Update app.py

app.py

@@ -7,54 +7,72 @@ from transformers import pipeline
 subprocess.check_call([sys.executable, "-m", "spacy", "download", "en_core_web_sm"])
 nlp = spacy.load("en_core_web_sm")
 spell = SpellChecker(language='en')
-para = pipeline("text2text-generation", model="Vamsi/T5_Paraphrase_Paws", device=-1)
+
+# load model once
+para = pipeline(
+    "text2text-generation",
+    model="Vamsi/T5_Paraphrase_Paws",
+    device=-1,
+    max_new_tokens=1024  # ← allow longer outputs per chunk
+)
 
 def simple_correct(text):
     return " ".join(spell.correction(w) or w for w in text.split())
 
 def chunk_text(text, max_words=380):
+    """Split into exact word-length chunks."""
     words = text.split()
     for i in range(0, len(words), max_words):
         yield " ".join(words[i:i + max_words])
 
 def humanize(text, tone, strength, freeze):
     locked = [(ent.text, ent.label_) for ent in nlp(text).ents] if freeze else []
-    chunks_out = []
+    out_chunks = []
+
     for chunk in chunk_text(text, 380):
-        paraphrased = para(chunk, max_length=512, do_sample=True,
-                           temperature=0.7 * strength / 10)[0]['generated_text']
+        paraphrased = para(
+            chunk,
+            max_new_tokens=1024,
+            do_sample=True,
+            temperature=0.7 * strength / 10
+        )[0]['generated_text']
+
         paraphrased = simple_correct(paraphrased)
+
+        # restore locked entities
         for ent, label in locked:
             paraphrased = re.sub(re.escape(ent), ent, paraphrased, flags=re.IGNORECASE)
-        chunks_out.append(paraphrased)
-    full_text = "\n\n".join(chunks_out)
 
-    # save downloadable file
-    file_path = "/tmp/humanized.txt"
+        out_chunks.append(paraphrased)
+
+    full_text = " ".join(out_chunks)
+
+    # downloadable file
+    file_path = "/tmp/full_humanized.txt"
     with open(file_path, "w", encoding="utf-8") as f:
         f.write(full_text)
+
     return full_text, file_path
 
-with gr.Blocks(title="AI Humanizer – Unlimited Words") as demo:
-    gr.Markdown("## AI Humanizer – Unlimited Words")
+with gr.Blocks(title="AI Humanizer – Exact Length") as demo:
+    gr.Markdown("## AI Humanizer – Exact Length, No Compression")
     with gr.Row():
         with gr.Column():
-            text_in = gr.Textbox(label="Paste your LONG text", lines=15, max_lines=None)
+            text_in = gr.Textbox(label="Paste your text (any length)", lines=15, max_lines=None)
             tone_dd = gr.Dropdown(["Casual", "Academic", "Marketing", "Legal", "Creative"], value="Casual", label="Tone")
             strength_sl = gr.Slider(1, 10, value=5, label="Strength 1-10")
             lock_cb = gr.Checkbox(label="Lock facts / dates / names")
             submit_btn = gr.Button("Humanize", variant="primary")
 
         with gr.Column():
-            text_out = gr.Textbox(label="FULL humanized text", lines=25, interactive=True)
+            text_out = gr.Textbox(label="Full humanized text (same word-count)", lines=25, interactive=True)
             with gr.Row():
-                copy_btn = gr.Button("📋 Copy Text")
+                copy_btn = gr.Button("📋 Copy")
                 file_out = gr.File(label="Download .txt")
 
     submit_btn.click(humanize, inputs=[text_in, tone_dd, strength_sl, lock_cb],
                      outputs=[text_out, file_out])
 
-    # JavaScript one-line copy
     copy_btn.click(None, text_out, None,
                    js="(txt) => navigator.clipboard.writeText(txt)")