--ink);
+  font:16px/1.45 var(--sans);
+  -webkit-font-smoothing:antialiased;
+}
+button,textarea,input { font:inherit; }
+a { color:inherit; }
+
+.wrap {
+  display:flex;
+  flex-direction:column;
+  width:min(calc(100% - 32px),1240px);
+  min-height:100vh;
+  min-height:100dvh;
+  margin:0 auto;
+  border-right:1px solid var(--border);
+  border-left:1px solid var(--border);
+  background:var(--paper);
+}
+
+.site-intro { margin-top:auto; padding:20px 30px 16px; }
+h1 { margin:0 0 7px; font:400 clamp(40px,4.3vw,51px)/.98 var(--serif); letter-spacing:.002em; }
+h1 em { color:var(--purple); font-weight:400; font-style:italic; }
+.sub { max-width:900px; margin:0; color:var(--neutral-500); font-size:15px; line-height:1.45; }
+.sub strong { color:var(--neutral-700); font-weight:600; }
+
+.model-strip {
+  display:grid;
+  grid-template-columns:48px 280px minmax(240px,1fr);
+  align-items:center;
+  column-gap:16px;
+  min-height:76px;
+  padding:12px 30px;
+  border-top:1px solid var(--border);
+  border-bottom:1px solid var(--border);
+  background:linear-gradient(90deg,rgba(108,79,224,.018),rgba(205,130,240,.035),rgba(108,79,224,.018));
+  transition:background .3s ease;
+}
+.model-logo-frame { position:relative; display:block; width:48px; height:48px; overflow:hidden; }
+.model-logo { position:absolute; top:50%; left:50%; display:block; width:56px; height:auto; max-width:none; transform:translate(-50%,-50%); image-rendering:pixelated; }
+.model-copy { display:flex; min-width:0; align-items:center; justify-content:center; }
+.model-name-row { display:flex; min-height:25px; align-items:center; gap:10px; white-space:nowrap; }
+.model-copy a { font:400 22px/1.15 var(--serif); text-decoration:none; }
+.model-copy a:hover { color:var(--purple); }
+.model-variant { display:inline-flex; height:18px; align-items:center; box-sizing:border-box; padding:0 7px; border:1px solid rgba(108,79,224,.14); border-radius:999px; background:rgba(108,79,224,.05); color:var(--purple); font:700 9px/1 var(--mono); letter-spacing:.035em; text-transform:uppercase; }
+.load-block { min-width:0; }
+.load-meta { display:flex; align-items:center; justify-content:space-between; gap:14px; margin-bottom:8px; color:var(--neutral-500); font-size:12px; font-weight:600; }
+#load-percent { font:600 11px/1 var(--mono); font-variant-numeric:tabular-nums; }
+.load-track { height:7px; overflow:hidden; border-radius:999px; background:rgba(108,79,224,.09); }
+.load-track span { display:block; width:0; height:100%; border-radius:inherit; background:linear-gradient(90deg,var(--purple),var(--lilac)); transition:width .18s ease; }
+.is-ready .load-track { background:rgba(45,189,112,.09); }
+.is-ready .load-track span { background:var(--green); }
+.is-ready .load-meta { color:#188e50; }
+.is-error .load-track span { background:var(--red); }
+.is-error .load-meta { color:var(--red); }
+
+main { display:flex; min-height:0; flex:0 0 auto; flex-direction:column; margin-bottom:auto; padding:18px 30px 22px; }
+.pii-workspace { display:grid; grid-template-columns:minmax(360px,1fr) minmax(420px,1.05fr); align-items:stretch; gap:18px; }
+.workspace-card { position:relative; display:flex; min-width:0; flex-direction:column; overflow:hidden; border:1px solid var(--border); border-radius:12px; background:var(--paper); padding:20px; }
+.panel-header { display:flex; align-items:flex-start; justify-content:space-between; gap:16px; margin-bottom:15px; }
+.panel-header h2 { margin:0 0 5px; color:var(--neutral-500); font:600 12px/1.25 var(--mono); letter-spacing:.08em; text-transform:uppercase; }
+.panel-header p { margin:0; color:var(--neutral-500); font-size:13px; line-height:1.4; }
+
+.editor-stage { position:relative; flex:1; min-height:clamp(230px,26dvh,320px); overflow:hidden; border:1px solid var(--border); border-radius:10px; background:var(--paper); transition:border-color .2s,box-shadow .2s; }
+.editor-stage:focus-within { border-color:var(--purple); box-shadow:0 0 0 3px rgba(108,79,224,.11); }
+#input { width:100%; height:100%; min-height:clamp(230px,26dvh,320px); padding:16px 17px; border:0; outline:0; background:transparent; color:var(--ink); font-size:16px; line-height:1.62; resize:none; }
+#input::placeholder { color:var(--neutral-300); }
+
+.output-stage { position:relative; flex:1; min-height:clamp(230px,26dvh,320px); overflow:auto; border:1px solid var(--border); border-radius:10px; background:var(--paper); }
+.output-body { padding:16px 17px; font-size:16px; line-height:1.68; white-space:pre-wrap; word-break:break-word; }
+.output-body.is-redacted { font-family:var(--mono); font-size:14px; line-height:1.7; color:var(--neutral-700); }
+.output-empty { padding:16px 17px; color:var(--neutral-300); }
+.is-detecting .output-stage { border-color:rgba(108,79,224,.28); box-shadow:0 0 0 3px rgba(108,79,224,.06); }
+
+/* marker-pen highlight, colored per PII domain */
+.pii { position:relative; z-index:0; display:inline-block; color:inherit; isolation:isolate; cursor:default; }
+.pii::before {
+  content:""; position:absolute; z-index:-1; top:.13em; right:-.07em; bottom:.09em; left:-.07em;
+  border-radius:3px 5px 2px 4px;
+  background:
+    radial-gradient(ellipse at 16% 40%,rgba(255,255,255,.38) 0 7%,transparent 8% 100%),
+    repeating-linear-gradient(176deg,
+      color-mix(in srgb,var(--pii-color) 80%,white) 0 2px,
+      color-mix(in srgb,var(--pii-color) 92%,transparent) 2px 5px,
+      color-mix(in srgb,var(--pii-color) 74%,white) 5px 8px);
+  box-shadow:inset 0 1px rgba(255,255,255,.34);
+  clip-path:polygon(0 8%,3% 2%,27% 6%,52% 1%,78% 5%,100% 0,98% 94%,73% 98%,43% 95%,17% 100%,1% 94%);
+  mix-blend-mode:multiply; opacity:.92;
+}
+.pii.is-animated::before { transform:scaleX(0) rotate(-.3deg); transform-origin:left center; animation:marker-stroke .6s cubic-bezier(.18,.72,.18,1) both; animation-delay:calc(var(--i)*45ms); }
+.pii .pii-tag { display:inline-block; margin-left:.28em; padding:0 5px; border-radius:999px; background:color-mix(in srgb,var(--pii-color) 62%,white); color:var(--neutral-700); font:700 8.5px/1.7 var(--mono); letter-spacing:.03em; text-transform:uppercase; vertical-align:.09em; }
+.redacted-tag { padding:0 5px; border-radius:4px; background:var(--neutral-100); border:1px solid var(--border); color:var(--purple); font-weight:600; }
+
+.view-toggle { display:inline-flex; padding:3px; border:1px solid var(--border); border-radius:999px; background:var(--neutral-50); }
+.view-toggle button { padding:5px 11px; border:0; border-radius:999px; background:transparent; color:var(--neutral-500); font:600 11px/1 var(--mono); letter-spacing:.03em; text-transform:uppercase; cursor:pointer; transition:all .15s; }
+.view-toggle button.active { background:var(--ink); color:var(--paper); }
+
+.controls { margin-top:14px; }
+
+.legend { display:flex; flex-wrap:wrap; gap:6px; margin-top:12px; }
+.legend-chip { display:inline-flex; align-items:center; gap:6px; padding:4px 9px 4px 6px; border:1px solid var(--border); border-radius:999px; background:var(--paper); color:var(--neutral-700); font:600 11px/1 var(--sans); }
+.legend-swatch { width:11px; height:11px; border-radius:4px; background:var(--pii-color); box-shadow:inset 0 0 0 1px rgba(0,0,0,.05); }
+.legend-count { color:var(--neutral-500); font:600 10px/1 var(--mono); }
+
+.examples-panel { display:flex; align-items:center; gap:15px; margin-top:16px; padding:15px 17px; overflow:hidden; border:1px solid var(--border); border-radius:11px; background:var(--neutral-50); }
+.examples-label { flex:none; color:var(--neutral-500); font:600 11px/1.3 var(--mono); letter-spacing:.07em; text-transform:uppercase; }
+.example-list { display:flex; flex-wrap:wrap; gap:7px; min-width:0; padding-top:2px; }
+.ex { display:inline-flex; align-items:center; padding:8px 12px; border:1px solid var(--border); border-radius:999px; background:var(--paper); color:var(--neutral-700); font-size:12px; line-height:1.15; cursor:pointer; transition:background .15s,border-color .15s,color .15s,transform .15s,box-shadow .15s; }
+.ex:hover { border-color:rgba(108,79,224,.28); color:var(--purple); transform:translateY(-1px); box-shadow:0 5px 12px rgba(54,42,105,.08); }
+.ex.active { border-color:var(--ink); background:var(--ink); color:var(--paper); box-shadow:none; }
+.sr-only { position:absolute!important; width:1px; height:1px; padding:0; margin:-1px; overflow:hidden; clip:rect(0,0,0,0); white-space:nowrap; border:0; }
+
+@keyframes marker-stroke { 0% { opacity:0; transform:scaleX(0) rotate(-.3deg); } 14% { opacity:.82; } 76% { opacity:1; transform:scaleX(1.03) rotate(-.3deg); } 100% { opacity:.92; transform:scaleX(1) rotate(-.3deg); } }
+
+@media (max-width:980px) { .pii-workspace { grid-template-columns:1fr; } }
+@media (max-width:720px) {
+  .wrap { width:100%; border:0; }
+  .site-intro { padding:20px; }
+  h1 { font-size:40px; }
+  .sub { font-size:14px; }
+  .model-strip { grid-template-columns:48px minmax(0,1fr); column-gap:16px; row-gap:10px; padding:12px 20px; }
+  .load-block { grid-column:1/-1; }
+  main { padding:15px 20px 20px; }
+  .examples-panel { align-items:flex-start; flex-direction:column; }
+}
+@media (max-width:460px) {
+  .model-name-row { align-items:flex-start; flex-direction:column; gap:2px; }
+  .panel-header { flex-direction:column; }
+}
+@media (prefers-reduced-motion:reduce) {
+  *,*::before,*::after { animation-duration:.01ms!important; animation-iteration-count:1!important; transition-duration:.01ms!important; }
+}