Update app.py

app.py

@@ -1,6 +1,7 @@
 import streamlit as st
 import json
 import os
+import re
 from transformers import pipeline
 
 STYLE_SAMPLES_FILE = "style_samples.json"
@@ -11,29 +12,69 @@ def load_style_samples():
             return json.load(f)
     return []
 
+def dedupe_sentences(text: str) -> str:
+    # Remove verbatim repeated sentences, keep order
+    parts = re.split(r'(?<=[.!?])\s+', text.strip())
+    seen = set()
+    out = []
+    for p in parts:
+        norm = re.sub(r'\s+', ' ', p.strip().lower())
+        if norm and norm not in seen:
+            seen.add(norm)
+            out.append(p.strip())
+    return " ".join(out)
+
 @st.cache_resource(show_spinner=False)
 def load_pipeline():
-    # CPU-friendly seq2seq model; swap later if you upgrade hardware
+    # CPU-friendly model; swap later to a stronger instruct model if available
     model_id = "google/flan-t5-base"
-    # Avoid device_map to prevent Accelerate requirement on Spaces CPU
     gen_pipe = pipeline(
         task="text2text-generation",
         model=model_id
+        # Note: no device_map to avoid Accelerate requirement on CPU Spaces
     )
     return gen_pipe
 
+def build_prompt(topic, audience, tone, length, style_example_text):
+    # Structured prompt reduces looping and anchors the model
+    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 BODY paragraphs; sentences under 20 words.\n"
+        "- Add 3–5 specific insights or steps; bullets allowed.\n"
+        "- End with a clear CTA inviting comments.\n\n"
+        "Constraints:\n"
+        "- Do NOT repeat sentences or 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 (optional):\n{style_example_text}\n\n"
+        "Output format (use these headers exactly):\n"
+        "HOOK:\n"
+        "BODY:\n"
+        "TAKEAWAY:\n"
+        "CTA:\n"
+    )
+
+# Load resources
 pipe = load_pipeline()
 style_samples = load_style_samples()
 
+# UI
 st.set_page_config(page_title="LinkedIn Post Generator", layout="centered")
 st.title("🔗 LinkedIn Post Generator (Hugging Face)")
-st.write("Generate LinkedIn posts with few-shot style guidance.")
+st.write("Generate concise, structured LinkedIn posts with few-shot style guidance.")
 
 with st.form("gen_form"):
     topic = st.text_input("Post Topic", "Generative AI for Business")
     tone = st.selectbox("Tone", ["Professional", "Friendly", "Inspirational", "Technical", "Concise"])
     audience = st.text_input("Audience", "Startup founders")
-    length = st.slider("Length (approx words)", 30, 400, 120, 10)
+    length = st.slider("Length (approx words)", 40, 300, 120, 10)
+
     use_sample = st.selectbox(
         "Style Sample (optional)",
         ["None"] + [f"Sample {i+1}" for i in range(len(style_samples))]
@@ -43,36 +84,11 @@ with st.form("gen_form"):
     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)
+        repetition_penalty = st.slider("Repetition penalty", 1.0, 2.0, 1.2, 0.05)
         no_repeat_ngram_size = st.slider("No-repeat n-gram size", 1, 6, 3, 1)
 
     submitted = st.form_submit_button("Generate Post")
 
-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
@@ -89,20 +105,21 @@ if submitted:
             try:
                 outputs = pipe(
                     prompt,
-                    max_new_tokens=length + 120,     # give space for sections
+                    max_new_tokens=length + 120,
                     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
+                # Handle list/dict return variants
                 if isinstance(outputs, list) and outputs and "generated_text" in outputs[0]:
-                    result = outputs[0]["generated_text"].strip()
+                    raw = outputs[0]["generated_text"].strip()
                 elif isinstance(outputs, dict) and "generated_text" in outputs:
-                    result = outputs["generated_text"].strip()
+                    raw = outputs["generated_text"].strip()
                 else:
-                    result = str(outputs)
+                    raw = str(outputs)
 
+                result = dedupe_sentences(raw)
                 st.success("Here's your LinkedIn post:")
                 st.write(result)
                 st.download_button("Download post as .txt", result, file_name="linkedin_post.txt")
@@ -110,7 +127,7 @@ if submitted:
                 st.error(f"Error generating post: {e}")
 
 st.markdown("---")
-st.write("Upload a JSON array of style samples (strings). This overwrites existing samples.")
+st.write("Upload a JSON array of style sample strings (overwrites existing).")
 file = st.file_uploader("Upload style_samples.json", type=["json"])
 if file:
     try: