Upload 3 files

README.md

@@ -9,3 +9,4 @@ app_file: app.py
 pinned: false
 ---
 
+

app.py

@@ -1,20 +1,17 @@
-import os, re, gradio as gr
-from transformers import AutoTokenizer, AutoModelForSequenceClassification, pipeline
+import os, re, time
+import gradio as gr
 
+# Keep it light on CPU and quiet
 os.environ.setdefault("TOKENIZERS_PARALLELISM", "false")
 
 URL_MODEL_ID = "CrabInHoney/urlbert-tiny-v4-malicious-url-classifier"
 URL_LABEL_MAP = {"LABEL_0":"benign","LABEL_1":"defacement","LABEL_2":"malware","LABEL_3":"phishing"}
 URL_RE = re.compile(r"""(?xi)\b(?:https?://|www\.)[a-z0-9\-._~%]+(?:/[^\s<>"']*)?""")
 
-_url_pipe = None
-def _get_url_pipe():
-    global _url_pipe
-    if _url_pipe is None:
-        tok = AutoTokenizer.from_pretrained(URL_MODEL_ID)
-        mdl = AutoModelForSequenceClassification.from_pretrained(URL_MODEL_ID)
-        _url_pipe = pipeline("text-classification", model=mdl, tokenizer=tok, device=-1, top_k=None)
-    return _url_pipe
+_pipe = None  # lazy-initialized transformers pipeline
+
+def _extract_urls(text: str):
+    return sorted(set(m.group(0) for m in URL_RE.finditer(text or "")))
 
 def _pretty(raw, id2label):
     if id2label:
@@ -23,55 +20,60 @@ def _pretty(raw, id2label):
         if k in id2label: return id2label[k]
     return URL_LABEL_MAP.get(raw, raw)
 
-def _extract_urls(text: str):
-    return sorted(set(m.group(0) for m in URL_RE.finditer(text or "")))
-
-def _classify_url(u: str):
-    pipe = _get_url_pipe()
-    id2label = getattr(pipe.model.config, "id2label", None)
-    scores = sorted(pipe(u)[0], key=lambda s: s["score"], reverse=True)
-    top = scores[0]
-    lbl = _pretty(top["label"], id2label)
-    conf_pct = round(100*float(top["score"]), 2)
-    return lbl, conf_pct
-
-def analyze(text: str):
-    text = (text or "").strip()
+def analyze(input_text: str):
+    """
+    Returns tuple for Interface: verdict, top_label, top_conf, table
+    """
+    t0 = time.time()
+    text = (input_text or "").strip()
     if not text:
         return "Paste an email body or a URL.", "", "", []
+
     urls = [text] if (text.lower().startswith(("http://","https://","www.")) and " " not in text) else _extract_urls(text)
     if not urls:
         return "No URLs detected in the text.", "", "", []
+
+    # Lazy import + pipeline creation (keeps app startup instant)
+    global _pipe
+    if _pipe is None:
+        from transformers import AutoTokenizer, AutoModelForSequenceClassification, pipeline
+        tok = AutoTokenizer.from_pretrained(URL_MODEL_ID)
+        mdl = AutoModelForSequenceClassification.from_pretrained(URL_MODEL_ID)
+        _pipe = pipeline("text-classification", model=mdl, tokenizer=tok, device=-1, top_k=None)
+
+    id2label = getattr(_pipe.model.config, "id2label", None)
     rows, unsafe, top_label, top_conf = [], False, "", ""
+
     for i, u in enumerate(urls, 1):
-        try:
-            lbl, conf = _classify_url(u)
-        except Exception as e:
-            rows.append([u, f"Error: {type(e).__name__}", "—"])
-            continue
+        scores = sorted(_pipe(u)[0], key=lambda s: s["score"], reverse=True)
+        top = scores[0]
+        lbl = _pretty(top["label"], id2label)
+        conf = round(100*float(top["score"]), 2)
         rows.append([u, lbl, conf])
         if i == 1:
             top_label, top_conf = lbl, f"{conf:.2f}%"
         if lbl.lower() in {"phishing","malware","defacement"}:
             unsafe = True
+
     verdict = "🔴 UNSAFE (links flagged)" if unsafe else "🟢 SAFE (all links benign)"
     return verdict, top_label, top_conf, rows
 
-with gr.Blocks() as demo:
-    gr.Markdown("# 🛡️ Phishing Detector (via Link Analysis)")
-    gr.Markdown("Paste an **email** or a **URL**. We extract links and classify each one.")
-    inp = gr.Textbox(label="Email or URL", lines=6)
-    btn = gr.Button("Analyze", variant="primary")
-    verdict = gr.Markdown()
-    label = gr.Textbox(label="Prediction", interactive=False)
-    conf = gr.Textbox(label="Confidence", interactive=False)
-    table = gr.Dataframe(headers=["URL","Prediction","Confidence (%)"],
-                         datatype=["str","str","number"],
-                         row_count=(0,"dynamic"),
-                         col_count=(3,"fixed"),
-                         interactive=False,
-                         label="Per-link results")
-    btn.click(analyze, [inp], [verdict, label, conf, table], show_progress=True)
+demo = gr.Interface(
+    fn=analyze,
+    inputs=gr.Textbox(lines=6, label="Email or URL", placeholder="Paste a URL or a full email…"),
+    outputs=[
+        gr.Markdown(label="Verdict"),
+        gr.Textbox(label="Prediction", interactive=False),
+        gr.Textbox(label="Confidence", interactive=False),
+        gr.Dataframe(headers=["URL","Prediction","Confidence (%)"],
+                     datatype=["str","str","number"],
+                     row_count=(0,"dynamic"), col_count=(3,"fixed"),
+                     interactive=False, label="Per-link results")
+    ],
+    title="🛡️ Phishing Detector (via Link Analysis)",
+    description="We extract links from your text and classify each with a compact malicious-URL model."
+)
 
 if __name__ == "__main__":
-    demo.launch(server_name="0.0.0.0", server_port=7860, show_error=True)
+    # Let Spaces decide host/port; keep defaults for maximum compatibility
+    demo.launch()

requirements.txt

@@ -3,3 +3,4 @@ transformers==4.55.2
 
 --extra-index-url https://download.pytorch.org/whl/cpu
 torch==2.4.0+cpu
+