更新训练配置

AITrain/train_npc_dialogue_lora.py

@@ -210,10 +210,11 @@ def create_lora_config():
         task_type=TaskType.CAUSAL_LM,
         target_modules=["q_proj", "k_proj", "v_proj", "gate_proj", "up_proj", "down_proj"],
         inference_mode=False,
-        r=8,  # 增加rank以提高表达能力
-        lora_alpha=8,  # alpha = 2 * r
-        lora_dropout=0.1,
-        modules_to_save=["lm_head", "embed_tokens"]
+        r=8,  # rank
+        lora_alpha=8,  # 降低alpha值以增加稳定性
+        lora_dropout=0.05,  # 降低dropout以减少不稳定性
+        # 移除modules_to_save以避免embed_tokens参数的NaN问题
+        # modules_to_save=["lm_head", "embed_tokens"]
     )
     return config
 
@@ -231,7 +232,10 @@ def prepare_dataset(data_path, tokenizer):
     
     # 转换为Dataset格式
     dataset = Dataset.from_list(data)
-    
+    #过滤 None 和空字符串（推荐）
+    dataset = dataset.filter(
+        lambda example: example.get("output") not in [None, ""]
+    )
     # 应用预处理函数
     tokenized_dataset = dataset.map(
         lambda example: process_func(example, tokenizer),
@@ -280,17 +284,32 @@ def train_lora_model(model_path, data_path, output_dir):
     # 3. 应用LoRA
     model = get_peft_model(model, lora_config)
     
-    # 4. 启用梯度计算
+    # 4. 数值稳定性初始化 - 初始化LoRA权重
+    for name, param in model.named_parameters():
+        if param.requires_grad:
+            if 'lora_A' in name:
+                # LoRA A矩阵使用正态分布初始化
+                torch.nn.init.normal_(param, mean=0.0, std=0.01)
+            elif 'lora_B' in name:
+                # LoRA B矩阵初始化为0
+                torch.nn.init.zeros_(param)
+            
+            # 检查初始化后是否有异常值
+            if torch.isnan(param).any() or torch.isinf(param).any():
+                logger.error(f"Abnormal values detected in parameter {name} after initialization")
+                torch.nn.init.normal_(param, mean=0.0, std=0.001)
+    
+    # 5. 启用梯度计算
     for param in model.parameters():
         if param.requires_grad:
             param.requires_grad_(True)
     
     model.config.use_cache = False  # 关闭缓存以节省显存
     
-    # 5. 准备数据集
+    # 6. 准备数据集
     train_preparedataset = prepare_dataset(data_path, tokenizer)
     
-    # 6. 配置训练参数 - 针对3080显卡优化
+    # 7. 配置训练参数 - 针对3080显卡优化
     training_args = TrainingArguments(
         output_dir=output_dir,
         per_device_train_batch_size=2,  # 减小batch size
@@ -298,9 +317,9 @@ def train_lora_model(model_path, data_path, output_dir):
         logging_steps=10,
         num_train_epochs=3,  # 增加训练轮数以充分学习角色特征
         save_steps=50,
-        learning_rate=1e-5,  # 降低学习率以增加稳定性
+        learning_rate=5e-6,  # 进一步降低学习率
         warmup_ratio=0.1,
-        max_grad_norm=1.0,   # 保持梯度裁剪
+        max_grad_norm=0.5,   # 更严格的梯度裁剪
         save_on_each_node=True,
         gradient_checkpointing=True,
         gradient_checkpointing_kwargs={"use_reentrant": True},
@@ -310,6 +329,8 @@ def train_lora_model(model_path, data_path, output_dir):
         bf16=True,  # 显式启用bf16以匹配模型加载类型
         #fp16=False, # 确保fp16被禁用
         save_total_limit=3,  # 只保留最新的3个检查点
+        adam_epsilon=1e-8,  # 增加数值稳定性
+        weight_decay=0.01,  # 添加权重衰减
     )
     
     #添加swan监测
@@ -322,7 +343,7 @@ def train_lora_model(model_path, data_path, output_dir):
     # 创建梯度监控回调
     gradient_monitor = GradientMonitorCallback()
     
-    # 7. 创建训练器
+    # 8. 创建训练器
     trainer = Trainer(
         model=model,
         args=training_args,
@@ -331,7 +352,7 @@ def train_lora_model(model_path, data_path, output_dir):
         callbacks=[swanlab_callback, gradient_monitor] # 添加梯度监控回调
     )
     
-    # 8. 开始训练
+    # 9. 开始训练
     print("Starting training...")
     logger.info("Starting training...")
     
@@ -344,7 +365,7 @@ def train_lora_model(model_path, data_path, output_dir):
         logger.error(traceback.format_exc())
         raise
     
-    # 9. 保存最终模型
+    # 10. 保存最终模型
     final_output_dir = os.path.join(output_dir, "final_model")
     trainer.save_model(final_output_dir)
     tokenizer.save_pretrained(final_output_dir)