diff --git a/ML/Training/Runner/runner_main.cpp b/ML/Training/Runner/runner_main.cpp
index c29f7e2..a8202c1 100644
--- a/ML/Training/Runner/runner_main.cpp
+++ b/ML/Training/Runner/runner_main.cpp
@@ -360,20 +360,18 @@ int main(int argc, char** argv)
 		return runServer(argv[2], argv[3], argv[4]);
 	}
 
-	// 默认单次模式
-	std::string dataDir = exeDir + "\\..\\..\\nmWTAI-ML\\data\\temp";
-
-	std::string datasetPath = dataDir + "\\dataset.bin";
-	std::string paramsPath  = dataDir + "\\params.bin";
-	std::string resultPath  = dataDir + "\\result.bin";
-	std::string dllPath     = exeDir + "\\HX_NWTM.dll";
-	std::string licPath     = exeDir + "\\..\\..\\..\\Bin\\Res\\license\\HXNWTM_license.dat";
-
-	if (argc >= 2) datasetPath = argv[1];
-	if (argc >= 3) paramsPath  = argv[2];
-	if (argc >= 4) resultPath  = argv[3];
-	if (argc >= 5) dllPath     = argv[4];
-	if (argc >= 6) licPath     = argv[5];
+	// Single-run mode requires explicit paths. Do not fall back to exeDir\\HX_NWTM.dll,
+	// because Runner and HX_NWTM.dll must be built against the same PEBI header version.
+	if (argc < 6) {
+		std::cerr << "Usage: runner.exe <dataset.bin> <params.bin> <result.bin> <HX_NWTM.dll> <license.dat>\n";
+		return 2;
+	}
+
+	std::string datasetPath = argv[1];
+	std::string paramsPath  = argv[2];
+	std::string resultPath  = argv[3];
+	std::string dllPath     = argv[4];
+	std::string licPath     = argv[5];
 
 	if (!fileExistsA(datasetPath)) { std::cerr << "ERROR: dataset not found\n"; return 10; }
 	if (!fileExistsA(paramsPath))  { std::cerr << "ERROR: params not found\n";  return 11; }
diff --git a/ML/nmWTAI-ML/configs/data_gen_family_random_v2_q.yaml b/ML/nmWTAI-ML/configs/data_gen_family_random_v2_q.yaml
index 108c29e..56487a5 100644
--- a/ML/nmWTAI-ML/configs/data_gen_family_random_v2_q.yaml
+++ b/ML/nmWTAI-ML/configs/data_gen_family_random_v2_q.yaml
@@ -12,7 +12,7 @@ paths:
 cpp:
   training_exe: "../Training/Release/training.exe"
   runner_exe: "../Training/Release/runner.exe"
-  hx_dll: "../Training/Release/HX_NWTM.dll"
+  hx_dll: "../../3rd/Pebi/V1/bin/HX_NWTM.dll"
   license_dat: "../../Bin/Res/license/HXNWTM_license.dat"
 
 dataset_runtime:
@@ -26,7 +26,7 @@ streaming_hdf5:  # HDF5 样本文件流式写入设置
   compression: null  # HDF5 压缩方式，null 表示不压缩
 
 parallel:  # 并行样本生成设置
-  n_workers: 36  # 并行生成样本的工作进程数量
+  n_workers: 12  # 并行生成样本的工作进程数量
   max_in_flight: 48  # 同时提交但尚未完成的最大任务数
   checkpoint_every_n: 5000  # 每生成多少条样本输出一次检查点
   checkpoint_every_sec: 180.0  # 每隔多少秒输出一次检查点
diff --git a/ML/nmWTAI-ML/requirements.txt b/ML/nmWTAI-ML/requirements.txt
index efb0bd4..173a7c5 100644
--- a/ML/nmWTAI-ML/requirements.txt
+++ b/ML/nmWTAI-ML/requirements.txt
@@ -7,3 +7,4 @@ numpy==2.4.3
 PyYAML==6.0.3
 scikit-learn==1.8.0
 torch==2.11.0
+tqdm==4.67.3
diff --git a/ML/nmWTAI-ML/src/data/dataset_generation.py b/ML/nmWTAI-ML/src/data/dataset_generation.py
index 779afd6..658b3e4 100644
--- a/ML/nmWTAI-ML/src/data/dataset_generation.py
+++ b/ML/nmWTAI-ML/src/data/dataset_generation.py
@@ -433,7 +433,14 @@ def _worker_simulate_parallel(args):
             _RESULT_BIN.unlink()
 
         # 每个 worker 使用独立临时目录，避免并行运行时 params.bin/result.bin 相互覆盖。
-        cmd = [str(cfg.runner_exe), str(cfg.dataset_bin), str(_PARAMS_BIN), str(_RESULT_BIN)]
+        cmd = [
+            str(cfg.runner_exe),
+            str(cfg.dataset_bin),
+            str(_PARAMS_BIN),
+            str(_RESULT_BIN),
+            str(cfg.hx_dll),
+            str(cfg.license_dat),
+        ]
         result = subprocess.run(
             cmd,
             cwd=str(cfg.runner_exe.parent),
@@ -472,8 +479,8 @@ def _worker_simulate_parallel(args):
         sec = int(np.clip(int(sch.sectionIndex), 1, max(len(sch.timeQ), 1)))
         enc = encode_schedule_to_timegrid(
             cfg,
-            sectionIndex=sec,
-            timeQ=sch.timeQ,
+            section_index=sec,
+            time_q=sch.timeQ,
             q=sch.q,
             n_sections=len(sch.timeQ),
         )