Fix examples

.dockerignore

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

README.md

@@ -1,13 +1,22 @@
----
+---
 title: English spellchecker
-emoji: ✍️
-colorFrom: purple
-colorTo: gray
-sdk: docker
-app_port: 7860
-header: mini
-pinned: false
-private: true
-short_description: Fix typos with a tiny local model on CPU
----
-
+emoji: ✍️
+colorFrom: purple
+colorTo: gray
+sdk: docker
+app_port: 7860
+header: mini
+pinned: false
+private: true
+---
+
+# English spellchecker
+
+A focused grammar and spelling correction demo powered by [`LiquidAI/LFM2.5-Spellchecker-350M`](https://huggingface.co/LiquidAI/LFM2.5-Spellchecker-350M).
+
+The FastAPI backend loads the private model on CPU and exposes word-level corrections to the custom frontend. The interface checks text automatically, visualizes edits in context, provides confidence and iteration controls, and includes concise examples.
+
+- `server.py` — model loading, correction API, health endpoint, and static serving
+- `static/index.html` — application structure and interaction logic
+- `static/style.css` — shared LiquidAI demo design and animations
+- `Dockerfile` and `requirements.txt` — reproducible CPU runtime

server.py

@@ -1,5 +1,12 @@
 """FastAPI backend for the LFM2.5 Spellchecker demo (Docker Space).
 
+Loads the published model from the Hub (pinned to `main`, so the Space always serves the current best
+checkpoint), exposes POST /api/correct, and serves the static frontend in static/. No Gradio.
+
+The model repo is private, so HF_TOKEN (a Space secret) is needed to pull it. Pinned library versions
+(see requirements.txt) match the environment the model was validated against — the encoder's custom
+bidirectional-mask code is sensitive to the transformers version.
+
     uvicorn server:app --host 0.0.0.0 --port 7860
 """
 import difflib
@@ -13,43 +20,17 @@ from fastapi.staticfiles import StaticFiles
 from pydantic import BaseModel
 from transformers import AutoModel
 
-
-def _effective_cpus():
-    """Cores actually granted to this container (cgroup quota), not the host core count.
-    os.cpu_count() returns the host's cores, so torch would oversubscribe the 2-vCPU Space."""
-    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
-
-
-_CPUS = _effective_cpus()
-torch.set_num_threads(_CPUS)
-try:
-    torch.set_num_interop_threads(1)
-except RuntimeError:
-    pass
-
-MODEL_ID = os.environ.get("SPELLCHECKER_MODEL", "LiquidAI/LFM2.5-Encoder-350M-Spellchecker")
-MODEL_REV = os.environ.get("SPELLCHECKER_REVISION", "main")
+MODEL_ID = os.environ.get("SPELLCHECKER_MODEL", "LiquidAI/LFM2.5-Spellchecker-350M")
+# Pin to the EXACT published commit so the container can never serve stale cached weights/remote-code
+# (the bug we hit: a rebuild kept serving old, tagger-only behaviour). Bump on each publish, or override.
+MODEL_REV = os.environ.get("SPELLCHECKER_REVISION", "65a4a90af31205d2f7ef66b6a68d7b3d276adfdd")
 STATIC = os.path.join(os.path.dirname(os.path.abspath(__file__)), "static")
 
-print(f"[server] loading {MODEL_ID}@{MODEL_REV} on {_CPUS} CPU thread(s) ...", flush=True)
+print(f"[server] loading {MODEL_ID}@{MODEL_REV} ...", flush=True)
+# fp16 (half the memory; the published weights are fp16). Casting the whole model uniformly avoids the
+# mixed-dtype error — torch 2.12 runs fp16 matmul on CPU fine.
 _model = AutoModel.from_pretrained(MODEL_ID, revision=MODEL_REV, trust_remote_code=True,
-                                   token=os.environ.get("HF_TOKEN")).float().eval()
+                                   token=os.environ.get("HF_TOKEN")).half().eval()
 _mem_bytes = sum(t.numel() * t.element_size() for t in (*_model.parameters(), *_model.buffers()))
 _mem_human = (f"{_mem_bytes / 1024**3:.2f} GB" if _mem_bytes >= 1024**3
               else f"{_mem_bytes / 1024**2:.0f} MB")
@@ -66,9 +47,9 @@ _ATTACH_LEFT = re.compile(r'\s+([.,!?;:%)\]}»…])')
 _ATTACH_RIGHT = re.compile(r'([(\[{«])\s+')
 _NT = re.compile(r"(\w+?)(n['’]t)\b", re.I)                       # don't->do n't, can't->ca n't
 _CONTR = re.compile(r"(\w)(['’](?:s|re|ve|ll|d|m))\b", re.I)      # let's->let 's, I'm->I 'm
-_REJOIN_NT = re.compile(r"\s+(n['’]t)\b", re.I)                   # do n't->don't
-_REJOIN_CONTR = re.compile(r"\s+(['’](?:s|re|ve|ll|d|m))\b", re.I)  # let 's->let's
-_REJOIN_SPLIT_APOSTROPHE = re.compile(r"(?<=\w)\s+(['’])\s+(?=(?:s|t|d|m|re|ve|ll)\b)", re.I)
+_REJOIN_NT = re.compile(r"\s+(n['’]t)\b", re.I)                   # do n't->don't
+_REJOIN_CONTR = re.compile(r"\s+(['’](?:s|re|ve|ll|d|m))\b", re.I)  # let 's->let's
+_REJOIN_SPLIT_APOSTROPHE = re.compile(r"(?<=\w)\s+(['’])\s+(?=(?:s|t|d|m|re|ve|ll)\b)", re.I)
 
 
 def tokenize(text: str) -> str:
@@ -78,15 +59,18 @@ def tokenize(text: str) -> str:
     return re.sub(r"\s+", " ", text).strip()
 
 
-def detok(text: str) -> str:
-    text = _ATTACH_LEFT.sub(r"\1", text)
-    text = _ATTACH_RIGHT.sub(r"\1", text)
-    text = _REJOIN_NT.sub(r"\1", text)                           # do n't->don't
-    text = _REJOIN_CONTR.sub(r"\1", text)                        # let 's->let's
-    text = _REJOIN_SPLIT_APOSTROPHE.sub(r"\1", text)              # don ' t->don't
-    return re.sub(r"\s+", " ", text).strip()
+def detok(text: str) -> str:
+    text = _ATTACH_LEFT.sub(r"\1", text)
+    text = _ATTACH_RIGHT.sub(r"\1", text)
+    text = _REJOIN_NT.sub(r"\1", text)                           # do n't->don't
+    text = _REJOIN_CONTR.sub(r"\1", text)                        # let 's->let's
+    text = _REJOIN_SPLIT_APOSTROPHE.sub(r"\1", text)              # don ' t->don't
+    return re.sub(r"\s+", " ", text).strip()
 
 
+# Startup self-test over the REAL user path (tokenize -> correct -> detok), logged + exposed at
+# /api/health: a correctly-deployed full system leaves this clean sentence UNCHANGED. If it changes,
+# the deploy is wrong (stale model, reranker inactive, or contraction handling broken) — no guessing.
 _PROBE_IN = "That's a fair point, let's discuss it tomorrow."
 try:
     _PROBE_OUT = detok(_model.correct([tokenize(_PROBE_IN)], max_iter=3)[0])

static/index.html

@@ -7,22 +7,29 @@
 <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>
+  (() => {
+    const sign = new URLSearchParams(location.search).get("__sign");
+    const suffix = sign ? `?__sign=${encodeURIComponent(sign)}` : "";
+    window.__STATIC_ASSET_SUFFIX = suffix;
+    document.write(`<link rel="stylesheet" href="./style.css${suffix}">`);
+  })();
+</script>
 </head>
 <body>
 <div class="wrap">
   <header class="site-intro">
     <h1>English <em>spellchecker</em></h1>
-    <p class="sub">Correct English spelling and grammar with LFM2.5.</p>
+    <p class="sub">Correct English spelling and grammar privately with LFM2.5.</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">
+      <img id="model-logo" class="model-logo" width="56" height="59" alt="">
     </span>
     <div class="model-copy">
       <div class="model-name-row">
-        <a href="https://huggingface.co/LiquidAI/LFM2.5-Encoder-350M-Spellchecker" target="_blank" rel="noopener">LFM2.5 Spellchecker</a>
+        <a href="https://huggingface.co/LiquidAI/LFM2.5-Spellchecker-350M" target="_blank" rel="noopener">LFM2.5 Spellchecker</a>
         <span class="model-variant">350M parameters</span>
       </div>
     </div>
@@ -33,7 +40,9 @@
       </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>
+    <span id="gpu" hidden></span>
   </section>
+  <script>document.getElementById("model-logo").src = `./lliquid.gif${window.__STATIC_ASSET_SUFFIX || ""}`;</script>
 
   <main>
     <section class="spell-workspace" id="spell-workspace" aria-label="Spellchecker">
@@ -157,13 +166,10 @@ const EXAMPLES = [
 ];
 
 const $ = id => document.getElementById(id);
-// Private Space auth: HF signs the embedded iframe URL with ?__sign=<token>.
-// Incognito / Safari block the third-party .hf.space cookie, so carry that token
-// on same-origin API requests, else /api/* returns 404.
-const __sign = new URLSearchParams(location.search).get("__sign");
+const sign = new URLSearchParams(location.search).get("__sign");
 const signedUrl = path => {
   const url = new URL(path, location.href);
-  if (__sign) url.searchParams.set("__sign", __sign);
+  if (sign) url.searchParams.set("__sign", sign);
   return url;
 };
 const input = $("input"), output = $("output"), copy = $("copy"), workspace = $("spell-workspace");