Update app.py

app.py

@@ -1,14 +1,35 @@
 import gradio as gr
 from transformers import pipeline
 
-MODEL_ID = "google/flan-t5-small"
+MODEL_ID = "google/flan-t5-small"   # swap to flan-t5-base if you have more CPU
 pipe = pipeline("text2text-generation", model=MODEL_ID)
 
-def infer(prompt, max_new_tokens=128, temperature=0.7, top_p=0.9):
+def infer(prompt, max_new_tokens=96, num_beams=4, length_penalty=0.9):
     if not prompt.strip():
         return "Please type something to generate."
-    out = pipe(prompt, max_new_tokens=int(max_new_tokens), do_sample=True, temperature=float(temperature), top_p=float(top_p))
+    # Beam search is deterministic and tends to avoid loops on T5.
+    out = pipe(
+        prompt,
+        max_new_tokens=int(max_new_tokens),
+        num_beams=int(num_beams),
+        early_stopping=True,
+        length_penalty=float(length_penalty),   # <1.0 favors shorter, reduces rambling
+        no_repeat_ngram_size=3                  # prevents 3-gram loops like "blue sky with"
+    )
     return out[0]["generated_text"]
 
-demo = gr.Interface(fn=infer, inputs=[gr.Textbox(lines=3, label="Instruction"), gr.Slider(16,256,128,step=8,label="Max new tokens"), gr.Slider(0,1.5,0.7,step=0.05,label="Temperature"), gr.Slider(0.1,1.0,0.9,step=0.05,label="Top‑p")], outputs=gr.Textbox(lines=8, label="Output"), title="Mini LLM (Local)")
-demo.launch()
+demo = gr.Interface(
+    fn=infer,
+    inputs=[
+        gr.Textbox(lines=3, label="Instruction",
+                   placeholder="Explain in one paragraph: Why is the sky blue?"),
+        gr.Slider(32, 192, 96, step=8, label="Max new tokens"),
+        gr.Slider(2, 8, 4, step=1, label="Num beams"),
+        gr.Slider(0.6, 1.3, 0.9, step=0.05, label="Length penalty")
+    ],
+    outputs=gr.Textbox(lines=10, label="Output"),
+    title="FLAN-T5 (Deterministic) — Mini LLM"
+)
+
+if __name__ == "__main__":
+    demo.launch()