Update app.py

app.py

@@ -1,49 +1,34 @@
-import os, textwrap, torch, gradio as gr
-import torch, textwrap, gradio as gr
-from transformers import (
-    AutoTokenizer,
-    AutoModelForCausalLM,
-    BitsAndBytesConfig,
-    pipeline,
-)
-AUTH = os.environ.get("HF_TOKEN")  
-MODEL_ID = "mistralai/Mixtral-8x7B-Instruct-v0.1"   # FP16 weights
-bnb_cfg  = BitsAndBytesConfig(
-    load_in_4bit=True,
-    bnb_4bit_compute_dtype=torch.float16,   # keeps mat-mul fast
-)
+# app.py  – CPU-only summariser for Hugging Face Spaces (free tier)
 
-tok   = AutoTokenizer.from_pretrained(MODEL_ID, token=AUTH, use_fast=True)
-model = AutoModelForCausalLM.from_pretrained(
-    MODEL_ID,
-    token=AUTH,
-    device_map="auto",
-    trust_remote_code=True,
-    quantization_config=bnb_cfg,            # perfectly fine here
-)
+import textwrap, gradio as gr
+from transformers import pipeline
 
-prompt_tpl = (
-    "Summarise the following transcript in short in 1 or 2 paragraph and point wise and don't miss any key information cover all"
-)
+# 1️⃣  small, open model that needs no access-token
+MODEL_ID = "sshleifer/distilbart-cnn-12-6"
+summariser = pipeline("summarization", model=MODEL_ID, device=-1)  # -1 = CPU
 
-gen = pipeline("text-generation", model=model, tokenizer=tok,
-               max_new_tokens=256, temperature=0.3)
-
-MAX_CHUNK = 6_000   # ≈ 4 k tokens
+# 2️⃣  rough char limit that maps to the model’s 1 024-token window
+MAX_CHUNK = 3_500
 
 def summarize(txt: str) -> str:
-    parts = textwrap.wrap(txt, MAX_CHUNK, break_long_words=False)
+    """Chunk long transcripts, summarise each, then summarise the summaries."""
+    chunks = textwrap.wrap(txt, MAX_CHUNK, break_long_words=False)
     partials = [
-        gen(prompt_tpl.format(chunk=p))[0]["generated_text"]
-        .split("### Summary:")[-1].strip()
-        for p in parts
+        summariser(c, max_length=160, min_length=30, do_sample=False)[0]["summary_text"]
+        for c in chunks
     ]
-    return gen(prompt_tpl.format(chunk=" ".join(partials)))[0]["generated_text"]\
-             .split("### Summary:")[-1].strip()
+    first_pass = " ".join(partials)
+    # if we had to chunk, do a second pass to get a coherent overall summary
+    if len(chunks) > 1:
+        first_pass = summariser(first_pass, max_length=180, min_length=40, do_sample=False)[0]["summary_text"]
+    return first_pass
+
+demo = gr.Interface(
+    fn=summarize,
+    inputs=gr.Textbox(lines=20, label="Transcript"),
+    outputs="text",
+    title="Free Transcript Summariser – DistilBART-CNN",
+)
 
-demo = gr.Interface(fn=summarize,
-                    inputs=gr.Textbox(lines=20, label="Transcript"),
-                    outputs="text",
-                    title="Free Transcript Summariser – Mixtral-8×7B")
 if __name__ == "__main__":
-    demo.launch()
+    demo.launch()