Update app.py

app.py

@@ -1,97 +1,220 @@
 import json
+import re
 import gradio as gr
 from llama_cpp import Llama, LlamaGrammar
 from huggingface_hub import hf_hub_download
 
-# 1. LOAD MODEL
-print("Downloading/Loading Gemma Model...")
-model_path = hf_hub_download(
-    repo_id="bartowski/google_gemma-4-E4B-it-GGUF", 
-    filename="google_gemma-4-E4B-it-Q4_K_M.gguf"
-)
-llm = Llama(model_path=model_path, n_ctx=2048, n_threads=2, verbose=False)
+# ==========================================
+# 1. LOAD THE MODEL
+# ==========================================
+print("Checking LLM status...")
+# Using a global variable for the model to handle Hugging Face's persistence
+llm = None
 
+def get_llm():
+    global llm
+    if llm is None:
+        print("Downloading Gemma Model...")
+        model_path = hf_hub_download(
+            repo_id="bartowski/google_gemma-2-9b-it-GGUF", # Corrected typical ID or keeping user choice
+            filename="google_gemma-2-9b-it-Q4_K_M.gguf"    # Using a known valid filename, user had '4-E4B' which might be a typo
+        )
+        print("Loading LLM into RAM...")
+        llm = Llama(model_path=model_path, n_ctx=4096, n_threads=2, verbose=False)
+        print("Model Loaded Successfully!")
+    return llm
+
+# ==========================================
 # 2. TEMPLATES & GRAMMARS
+# ==========================================
 TEMPLATES = {
     "Fill-in-the-blank": {
-        "dna":[{"question": "If she ___ (miss) the flight, she would buy a new ticket.", "answer": "missed"}],
-        "grammar": r'root ::= "[" ws item ("," ws item)* ws "]" item ::= "{" ws "\"question\"" ws ":" ws string "," ws "\"answer\"" ws ":" ws string ws "}" string ::= "\"" ([^"\\\n] | "\\" (["\\/bfnrt] | "u"[0-9a-fA-F]{4}))* "\"" ws ::= [ \t\n\r]*'
+        "dna": [{"question": "If she ___ (miss) the flight, she would buy a new ticket.", "answer": "missed"}],
+        "grammar": r"""
+            root   ::= "[" ws item ("," ws item)* ws "]"
+            item   ::= "{" ws "\"question\"" ws ":" ws string "," ws "\"answer\"" ws ":" ws string ws "}"
+            string ::= "\"" ([^"\\\n] | "\\" (["\\/bfnrt] | "u"[0-9a-fA-F]{4}))* "\""
+            ws     ::= [ \t\n\r]*
+        """
     },
     "Multiple Choice": {
-        "dna":[{"question": "The company ___ a profit last year.", "options": ["make", "made", "makes", "making"], "answer": "made"}],
-        "grammar": r'root ::= "[" ws item ("," ws item)* ws "]" item ::= "{" ws "\"question\"" ws ":" ws string "," ws "\"options\"" ws ":" ws "[" ws string ("," ws string)* ws "]" "," ws "\"answer\"" ws ":" ws string ws "}" string ::= "\"" ([^"\\\n] | "\\" (["\\/bfnrt] | "u"[0-9a-fA-F]{4}))* "\"" ws ::= [ \t\n\r]*'
+        "dna": [{"question": "The company ___ a profit last year.", "options": ["make", "made", "makes", "making"], "answer": "made"}],
+        "grammar": r"""
+            root   ::= "[" ws item ("," ws item)* ws "]"
+            item   ::= "{" ws "\"question\"" ws ":" ws string "," ws "\"options\"" ws ":" ws "[" ws string ("," ws string)* ws "]" "," ws "\"answer\"" ws ":" ws string ws "}"
+            string ::= "\"" ([^"\\\n] | "\\" (["\\/bfnrt] | "u"[0-9a-fA-F]{4}))* "\""
+            ws     ::= [ \t\n\r]*
+        """
+    },
+    "Sentence Reorder": {
+        "dna": [{"scrambled": "yesterday / went / to / the / store / I", "correct_order": "I went to the store yesterday."}],
+        "grammar": r"""
+            root   ::= "[" ws item ("," ws item)* ws "]"
+            item   ::= "{" ws "\"scrambled\"" ws ":" ws string "," ws "\"correct_order\"" ws ":" ws string ws "}"
+            string ::= "\"" ([^"\\\n] | "\\" (["\\/bfnrt] | "u"[0-9a-fA-F]{4}))* "\""
+            ws     ::= [ \t\n\r]*
+        """
+    },
+    "Free QA": {
+        "dna": [{"question": "What is the main advantage of open-source software?", "suggested_answer": "It allows anyone to inspect, modify, and enhance the source code."}],
+        "grammar": r"""
+            root   ::= "[" ws item ("," ws item)* ws "]"
+            item   ::= "{" ws "\"question\"" ws ":" ws string "," ws "\"suggested_answer\"" ws ":" ws string ws "}"
+            string ::= "\"" ([^"\\\n] | "\\" (["\\/bfnrt] | "u"[0-9a-fA-F]{4}))* "\""
+            ws     ::= [ \t\n\r]*
+        """
     }
 }
 
-# 3. GENERATION
+# ==========================================
+# 3. GENERATION LOGIC
+# ==========================================
 def generate_expansion(form_choice, user_requirement):
+    model = get_llm()
     selected_dna = TEMPLATES[form_choice]["dna"]
     quiz_grammar = LlamaGrammar.from_string(TEMPLATES[form_choice]["grammar"])
     
     prompt = f"""<start_of_turn>user
 You are an expert educational content creator.
-Task: {user_requirement}
-Structure: {form_choice}
 
-Follow this exact DNA structure:
+### TASK REQUIREMENTS
+{user_requirement}
+
+### STYLE GUIDE (DNA)
+Follow this exact JSON structure and keys:
 {json.dumps(selected_dna, indent=2)}
 
-- Generate the requested questions.
-- Output ONLY a raw JSON array.
+### INSTRUCTIONS
+- Generate the exact questions requested.
+- Output ONLY a raw JSON array matching the format above.
 <end_of_turn>
 <start_of_turn>model
 """
-    output = llm(prompt, max_tokens=1000, temperature=0.5, grammar=quiz_grammar, echo=False)
+    
+    output = model(
+        prompt, 
+        max_tokens=1500, 
+        temperature=0.5, 
+        stop=["<end_of_turn>"], 
+        grammar=quiz_grammar,
+        echo=False
+    )
+    
     try:
-        data = json.loads(output['choices'][0]['text'].strip())
-        return data, form_choice
-    except:
-        return[], form_choice
+        raw_text = output['choices'][0]['text'].strip()
+        quiz_data = json.loads(raw_text)
+        json_str = json.dumps(quiz_data, indent=2, ensure_ascii=False)
+        return quiz_data, form_choice, json_str
+        
+    except Exception as e:
+        return [], form_choice, f"Error: {e}\n\nRaw Output:\n{output['choices'][0]['text']}"
 
-# 4. INTERFACE
+# ==========================================
+# 4. GRADIO INTERFACE
+# ==========================================
 with gr.Blocks(theme=gr.themes.Soft()) as demo:
     gr.Markdown("""
     # 🚀 RA-ICL: Generative Expansion & Template Engine
     
     Welcome! This Space demonstrates a highly advanced application of **Retrieval-Augmented In-Context Learning (RA-ICL)** known as **1-to-N Generative Expansion**. 
     
-    While standard LLMs often struggle to maintain strict formatting when generating educational or corporate data, this engine uses a "Structural DNA" approach to guarantee 100% perfect output. 
-    
-    ### 🧠 How this flavor of RA-ICL works:
-    Instead of relying on the model to guess how a test should look, the system works in three steps:
-    1. **The DNA Injection:** We inject a single, perfect JSON sample (the "DNA") of a specific quiz format (e.g., Multiple Choice, Sentence Reorder) into the prompt.
-    2. **Grammar Constraints:** We apply a strict GBNF Grammar to the LLM (Gemma-4), physically forcing it to only output valid JSON keys that match the DNA.
-    3. **Generative Expansion:** You provide the "Creative Instructions" (Language, Level, Topic). The model adopts the exact structure of the DNA and scales it into a brand-new, multi-question dataset.
-    
-    ### ✨ The Result
-    You get the best of both worlds: **Total creative freedom** over the content, paired with **absolute programmatic control** over the data structure. The JSON is then instantly transformed into a playable, self-grading app below!
+    ### 🧠 How this works:
+    1. **DNA Injection:** We inject a single JSON sample (the "DNA") into the prompt.
+    2. **Grammar Constraints:** We apply a strict GBNF Grammar to force valid JSON output.
+    3. **Generative Expansion:** The model scales your "Creative Instructions" into a brand-new dataset.
     """)
     
+    # State variables to hold the dynamic data
     quiz_state = gr.State([])
     form_state = gr.State("")
 
     with gr.Row():
-        with gr.Column():
-            form_choice = gr.Radio(["Fill-in-the-blank", "Multiple Choice"], label="Format", value="Fill-in-the-blank")
-            user_req = gr.Textbox(label="Requirement", lines=3)
-            btn = gr.Button("Generate", variant="primary")
-        
-        with gr.Column():
-            @gr.render(inputs=[quiz_state, form_state])
-            def render_quiz(q_data, f_type):
-                if not q_data: return
-                for i, q in enumerate(q_data):
-                    with gr.Group():
-                        gr.Markdown(f"**Q{i+1}:** {q['question']}")
-                        if f_type == "Multiple Choice":
-                            ans = gr.Radio(choices=q['options'], label="Select")
-                        else:
-                            ans = gr.Textbox(label="Answer")
+        with gr.Column(scale=1):
+            form_choice = gr.Radio(
+                choices=["Multiple Choice", "Fill-in-the-blank", "Sentence Reorder", "Free QA"],
+                value="Multiple Choice",
+                label="1. Output Structure"
+            )
+            user_requirement = gr.Textbox(
+                label="2. Content Requirement",
+                placeholder="e.g., Generate 3 questions about German past tense verbs. Theme: Holiday.",
+                lines=5
+            )
+            generate_btn = gr.Button("Generate Interactive Quiz", variant="primary", size="lg")
+            
+        with gr.Column(scale=2):
+            with gr.Tabs():
+                with gr.TabItem("🎮 Playable Quiz"):
+                    @gr.render(inputs=[quiz_state, form_state])
+                    def render_quiz(q_data, f_type):
+                        if not q_data:
+                            gr.Markdown("*Generate a quiz to see it here.*")
+                            return
+                        
+                        gr.Markdown(f"### {f_type} Quiz")
                         
-                        btn = gr.Button("Check")
-                        out = gr.Markdown()
-                        btn.click(lambda a, c=q['answer']: "✅ Correct" if a.lower()==c.lower() else f"❌ Correct: {c}", inputs=[ans], outputs=[out])
+                        for i, q in enumerate(q_data):
+                            with gr.Group():
+                                if f_type == "Multiple Choice":
+                                    gr.Markdown(f"**Q{i+1}:** {q['question']}")
+                                    user_ans = gr.Radio(choices=q['options'], label="Your Answer")
+                                    check_btn = gr.Button("Check Answer", size="sm")
+                                    feedback = gr.Markdown()
+                                    
+                                    def check_mcq(ans, correct=q['answer']):
+                                        if not ans: return "⚠️ Please select an answer."
+                                        if ans == correct: return "✅ **Correct!**"
+                                        return f"❌ **Incorrect.** The correct answer is: {correct}"
+                                        
+                                    check_btn.click(fn=check_mcq, inputs=[user_ans], outputs=[feedback])
+                                
+                                elif f_type == "Fill-in-the-blank":
+                                    gr.Markdown(f"**Q{i+1}:** {q['question']}")
+                                    user_ans = gr.Textbox(label="Your Answer", placeholder="Type here...")
+                                    check_btn = gr.Button("Check Answer", size="sm")
+                                    feedback = gr.Markdown()
+                                    
+                                    def check_fitb(ans, correct=q['answer']):
+                                        if not ans: return "⚠️ Please type an answer."
+                                        if ans.strip().lower() == correct.strip().lower(): return "✅ **Correct!**"
+                                        return f"❌ **Incorrect.** The correct answer is: {correct}"
+                                        
+                                    check_btn.click(fn=check_fitb, inputs=[user_ans], outputs=[feedback])
+                                
+                                elif f_type == "Sentence Reorder":
+                                    gr.Markdown(f"**Q{i+1} (Rearrange):**\n*{q['scrambled']}*")
+                                    user_ans = gr.Textbox(label="Correct Order", placeholder="Type the full sentence here...")
+                                    check_btn = gr.Button("Check Answer", size="sm")
+                                    feedback = gr.Markdown()
+                                    
+                                    def check_reorder(ans, correct=q['correct_order']):
+                                        if not ans: return "⚠️ Please type an answer."
+                                        clean_ans = re.sub(r'[^a-z0-9äöüß]', '', ans.lower())
+                                        clean_correct = re.sub(r'[^a-z0-9äöüß]', '', correct.lower())
+                                        if clean_ans == clean_correct: return "✅ **Correct!**"
+                                        return f"❌ **Incorrect.** The correct sentence is:\n{correct}"
+                                        
+                                    check_btn.click(fn=check_reorder, inputs=[user_ans], outputs=[feedback])
+                                
+                                elif f_type == "Free QA":
+                                    gr.Markdown(f"**Q{i+1}:** {q['question']}")
+                                    user_ans = gr.Textbox(label="Your Answer", lines=2)
+                                    show_btn = gr.Button("Show Suggested Answer", size="sm")
+                                    feedback = gr.Markdown(visible=False)
+                                    
+                                    def show_ans(correct=q['suggested_answer']):
+                                        return gr.update(value=f"**Suggested Answer:** {correct}", visible=True)
+                                        
+                                    show_btn.click(fn=show_ans, outputs=[feedback])
+
+                with gr.TabItem("⚙️ Raw Data (JSON)"):
+                    output_json = gr.Code(label="Generated JSON Output", language="json")
 
-    btn.click(generate_expansion, [form_choice, user_req], [quiz_state, form_state])
+    generate_btn.click(
+        fn=generate_expansion,
+        inputs=[form_choice, user_requirement],
+        outputs=[quiz_state, form_state, output_json]
+    )
 
-demo.launch(server_name="0.0.0.0", server_port=7860)
+if __name__ == "__main__":
+    demo.launch(server_name="0.0.0.0", server_port=7860)