feat: 新增情感分析和聊天助手应用，重构主应用

新增gradio_app.py实现电影评论情感分析功能
新增app02.py实现基于SmolLM2-135M的聊天助手
重构app.py为模型对比测试脚本
新增test.py实现LoRA微调测试

app.py

@@ -1,137 +1,23 @@
-import torch
-import gradio as gr
-from transformers import AutoModelForCausalLM, AutoTokenizer
-import threading
-import time
+from transformers import pipeline
 
-# Global variables for model and tokenizer
-model = None
-tokenizer = None
-model_loaded = False
-checkpoint = "HuggingFaceTB/SmolLM2-135M"
-device = "cuda" if torch.cuda.is_available() else "cpu"
+safe_pipe = pipeline(
+    "text-generation",
+    model="meta-llama/Llama-2-7b-chat-hf",
+    torch_dtype="auto",
+    device_map="auto"
+)
 
-def load_model():
-    """Load the model and tokenizer"""
-    global model, tokenizer, model_loaded
-    try:
-        print("Loading model...")
-        tokenizer = AutoTokenizer.from_pretrained(checkpoint)
-        model = AutoModelForCausalLM.from_pretrained(checkpoint).to(device)
-        model_loaded = True
-        print("Model loaded successfully!")
-    except Exception as e:
-        print(f"Error loading model: {e}")
-        model_loaded = False
+naive_pipe = pipeline(
+    "text-generation",
+    model="microsoft/DialoGPT-medium",
+    torch_dtype="auto",
+    device_map="auto"
+)
 
-def generate_response(message, history, temperature, max_length, top_p):
-    """Generate response from the model"""
-    global model, tokenizer, model_loaded
-    
-    if not model_loaded:
-        return "模型尚未加载完成，请稍等..."
-    
-    try:
-        # Tokenize input
-        inputs = tokenizer.encode(message, return_tensors="pt").to(device)
-        
-        # Generate
-        with torch.no_grad():
-            outputs = model.generate(
-                inputs,
-                max_length=max_length,
-                temperature=temperature,
-                top_p=top_p,
-                do_sample=True,
-                pad_token_id=tokenizer.eos_token_id
-            )
-        
-        # Decode response
-        response = tokenizer.decode(outputs[0], skip_special_tokens=True)
-        return response.strip()
-        
-    except Exception as e:
-        return f"生成回复时出错: {str(e)}"
+safe_out = safe_pipe(prompt, max_new_tokens=100, do_sample=False)[0]["generated_text"]
+naive_out = naive_pipe(prompt, max_new_tokens=100, do_sample=False)[0]["generated_text"]
 
-def chat_interface(message, history, temperature, max_length, top_p):
-    """Chat interface for Gradio"""
-    response = generate_response(message, history, temperature, max_length, top_p)
-    history.append([message, response])
-    return "", history
-
-# Load model in background
-loading_thread = threading.Thread(target=load_model)
-loading_thread.start()
-
-# Create Gradio interface
-with gr.Blocks(title="AI Chat Assistant") as demo:
-    gr.Markdown("# 🤖 AI Chat Assistant")
-    gr.Markdown("基于 SmolLM2-135M 模型的聊天助手")
-    
-    with gr.Row():
-        with gr.Column(scale=3):
-            chatbot = gr.Chatbot(
-                value=[],
-                height=500,
-                show_label=False
-            )
-            
-            with gr.Row():
-                msg = gr.Textbox(
-                    placeholder="输入您的消息...",
-                    show_label=False,
-                    scale=4
-                )
-                send_btn = gr.Button("发送", scale=1)
-            
-            clear_btn = gr.Button("清空对话")
-        
-        with gr.Column(scale=1):
-            gr.Markdown("### 参数设置")
-            temperature = gr.Slider(
-                minimum=0.1,
-                maximum=2.0,
-                value=0.7,
-                step=0.1,
-                label="Temperature"
-            )
-            max_length = gr.Slider(
-                minimum=100,
-                maximum=2000,
-                value=1000,
-                step=100,
-                label="最大长度"
-            )
-            top_p = gr.Slider(
-                minimum=0.1,
-                maximum=1.0,
-                value=0.95,
-                step=0.05,
-                label="Top-p"
-            )
-    
-    # Event handlers
-    send_btn.click(
-        chat_interface,
-        inputs=[msg, chatbot, temperature, max_length, top_p],
-        outputs=[msg, chatbot]
-    )
-    
-    msg.submit(
-        chat_interface,
-        inputs=[msg, chatbot, temperature, max_length, top_p],
-        outputs=[msg, chatbot]
-    )
-    
-    clear_btn.click(
-        lambda: ([], ""),
-        outputs=[chatbot, msg]
-    )
-
-if __name__ == "__main__":
-    demo.launch(
-        server_name="0.0.0.0",
-        server_port=7860,
-        share=True,
-        show_error=True
-    )
+print("=== 安全对齐模型回答 ===")
+print(safe_out)
+print("\n=== 无对齐模型回答 ===")
+print(naive_out)

app02.py

@@ -0,0 +1,137 @@
+import torch
+import gradio as gr
+from transformers import AutoModelForCausalLM, AutoTokenizer
+import threading
+import time
+
+# Global variables for model and tokenizer
+model = None
+tokenizer = None
+model_loaded = False
+checkpoint = "HuggingFaceTB/SmolLM2-135M"
+device = "cuda" if torch.cuda.is_available() else "cpu"
+
+def load_model():
+    """Load the model and tokenizer"""
+    global model, tokenizer, model_loaded
+    try:
+        print("Loading model...")
+        tokenizer = AutoTokenizer.from_pretrained(checkpoint)
+        model = AutoModelForCausalLM.from_pretrained(checkpoint).to(device)
+        model_loaded = True
+        print("Model loaded successfully!")
+    except Exception as e:
+        print(f"Error loading model: {e}")
+        model_loaded = False
+
+def generate_response(message, history, temperature, max_length, top_p):
+    """Generate response from the model"""
+    global model, tokenizer, model_loaded
+    
+    if not model_loaded:
+        return "模型尚未加载完成，请稍等..."
+    
+    try:
+        # Tokenize input
+        inputs = tokenizer.encode(message, return_tensors="pt").to(device)
+        
+        # Generate
+        with torch.no_grad():
+            outputs = model.generate(
+                inputs,
+                max_length=max_length,
+                temperature=temperature,
+                top_p=top_p,
+                do_sample=True,
+                pad_token_id=tokenizer.eos_token_id
+            )
+        
+        # Decode response
+        response = tokenizer.decode(outputs[0], skip_special_tokens=True)
+        return response.strip()
+        
+    except Exception as e:
+        return f"生成回复时出错: {str(e)}"
+
+def chat_interface(message, history, temperature, max_length, top_p):
+    """Chat interface for Gradio"""
+    response = generate_response(message, history, temperature, max_length, top_p)
+    history.append([message, response])
+    return "", history
+
+# Load model in background
+loading_thread = threading.Thread(target=load_model)
+loading_thread.start()
+
+# Create Gradio interface
+with gr.Blocks(title="AI Chat Assistant") as demo:
+    gr.Markdown("# 🤖 AI Chat Assistant")
+    gr.Markdown("基于 SmolLM2-135M 模型的聊天助手")
+    
+    with gr.Row():
+        with gr.Column(scale=3):
+            chatbot = gr.Chatbot(
+                value=[],
+                height=500,
+                show_label=False
+            )
+            
+            with gr.Row():
+                msg = gr.Textbox(
+                    placeholder="输入您的消息...",
+                    show_label=False,
+                    scale=4
+                )
+                send_btn = gr.Button("发送", scale=1)
+            
+            clear_btn = gr.Button("清空对话")
+        
+        with gr.Column(scale=1):
+            gr.Markdown("### 参数设置")
+            temperature = gr.Slider(
+                minimum=0.1,
+                maximum=2.0,
+                value=0.7,
+                step=0.1,
+                label="Temperature"
+            )
+            max_length = gr.Slider(
+                minimum=100,
+                maximum=2000,
+                value=1000,
+                step=100,
+                label="最大长度"
+            )
+            top_p = gr.Slider(
+                minimum=0.1,
+                maximum=1.0,
+                value=0.95,
+                step=0.05,
+                label="Top-p"
+            )
+    
+    # Event handlers
+    send_btn.click(
+        chat_interface,
+        inputs=[msg, chatbot, temperature, max_length, top_p],
+        outputs=[msg, chatbot]
+    )
+    
+    msg.submit(
+        chat_interface,
+        inputs=[msg, chatbot, temperature, max_length, top_p],
+        outputs=[msg, chatbot]
+    )
+    
+    clear_btn.click(
+        lambda: ([], ""),
+        outputs=[chatbot, msg]
+    )
+
+if __name__ == "__main__":
+    demo.launch(
+        server_name="0.0.0.0",
+        server_port=7860,
+        share=True,
+        show_error=True
+    )

gradio_app.py

@@ -0,0 +1,101 @@
+import gradio as gr
+from transformers import pipeline
+
+def load_model():
+    try:
+        print("[INFO] 开始加载模型...")
+        # 使用预训练的情感分析模型
+        model_name = "distilbert-base-uncased-finetuned-sst-2-english"
+        print(f"[INFO] 加载预训练模型: {model_name}")
+        
+        pipe = pipeline(
+            "sentiment-analysis",
+            model=model_name,
+            device=-1  # 使用CPU
+        )
+        print("[INFO] Pipeline创建完成")
+        return pipe
+    except Exception as e:
+        print(f"[ERROR] 模型加载失败: {str(e)}")
+        raise e
+
+def analyze_sentiment(text):
+    try:
+        print(f"[INFO] 收到文本分析请求: {text}")
+        result = pipe(text)
+        print(f"[INFO] 模型返回结果: {result}")
+        sentiment = result[0]['label']
+        confidence = result[0]['score']
+        response = f"情感类别: {sentiment}\n置信度: {confidence:.4f}"
+        print(f"[INFO] 返回分析结果: {response}")
+        return response
+    except Exception as e:
+        error_msg = f"错误: {str(e)}"
+        print(f"[ERROR] 分析过程出错: {error_msg}")
+        return error_msg
+
+print("[INFO] 正在加载模型，请稍候...")
+pipe = load_model()
+
+with gr.Blocks() as demo:
+    gr.Markdown("# 电影评论情感分析")
+    
+    with gr.Row():
+        text_input = gr.Textbox(
+            label="请输入电影评论",
+            placeholder="例如: I absolutely love this movie!",
+            lines=3
+        )
+    
+    with gr.Row():
+        analyze_btn = gr.Button("分析情感")
+    
+    with gr.Row():
+        output = gr.Textbox(label="分析结果")
+    
+    analyze_btn.click(
+        fn=analyze_sentiment,
+        inputs=text_input,
+        outputs=output
+    )
+    
+    gr.Examples(
+        examples=[
+            "I absolutely love this movie!",
+            "This film is terrible and a waste of time.",
+            "The acting was good but the story was predictable."
+        ],
+        inputs=text_input
+    )
+
+def check_network():
+    import socket
+    try:
+        # 测试与google dns的连接
+        socket.create_connection(("8.8.8.8", 53), timeout=3)
+        return True
+    except OSError:
+        return False
+
+if __name__ == "__main__":
+    try:
+        print("[INFO] 检查网络连接...")
+        if check_network():
+            print("[INFO] 网络连接正常")
+        else:
+            print("[WARNING] 网络连接可能不稳定，这可能会影响模型加载和公共URL访问")
+        
+        print("[INFO] 启动Gradio界面...")
+        demo.queue()  # 启用队列处理请求
+        server = demo.launch(
+            server_name="0.0.0.0",  # 允许外部访问
+            server_port=7860,      # 指定端口
+            share=True,            # 创建公共URL
+            show_error=True,       # 显示详细错误信息
+            show_api=False,        # 不显示API文档
+            favicon_path=None      # 禁用favicon请求
+        )
+        print(f"[INFO] Gradio服务器状态: {server}")
+    except Exception as e:
+        print(f"[ERROR] Gradio启动失败: {str(e)}")
+        raise e

test.py

@@ -0,0 +1,133 @@
+# =========================================================
+# 0. 一键安装依赖
+# =========================================================
+!pip install -q -U bitsandbytes         # 升级 4-bit 支持
+!pip install -q transformers datasets peft accelerate evaluate
+
+# =========================================================
+# 1. 必要的 import
+# =========================================================
+import torch
+from datasets import load_dataset
+from transformers import (
+    AutoTokenizer,
+    AutoModelForSequenceClassification,
+    TrainingArguments,
+    Trainer,
+    pipeline   # ← 新增这一行
+)
+from peft import LoraConfig, get_peft_model, TaskType
+
+# =========================================================
+# 2. 加载数据（IMDB 影评，50 k 条）
+# =========================================================
+ds = load_dataset("imdb")
+
+# =========================================================
+# 3. 加载模型 + LoRA
+# =========================================================
+checkpoint = "distilbert-base-uncased"
+print(f"\n[INFO] 开始加载模型和tokenizer: {checkpoint}")
+tokenizer = AutoTokenizer.from_pretrained(checkpoint)
+print(f"[INFO] Tokenizer加载完成，词表大小: {len(tokenizer)}")
+
+# 如果 pad_token 不存在，补一个
+if tokenizer.pad_token is None:
+    tokenizer.pad_token = tokenizer.eos_token
+    print(f"[INFO] 设置pad_token为eos_token: {tokenizer.pad_token}")
+
+print("\n[INFO] 开始加载基础模型...")
+model = AutoModelForSequenceClassification.from_pretrained(
+    checkpoint,
+    num_labels=2
+)
+print(f"[INFO] 基础模型加载完成，参数量: {sum(p.numel() for p in model.parameters()):,}")
+
+print("\n[INFO] 配置LoRA参数...")
+lora_config = LoraConfig(
+    task_type=TaskType.SEQ_CLS,
+    r=8,
+    lora_alpha=32,
+    lora_dropout=0.1,
+    target_modules=["q_lin", "v_lin"]   # DistilBERT 的 Q/V 投影
+)
+print(f"[INFO] LoRA配置: rank={lora_config.r}, alpha={lora_config.lora_alpha}, dropout={lora_config.lora_dropout}")
+
+model = get_peft_model(model, lora_config)
+print("\n[INFO] LoRA参数统计:")
+model.print_trainable_parameters()  # 看看训练参数量
+
+# =========================================================
+# 4. 数据预处理
+# =========================================================
+print("\n[INFO] 开始数据预处理...")
+def tok(batch):
+    return tokenizer(
+        batch["text"],
+        truncation=True,
+        padding="max_length",
+        max_length=256
+    )
+
+print(f"[INFO] 原始数据集大小: 训练集={len(ds['train'])}, 测试集={len(ds['test'])}")
+ds_tok = ds.map(tok, batched=True, remove_columns=["text"])
+ds_tok = ds_tok.rename_column("label", "labels").with_format("torch")
+print("[INFO] 数据tokenize完成")
+
+# 为了演示更快，只取 5 k 训练 + 1 k 验证
+train_ds = ds_tok["train"].select(range(5000))
+eval_ds  = ds_tok["test"].select(range(1000))
+print(f"[INFO] 最终使用数据集大小: 训练集={len(train_ds)}, 验证集={len(eval_ds)}")
+
+# =========================================================
+# 5. 训练参数 & 启动
+# =========================================================
+args = TrainingArguments(
+    output_dir="distilbert-lora-imdb",
+    per_device_train_batch_size=16,
+    per_device_eval_batch_size=16,
+    num_train_epochs=2,
+    learning_rate=2e-4,
+    fp16=torch.cuda.is_available(),
+    logging_steps=50,
+    eval_strategy="epoch",     # ← 旧版用 eval_strategy
+    save_strategy="epoch",     # ← 旧版用 save_strategy
+    report_to="none"
+)
+
+trainer = Trainer(
+    model=model,
+    args=args,
+    train_dataset=train_ds,
+    eval_dataset=eval_ds,
+    tokenizer=tokenizer
+)
+
+print("\n[INFO] 开始训练...")
+result = trainer.train()
+print(f"\n[INFO] 训练完成！")
+print(f"[INFO] 训练损失: {result.training_loss:.4f}")
+print(f"[INFO] 训练时长: {result.metrics['train_runtime']:.2f}秒")
+
+# =========================================================
+# 6. 保存与推理示例（可选）
+# =========================================================
+trainer.save_model("distilbert-lora-imdb")
+tokenizer.save_pretrained("distilbert-lora-imdb")
+
+# 本地推理
+print("\n[INFO] 加载训练好的模型进行推理...")
+from peft import PeftModel
+base = AutoModelForSequenceClassification.from_pretrained(
+    checkpoint,
+    num_labels=2
+)
+model_loaded = PeftModel.from_pretrained(base, "distilbert-lora-imdb")
+print("[INFO] 模型加载完成")
+
+pipe = pipeline("text-classification", model=model_loaded, tokenizer=tokenizer)
+test_text = "I absolutely love this movie!"
+print(f"\n[INFO] 测试文本: {test_text}")
+result = pipe(test_text)
+print(f"[INFO] 预测结果: {result}")
+print(f"[INFO] 情感类别: {result[0]['label']}, 置信度: {result[0]['score']:.4f}")