Update app.py

app.py

@@ -13,9 +13,9 @@ def load_style_samples():
 
 @st.cache_resource(show_spinner=False)
 def load_pipeline():
-    # Choose a lightweight, CPU-friendly model for Spaces
+    # CPU-friendly seq2seq model; swap later if you upgrade hardware
     model_id = "google/flan-t5-base"
-    # IMPORTANT: Do NOT pass device_map here to avoid needing accelerate
+    # Avoid device_map to prevent Accelerate requirement on Spaces CPU
     gen_pipe = pipeline(
         task="text2text-generation",
         model=model_id
@@ -27,7 +27,7 @@ style_samples = load_style_samples()
 
 st.set_page_config(page_title="LinkedIn Post Generator", layout="centered")
 st.title("🔗 LinkedIn Post Generator (Hugging Face)")
-st.write("Generate LinkedIn posts in your style using a compact open model.")
+st.write("Generate LinkedIn posts with few-shot style guidance.")
 
 with st.form("gen_form"):
     topic = st.text_input("Post Topic", "Generative AI for Business")
@@ -39,37 +39,70 @@ with st.form("gen_form"):
         ["None"] + [f"Sample {i+1}" for i in range(len(style_samples))]
     )
     custom_style = st.text_area("Or paste your own style sample (optional)")
+
+    with st.expander("Advanced generation settings"):
+        temperature = st.slider("Temperature", 0.1, 1.2, 0.7, 0.05)
+        top_p = st.slider("Top-p (nucleus)", 0.1, 1.0, 0.9, 0.05)
+        repetition_penalty = st.slider("Repetition penalty", 1.0, 2.0, 1.15, 0.05)
+        no_repeat_ngram_size = st.slider("No-repeat n-gram size", 1, 6, 3, 1)
+
     submitted = st.form_submit_button("Generate Post")
 
-prompt_style = ""
+def build_prompt(topic, audience, tone, length, style_example_text):
+    # Structured prompt to reduce repetition and produce LinkedIn-ready content
+    return (
+        "Task: Write a LinkedIn post.\n\n"
+        f"Topic: \"{topic}\"\n"
+        f"Audience: \"{audience}\"\n"
+        f"Tone: \"{tone}\"\n"
+        f"Target length: ~{length} words.\n\n"
+        "Style requirements:\n"
+        "- Start with a 1–2 line hook with a concrete claim or question.\n"
+        "- Use 2–3 short paragraphs; keep sentences under 20 words.\n"
+        "- Add 3–5 specific insights or steps (use bullet points if helpful).\n"
+        "- End with a clear CTA (ask a question or invite comments).\n\n"
+        "Constraints:\n"
+        "- No repeated sentences or filler phrases.\n"
+        "- Avoid clichés like “it's a great example of how we can make a difference in the world.”\n"
+        "- Use plain business English.\n\n"
+        f"Reference style:\n{style_example_text}\n\n"
+        "Output format:\n"
+        "HOOK:\n"
+        "BODY:\n"
+        "TAKEAWAY:\n"
+        "CTA:\n"
+    )
+
+style_example_text = ""
 if use_sample != "None":
     idx = int(use_sample.split()[1]) - 1
-    prompt_style += f"Sample style: {style_samples[idx]}\n"
+    style_example_text += f"Sample style:\n{style_samples[idx]}\n"
 if custom_style.strip():
-    prompt_style += f"User style: {custom_style}\n"
-
-prompt = (
-    f"Write a LinkedIn post on '{topic}'.\n"
-    f"Tone: {tone}. Audience: {audience}. Length: about {length} words.\n"
-    f"{prompt_style}"
-    "Write in engaging, natural language, and end with a strong call to action."
-)
+    style_example_text += f"Custom style:\n{custom_style}\n"
 
 if submitted:
     if not topic.strip():
         st.warning("Please enter a topic.")
     else:
+        prompt = build_prompt(topic, audience, tone, length, style_example_text)
         with st.spinner("Generating..."):
             try:
-                # flan-t5 uses text2text; pipeline returns list of dicts
-                outputs = pipe(prompt, max_new_tokens=length + 48)
-                # Support both list and dict return shapes defensively
-                if isinstance(outputs, list) and len(outputs) and "generated_text" in outputs[0]:
+                outputs = pipe(
+                    prompt,
+                    max_new_tokens=length + 120,     # give space for sections
+                    temperature=temperature,
+                    top_p=top_p,
+                    repetition_penalty=repetition_penalty,
+                    no_repeat_ngram_size=no_repeat_ngram_size
+                )
+                # Pipeline may return list or dict; handle both
+                if isinstance(outputs, list) and outputs and "generated_text" in outputs[0]:
                     result = outputs[0]["generated_text"].strip()
                 elif isinstance(outputs, dict) and "generated_text" in outputs:
                     result = outputs["generated_text"].strip()
                 else:
                     result = str(outputs)
+
                 st.success("Here's your LinkedIn post:")
                 st.write(result)
                 st.download_button("Download post as .txt", result, file_name="linkedin_post.txt")