Upload app.py with huggingface_hub

app.py

@@ -141,6 +141,7 @@ CRITICAL RULES:
 2. Separate the prompt context. Do NOT penalize or critique the user for constraints that were mentioned in the [INTERVIEW QUESTION].
 3. You MUST generate an Example Answer at the very end. Keep it extremely short.
 4. If the candidate's answer is very short, vague, or a variation of "I don't know", you MUST give a Grade of 1/10 and explicitly state they failed to provide an answer in the Cons.
+5. A concise answer is NOT a bad answer. Judge by correctness and relevance, NOT by length.
 
 GRADING SCALE (follow this strictly):
 - 9-10: The answer is correct, demonstrates clear understanding, and covers the key points. It does NOT need to be perfect or exhaustive. A real interviewer would be impressed.
@@ -150,6 +151,11 @@ GRADING SCALE (follow this strictly):
 
 IMPORTANT: Only list a Con if it is a genuine mistake or a significant missing concept. Do NOT list "nice-to-have" extras or alternative approaches as Cons. If the answer is strong, give it a 9 or 10.
 
+GRADING EXAMPLES:
+- Question: "How would you secure a REST API?" Answer: "I'd use HTTPS, JWT authentication, input validation, and rate limiting." -> Grade: 9/10 (concise but covers all key points)
+- Question: "How would you secure a REST API?" Answer: "I'd add authentication and maybe some encryption." -> Grade: 5/10 (too vague, missing specifics)
+- Question: "How would you secure a REST API?" Answer: "I'm not really sure, probably a firewall." -> Grade: 2/10 (shows lack of understanding)
+
 You MUST output exactly this format and nothing else:
 Grade: [1-10]/10
 Pros:
@@ -169,6 +175,26 @@ Example Answer:
     outputs = generator(messages, max_new_tokens=400, temperature=0.15, do_sample=True)
     raw_feedback = outputs[0]['generated_text'][-1]['content']
     
+    # Safety net: prevent unreasonably low grades for real answers
+    match = re.search(r'Grade:\s*(\d+)', raw_feedback)
+    if match:
+        model_score = int(match.group(1))
+        word_count = len(candidate_answer.split())
+        
+        # If the answer has real substance, enforce a minimum floor
+        if word_count >= 40:
+            min_grade = 4
+        elif word_count >= 20:
+            min_grade = 3
+        elif word_count >= 8:
+            min_grade = 2
+        else:
+            min_grade = 1
+        
+        if model_score < min_grade:
+            adjusted_score = min_grade
+            raw_feedback = re.sub(r'Grade:\s*\d+', f'Grade: {adjusted_score}', raw_feedback)
+    
     score_html = create_circular_progress(raw_feedback)
     clean_feedback = re.sub(r'Grade:.*?\n', '', raw_feedback).strip()