app.py

app.py

@@ -1,26 +1,25 @@
 import re
-import math
 import gradio as gr
 import torch
-from transformers import PreTrainedTokenizerFast, BartForConditionalGeneration
+from transformers import PreTrainedTokenizerFast, T5ForConditionalGeneration
 
-# ✅ 공개 KoBART 모델
-MODEL_NAME = "gogamza/kobart-base-v2"
+# ✅ KoT5 요약 모델
+MODEL_NAME = "psyche/KoT5-summarization"
 
 tokenizer = PreTrainedTokenizerFast.from_pretrained(MODEL_NAME)
-model = BartForConditionalGeneration.from_pretrained(MODEL_NAME)
+model = T5ForConditionalGeneration.from_pretrained(MODEL_NAME)
 
 # CPU 동적 양자화 적용
 try:
     model = torch.quantization.quantize_dynamic(
         model, {torch.nn.Linear}, dtype=torch.qint8
     )
-except Exception:
+except:
     pass
 
 model.eval()
 
-# ===== 유틸 함수 =====
+# ===== 유틸 =====
 def normalize_text(text: str) -> str:
     return re.sub(r"\s+", " ", text).strip()
 
@@ -55,17 +54,16 @@ def chunk_by_tokens(sentences, max_tokens=900):
         chunks.append(" ".join(cur))
     return chunks
 
-# ===== 요약 함수 =====
+# ===== 요약 =====
 def summarize_raw(text: str, min_len: int, max_len: int) -> str:
-    inputs = tokenizer([text], max_length=1024, truncation=True, return_tensors="pt")
+    input_ids = tokenizer.encode(text, return_tensors="pt", truncation=True, max_length=1024)
     with torch.no_grad():
         summary_ids = model.generate(
-            inputs["input_ids"],
+            input_ids,
             num_beams=4,
             min_length=min_len,
             max_length=max_len,
-            early_stopping=True,
-            no_repeat_ngram_size=3
+            early_stopping=True
         )
     return tokenizer.decode(summary_ids[0], skip_special_tokens=True)
 
@@ -77,8 +75,8 @@ def apply_style_prompt(text: str, mode: str, final: bool=False) -> str:
     else:
         inst = "다음 한국어 텍스트를 bullet 형태로 핵심만 요약하세요."
     if final:
-        inst += " 이 요약은 최종본입니다."
-    return f"{inst}\n\n[텍스트]\n{text}"
+        inst += " 원래 순서를 유지하며 문장 연결을 자연스럽게 하세요."
+    return f"{inst}\n\n{text}"
 
 def postprocess(summary: str, mode: str) -> str:
     s = summary.strip()
@@ -124,7 +122,7 @@ def ui_summarize(text, target_len, style):
     return summarize_long(text, int(target_len), mode)
 
 with gr.Blocks() as demo:
-    gr.Markdown("## 📝 KoBART 한국어 요약기 (공개 모델 gogamza/kobart-base-v2)")
+    gr.Markdown("## 📝 KoT5 한국어 요약기 (긴 문서 자동 분할 + 순서 보존)")
     with gr.Row():
         with gr.Column():
             input_text = gr.Textbox(label="원문 입력", lines=16)