feat(摘要生成): 升级摘要生成模型至Qwen2.5并优化提示工程

- 将摘要生成模型从Falconsai/text_summarization替换为Qwen/Qwen2.5-0.5B-Instruct
- 重构摘要生成逻辑，使用更灵活的提示模板和生成参数
- 修复app.py中commit消息显示格式问题
- 优化score_titles端点的代码格式

app.py

@@ -66,7 +66,7 @@ def check_hf_token(token):
                 # 构建消息内容
                 message = f"你好，{user_info.get('name')}，我已启动"
                 if commit_message:
-                    message += f"\n\n最近的Git提交: {commit_message}"
+                    message += f"\n\{commit_message}"
 
                 webhook_headers = {"Content-Type": "application/json"}
                 webhook_data = {
@@ -515,7 +515,11 @@ async def text_to_speech(
 
 
 @app.post("/text/score")
-async def score_titles(request_data: Dict[str, List[str]] = Body(..., description="Object containing titles list")):
+async def score_titles(
+    request_data: Dict[str, List[str]] = Body(
+        ..., description="Object containing titles list"
+    )
+):
     """
     为文章标题列表打分，返回标题和对应的分数。
 
@@ -524,19 +528,18 @@ async def score_titles(request_data: Dict[str, List[str]] = Body(..., descriptio
     try:
         # 获取标题列表
         titles = request_data.get("titles", [])
-        
+
         # 调用打分函数
         scores = score_article_titles(titles)
-        
+
         # 组合标题和分数
         results = [
-            {"title": title, "score": score}
-            for title, score in zip(titles, scores)
+            {"title": title, "score": score} for title, score in zip(titles, scores)
         ]
-        
+
         # 按分数倒序排序
         results.sort(key=lambda x: x["score"], reverse=True)
-        
+
         return {"results": results}
     except Exception as e:
         logger.error(f"Title scoring error: {e}")
@@ -547,7 +550,7 @@ async def score_titles(request_data: Dict[str, List[str]] = Body(..., descriptio
 async def generate_summary(
     text: str = Body(..., description="Text to summarize"),
     max_length: int = Body(300, description="Maximum length of the summary"),
-    min_length: int = Body(50, description="Minimum length of the summary")
+    min_length: int = Body(50, description="Minimum length of the summary"),
 ):
     """
     生成文本摘要。
@@ -559,7 +562,7 @@ async def generate_summary(
     try:
         # 调用摘要函数
         summary = summarize_text(text, max_length=max_length, min_length=min_length)
-        
+
         return {"summary": summary}
     except Exception as e:
         logger.error(f"Summarization error: {e}")

utils/summarization.py

@@ -1,22 +1,47 @@
-from transformers import pipeline
+from transformers import AutoModelForCausalLM, AutoTokenizer
+import torch
 
 def summarize_text(text, max_length=300, min_length=50):
     """
-    使用 Falconsai/text_summarization 生成摘要
-    :param text: 输入的文章内容（建议为英文）
-    :param max_length: 摘要最大长度（token数，约等于汉字数*0.5）
-    :param min_length: 摘要最小长度
+    使用 Qwen2.5-0.5B-Instruct 生成摘要
+    :param text: 输入的文章内容（支持中英文）
+    :param max_length: 摘要最大长度（汉字数）
+    :param min_length: 摘要最小长度（汉字数）
     :return: 生成的摘要
     """
-    # 加载Pipeline
-    summarizer = pipeline("summarization", model="Falconsai/text_summarization")
+    model_name = "Qwen/Qwen2.5-0.5B-Instruct"
+    
+    # 加载模型和分词器
+    model = AutoModelForCausalLM.from_pretrained(
+        model_name,
+        torch_dtype="auto",
+        device_map="auto"
+    )
+    tokenizer = AutoTokenizer.from_pretrained(model_name)
+    
+    # 构建提示
+    prompt = f"请将以下文章总结为{max_length}字以内的摘要：\n\n{text}"
+    messages = [{"role": "user", "content": prompt}]
+    
+    # 应用聊天模板
+    text = tokenizer.apply_chat_template(
+        messages,
+        tokenize=False,
+        add_generation_prompt=True
+    )
     
     # 生成摘要
-    result = summarizer(
-        text,
-        max_length=max_length,
-        min_length=min_length,
-        do_sample=False
+    model_inputs = tokenizer([text], return_tensors="pt").to(model.device)
+    generated_ids = model.generate(
+        **model_inputs,
+        max_new_tokens=max_length + 50,
+        temperature=0.3
     )
     
-    return result[0]['summary_text']
+    # 处理生成结果
+    generated_ids = [
+        output_ids[len(input_ids):] for input_ids, output_ids in zip(model_inputs.input_ids, generated_ids)
+    ]
+    
+    response = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
+    return response