Update app.py

app.py

@@ -5,10 +5,11 @@ Implements the 5-module architecture from the design document:
   1. UI Module            — Gradio web interface
   2. Validation & Flow    — Input validation and routing
   3. Pre-processing       — 6-Element structured prompt builder
-  4. Generation           — Gemma-4 LLM via Hugging Face
+  4. Generation           — Gemma-3 LLM via Hugging Face
   5. Output               — Parse raw LLM output into structured results
 """
  
+import os
 import re
 import torch
 import gradio as gr
@@ -16,17 +17,19 @@ from transformers import AutoTokenizer, AutoModelForCausalLM, pipeline
  
 # ---------------------------------------------------------------------------
 # MODULE 4 — Generation Module (LLM setup)
+# FIX: Correct model ID + read HF_TOKEN from environment (set as Space secret)
 # ---------------------------------------------------------------------------
-MODEL_ID = "google/gemma-4-it"
+MODEL_ID  = "google/gemma-3-4b-it"   # smallest Gemma-3 — works on free tier
+HF_TOKEN  = os.environ.get("HF_TOKEN")  # set this in Space Settings → Secrets
  
 _tokenizer = None
-_pipe = None
+_pipe      = None
  
 def get_tokenizer():
     """Returns the tokenizer, loading it on first call."""
     global _tokenizer
     if _tokenizer is None:
-        _tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)
+        _tokenizer = AutoTokenizer.from_pretrained(MODEL_ID, token=HF_TOKEN)
     return _tokenizer
  
 def get_pipeline():
@@ -36,15 +39,16 @@ def get_pipeline():
         tokenizer = get_tokenizer()
         model = AutoModelForCausalLM.from_pretrained(
             MODEL_ID,
-            torch_dtype=torch.bfloat16,   # efficient on modern GPUs/CPUs
-            device_map="auto",             # auto-detect GPU/CPU
+            token=HF_TOKEN,
+            torch_dtype=torch.bfloat16,
+            device_map="auto",
         )
         _pipe = pipeline(
             "text-generation",
             model=model,
             tokenizer=tokenizer,
             max_new_tokens=512,
-            do_sample=False,               # deterministic output
+            do_sample=False,
         )
     return _pipe
  
@@ -114,7 +118,6 @@ def parse_output(raw: str) -> dict:
     Returns a dict with keys: result, accuracy, summary, issues.
     Falls back gracefully if parsing fails.
     """
-    # Strip any echoed prompt (model sometimes repeats <start_of_turn>)
     if "<start_of_turn>model" in raw:
         raw = raw.split("<start_of_turn>model")[-1]
  
@@ -148,10 +151,7 @@ def parse_output(raw: str) -> dict:
  
 def format_for_display(parsed: dict) -> tuple[str, str, str]:
     """
-    Converts the parsed dict into three Gradio-friendly strings:
-      - verdict     (shown in a highlighted Textbox)
-      - metrics     (accuracy + summary)
-      - issues_text (bullet list)
+    Converts the parsed dict into three Gradio-friendly strings.
     """
     emoji   = "✅" if parsed["result"] == "PASS" else ("❌" if parsed["result"] == "FAIL" else "⚠️")
     verdict = f"{emoji}  {parsed['result']}"
@@ -167,7 +167,7 @@ def format_for_display(parsed: dict) -> tuple[str, str, str]:
 # MODULE 4 — Generation Module (inference call)
 # ---------------------------------------------------------------------------
 def generate(prompt: str) -> str:
-    outputs = get_pipeline()(prompt, return_full_text=False)  # FIX: was pipe() — undefined variable
+    outputs = get_pipeline()(prompt, return_full_text=False)
     return outputs[0]["generated_text"]
  
  
@@ -185,10 +185,14 @@ def validate_and_evaluate(description: str, code: str):
     if not code or not code.strip():
         return "", "", "", "⚠️ Please provide Python code to evaluate."
  
+    # --- Check token is available ---
+    if not HF_TOKEN:
+        return "", "", "", "❌ HF_TOKEN secret is not set. Add it in Space Settings → Variables and secrets."
+ 
     # --- Pre-processing ---
     prompt = build_prompt(description, code)
  
-    # Token length guard — warn user before sending oversized input to the model
+    # Token length guard
     token_count = len(get_tokenizer().encode(prompt))
     if token_count > 2048:
         return "", "", "", (
@@ -206,18 +210,18 @@ def validate_and_evaluate(description: str, code: str):
     parsed  = parse_output(raw_output)
     verdict, metrics, issues = format_for_display(parsed)
  
-    return verdict, metrics, issues, ""   # empty error = success
+    return verdict, metrics, issues, ""
  
  
 # ---------------------------------------------------------------------------
 # MODULE 1 — UI Module (Gradio)
 # ---------------------------------------------------------------------------
-with gr.Blocks(title="Python Code Evaluator — SE-Group1") as demo:  # FIX: theme moved to launch()
+with gr.Blocks(title="Python Code Evaluator — SE-Group1") as demo:
  
     gr.Markdown(
         """
         # 🐍 Python Code Evaluator
-        **COS60011 — SE-Group1** | Powered by Gemma-4 via Hugging Face
+        **COS60011 — SE-Group1** | Powered by Gemma-3 via Hugging Face
         Enter a **requirements description** and your **Python code**.
         The system will evaluate whether the code correctly implements the described behaviour.
         """
@@ -259,4 +263,5 @@ with gr.Blocks(title="Python Code Evaluator — SE-Group1") as demo:  # FIX: the
     )
  
 if __name__ == "__main__":
-    demo.launch(share=False, theme=gr.themes.Soft(primary_hue="blue"))  
+    demo.launch(share=False, theme=gr.themes.Soft(primary_hue="blue"))
+