Create app.py

app.py

@@ -0,0 +1,56 @@
+import gradio as gr
+import pandas as pd
+from transformers import AutoTokenizer, AutoModelForCausalLM
+import torch
+
+# 모델과 토크나이저 로드
+model_name = "meta-llama/Meta-Llama-3.1-8B"
+tokenizer = AutoTokenizer.from_pretrained(model_name)
+model = AutoModelForCausalLM.from_pretrained(model_name, torch_dtype=torch.float16, device_map="auto")
+
+# KMMLU 데이터셋 로드 (예시)
+df = pd.read_csv("kmmlu_sample.csv")
+
+def evaluate_model(question, choices):
+    prompt = f"질문: {question}\n\n선택지:\n"
+    for i, choice in enumerate(choices):
+        prompt += f"{chr(65 + i)}. {choice}\n"
+    prompt += "\n답변:"
+
+    inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
+    with torch.no_grad():
+        outputs = model.generate(**inputs, max_new_tokens=1, temperature=0.0)
+    
+    answer = tokenizer.decode(outputs[0][-1:], skip_special_tokens=True).strip()
+    return answer
+
+def run_kmmlu_test(subject):
+    subject_df = df[df['subject'] == subject]
+    correct = 0
+    total = len(subject_df)
+    
+    results = []
+    for _, row in subject_df.iterrows():
+        question = row['question']
+        choices = [row['A'], row['B'], row['C'], row['D']]
+        correct_answer = row['answer']
+        
+        model_answer = evaluate_model(question, choices)
+        is_correct = model_answer == correct_answer
+        
+        if is_correct:
+            correct += 1
+        
+        results.append(f"질문: {question}\n모델 답변: {model_answer}\n정답: {correct_answer}\n정확도: {'맞음' if is_correct else '틀림'}\n")
+    
+    accuracy = correct / total
+    summary = f"주제: {subject}\n정확도: {accuracy:.2%} ({correct}/{total})\n\n"
+    return summary + "\n".join(results)
+
+subjects = df['subject'].unique().tolist()
+
+iface = gr.Interface(
+    fn=run_kmmlu_test,
+    inputs=gr.Dropdown(choices=subjects, label="주제 선택"),
+    outputs="text",
+    title="Llama 3.1을 이용한