diff --git a/dataloader/dataloader.py b/dataloader/dataloader.py
index d2318a3..740bb47 100644
--- a/dataloader/dataloader.py
+++ b/dataloader/dataloader.py
@@ -11,7 +11,16 @@
 from tqdm import tqdm
 import re
 import os
-
+from utils.config import (
+    hotpot_qa_path,
+    mwh_qa_path,
+    trivia_qa_path,
+    cbt_path,
+    math_path,
+    gsm8k_math,
+    arc_path,
+    mmlu_path
+)
 split_map = {
     "train": {
         "hotpot_qa": "train",
@@ -45,7 +54,7 @@ def sample(self, count: int):
 class DataloaderForHotpotQA(BaseDataloader):
 
     def __init__(
-        self, dataset: str = "hotpot_qa", name: str = "distractor", split: str = "train"
+        self, dataset: str = hotpot_qa_path, name: str = "distractor", split: str = "train"
     ):
         self.dataset = dataset
         self.name = name
@@ -105,7 +114,7 @@ def sample_once(self):
 class DataloaderForMWHQA(BaseDataloader):
     def __init__(
         self,
-        dataset_path: str = "/home/test/test04/yuanjiarui/project/json_dataset/2MultiWikiQA",
+        dataset_path: str = mwh_qa_path,
         name: str = "",
         split: str = "train",
     ):
@@ -153,7 +162,7 @@ def sample_once(self):
 class DataloaderForTrivalQA(BaseDataloader):
     def __init__(
         self,
-        dataset: str = "/home/test/test04/yuanjiarui/project/huggingface_cache/trivia_qa_dataset",
+        dataset: str = trivia_qa_path,
         name: str = "rc",
         split: str = "train",
     ):
@@ -194,7 +203,7 @@ def sample_once(self):
 class DataloaderForCBT(BaseDataloader):
     def __init__(
         self,
-        dataset: str = "/home/test/test04/yuanjiarui/project/huggingface_cache/cbt_dataset",
+        dataset: str = cbt_path,
         name: str = "CN",
         split: str = "train",
     ):
@@ -230,7 +239,7 @@ def sample_once(self):
 class DataloaderForGSM8K(BaseDataloader):
     def __init__(
         self,
-        dataset: str = "/home/test/test04/yuanjiarui/project/huggingface_cache/gsm8k_dataset",
+        dataset: str = gsm8k_math,
         name: str = "main",
         split: str = "train",
     ):
@@ -263,7 +272,7 @@ def sample_once(self):
 class DataloaderForMATH(BaseDataloader):
     def __init__(
         self,
-        dataset: str = "/home/test/test04/yuanjiarui/project/json_dataset/MATH",
+        dataset: str = math_path,
         split="train",
     ):
         self.split = split
@@ -325,7 +334,7 @@ def sample_once(self):
 class DataloaderForARC(BaseDataloader):
     def __init__(
         self,
-        dataset: str = "/home/test/test04/yuanjiarui/project/huggingface_cache/arc_dataset",
+        dataset: str = arc_path,
         name: str = "ARC-Challenge",
         split: str = "train",
     ):
@@ -366,7 +375,7 @@ def sample_once(self):
 class DataloaderForMMLU(BaseDataloader):
     def __init__(
         self,
-        dataset: str = "/home/test/test04/yuanjiarui/project/huggingface_cache/mmlu_dataset",
+        dataset: str = mmlu_path,
         name: str = "all",
         split: str = "auxiliary_train",
     ):
diff --git a/train/dpo_recipes/arc.yaml b/train/dpo_recipes/arc.yaml
index 36b371b..ee8b437 100644
--- a/train/dpo_recipes/arc.yaml
+++ b/train/dpo_recipes/arc.yaml
@@ -19,9 +19,9 @@ sample_count: 0
 origin_dpo_yaml_path: ./alignment-handbook/recipes/Llama3-8b/arc_dpo/base.yaml
 mid_dpo_jsonl_root_path: ./results/arc_dpo
 mid_dpo_dataset_root_path: ./my_datasets/arc_dpo
-initial_dpo_min_value: 0
-initial_dpo_episilon: -100
-monte_sample_count: 10
+initial_dpo_min_value: 0.2
+initial_dpo_episilon: 0.45
+monte_sample_count: 10000
 cal_ppl: 1
 from_initial: False
 lambda1: -0.4
diff --git a/train/sft_dpo_recipes/arc.yaml b/train/sft_dpo_recipes/arc.yaml
index 4982fe7..40007f7 100644
--- a/train/sft_dpo_recipes/arc.yaml
+++ b/train/sft_dpo_recipes/arc.yaml
@@ -17,14 +17,14 @@ origin_sft_yaml_path: ./alignment-handbook/recipes/Llama3-8b/arc_sft_dpo/sft/bas
 mid_sft_jsonl_root_path: ./results/arc_sft_dpo/sft
 mid_sft_dataset_root_path: ./my_datasets/arc_sft_dpo/sft
 initial_episilon: 0.6
-sample_count: 10
+sample_count: 10000
 # dpo
 origin_dpo_yaml_path: ./alignment-handbook/recipes/Llama3-8b/arc_sft_dpo/dpo/base.yaml
 mid_dpo_jsonl_root_path: ./results/arc_sft_dpo/dpo
 mid_dpo_dataset_root_path: ./my_datasets/arc_sft_dpo/dpo
-initial_dpo_min_value: 0
-initial_dpo_episilon: -100
-monte_sample_count: 10
+initial_dpo_min_value: 0.2
+initial_dpo_episilon: 0.45
+monte_sample_count: 10000
 cal_ppl: 1
 lambda1: -0.5
 lambda2: 0.6
diff --git a/utils/config.py b/utils/config.py
index d68eff2..76a730c 100644
--- a/utils/config.py
+++ b/utils/config.py
@@ -7,3 +7,11 @@
 llama3_path_aistation = ""
 llama3_path_a800 = ""
 prompt_pool_path_a800 = ""
+hotpot_qa_path=""
+mwh_qa_path=""
+trivia_qa_path=""
+cbt_path=""
+math_path=""
+gsm8k_math=""
+arc_path=""
+mmlu_path=""
\ No newline at end of file