diff --git a/README.md b/README.md
index 04b62aef..25e8d7dd 100644
--- a/README.md
+++ b/README.md
@@ -275,6 +275,11 @@ We also provide the raw data exported from Weights & Biases for the detailed res
 - TextVQA (textvqa)
   - TextVQA Validation (textvqa_val)
   - TextVQA Test (textvqa_test)
+- VCR-Wiki (vcr_wiki)
+  - vcr English easy mode (vcr_wiki_en_easy)
+  - vcr English hard mode (vcr_wiki_en_hard)
+  - vcr Chinese easy mode (vcr_wiki_zh_easy)
+  - vcr Chinese hard mode (vcr_wiki_zh_hard)
 - VizWizVQA (vizwiz_vqa)
   - VizWizVQA Validation (vizwiz_vqa_val)
   - VizWizVQA Test (vizwiz_vqa_test)
diff --git a/lmms_eval/tasks/vcr/_default_template_vcr_yaml b/lmms_eval/tasks/vcr/_default_template_vcr_yaml
deleted file mode 100644
index 8a9095a3..00000000
--- a/lmms_eval/tasks/vcr/_default_template_vcr_yaml
+++ /dev/null
@@ -1,31 +0,0 @@
-dataset_path: vcr-org/VCR-wiki-en-hard
-dataset_kwargs:
-  token: True
-task: "vcr_wiki_en_hard"
-test_split: test
-output_type: generate_until
-doc_to_visual: !function utils.vcr_doc_to_visual
-doc_to_text: !function utils.vcr_doc_to_text
-doc_to_target: "answer"
-generation_kwargs:
-  max_new_tokens: 120
-  temperature: 0
-  top_p: 0
-  num_beams: 1
-  do_sample: false
-# The return value of process_results will be used by metrics
-process_results: !function utils.vcr_en_process_results
-# Note that the metric name can be either a registed metric function (such as the case for GQA) or a key name returned by process_results
-metric_list:
-  - metric: mme_percetion_score
-    aggregation: !function utils.vcr_en_process_results
-    higher_is_better: true
-  - metric: mme_cognition_score
-    aggregation: !function utils.vcr_en_process_results
-    higher_is_better: true
-model_specific_prompt_kwargs:
-  default:
-    pre_prompt: ""
-    post_prompt: "What is the covered texts in the image? Please restore the covered texts without outputting the explanations."
-metadata:
-  - version: 0.0.1
diff --git a/lmms_eval/tasks/vcr/vcr_wiki_en_easy.yaml b/lmms_eval/tasks/vcr/vcr_wiki_en_easy.yaml
deleted file mode 100644
index 1aa2745d..00000000
--- a/lmms_eval/tasks/vcr/vcr_wiki_en_easy.yaml
+++ /dev/null
@@ -1,31 +0,0 @@
-dataset_path: vcr-org/VCR-wiki-en-easy
-dataset_kwargs:
-  token: True
-task: "vcr_wiki_en_easy"
-test_split: test
-output_type: generate_until
-doc_to_visual: !function utils.vcr_doc_to_visual
-doc_to_text: !function utils.vcr_doc_to_text
-doc_to_target: "answer"
-generation_kwargs:
-  max_new_tokens: 120
-  temperature: 0
-  top_p: 0
-  num_beams: 1
-  do_sample: false
-# The return value of process_results will be used by metrics
-process_results: !function utils.vcr_en_process_results
-# Note that the metric name can be either a registed metric function (such as the case for GQA) or a key name returned by process_results
-metric_list:
-  - metric: mme_percetion_score
-    aggregation: !function utils.vcr_en_process_results
-    higher_is_better: true
-  - metric: mme_cognition_score
-    aggregation: !function utils.vcr_en_process_results
-    higher_is_better: true
-model_specific_prompt_kwargs:
-  default:
-    pre_prompt: ""
-    post_prompt: "What is the covered texts in the image? Please restore the covered texts without outputting the explanations."
-metadata:
-  - version: 0.0.1
diff --git a/lmms_eval/tasks/vcr/vcr_wiki_en_hard.yaml b/lmms_eval/tasks/vcr/vcr_wiki_en_hard.yaml
deleted file mode 100644
index 8a9095a3..00000000
--- a/lmms_eval/tasks/vcr/vcr_wiki_en_hard.yaml
+++ /dev/null
@@ -1,31 +0,0 @@
-dataset_path: vcr-org/VCR-wiki-en-hard
-dataset_kwargs:
-  token: True
-task: "vcr_wiki_en_hard"
-test_split: test
-output_type: generate_until
-doc_to_visual: !function utils.vcr_doc_to_visual
-doc_to_text: !function utils.vcr_doc_to_text
-doc_to_target: "answer"
-generation_kwargs:
-  max_new_tokens: 120
-  temperature: 0
-  top_p: 0
-  num_beams: 1
-  do_sample: false
-# The return value of process_results will be used by metrics
-process_results: !function utils.vcr_en_process_results
-# Note that the metric name can be either a registed metric function (such as the case for GQA) or a key name returned by process_results
-metric_list:
-  - metric: mme_percetion_score
-    aggregation: !function utils.vcr_en_process_results
-    higher_is_better: true
-  - metric: mme_cognition_score
-    aggregation: !function utils.vcr_en_process_results
-    higher_is_better: true
-model_specific_prompt_kwargs:
-  default:
-    pre_prompt: ""
-    post_prompt: "What is the covered texts in the image? Please restore the covered texts without outputting the explanations."
-metadata:
-  - version: 0.0.1
diff --git a/lmms_eval/tasks/vcr/vcr_wiki_zh_easy.yaml b/lmms_eval/tasks/vcr/vcr_wiki_zh_easy.yaml
deleted file mode 100644
index ba888daf..00000000
--- a/lmms_eval/tasks/vcr/vcr_wiki_zh_easy.yaml
+++ /dev/null
@@ -1,31 +0,0 @@
-dataset_path: vcr-org/VCR-wiki-zh-easy
-dataset_kwargs:
-  token: True
-task: "vcr_wiki_zh_easy"
-test_split: test
-output_type: generate_until
-doc_to_visual: !function utils.vcr_doc_to_visual
-doc_to_text: !function utils.vcr_doc_to_text
-doc_to_target: "answer"
-generation_kwargs:
-  max_new_tokens: 120
-  temperature: 0
-  top_p: 0
-  num_beams: 1
-  do_sample: false
-# The return value of process_results will be used by metrics
-process_results: !function utils.vcr_zh_process_results
-# Note that the metric name can be either a registed metric function (such as the case for GQA) or a key name returned by process_results
-metric_list:
-  - metric: mme_percetion_score
-    aggregation: !function utils.vcr_zh_process_results
-    higher_is_better: true
-  - metric: mme_cognition_score
-    aggregation: !function utils.vcr_zh_process_results
-    higher_is_better: true
-model_specific_prompt_kwargs:
-  default:
-    pre_prompt: ""
-    post_prompt: "图像中被覆盖的文本是什么？请在不输出解释的情况下还原被覆盖的文本。"
-metadata:
-  - version: 0.0.1
diff --git a/lmms_eval/tasks/vcr/vcr_wiki_zh_hard.yaml b/lmms_eval/tasks/vcr/vcr_wiki_zh_hard.yaml
deleted file mode 100644
index b6dc0f7f..00000000
--- a/lmms_eval/tasks/vcr/vcr_wiki_zh_hard.yaml
+++ /dev/null
@@ -1,31 +0,0 @@
-dataset_path: vcr-org/VCR-wiki-zh-hard
-dataset_kwargs:
-  token: True
-task: "vcr_wiki_zh_hard"
-test_split: test
-output_type: generate_until
-doc_to_visual: !function utils.vcr_doc_to_visual
-doc_to_text: !function utils.vcr_doc_to_text
-doc_to_target: "answer"
-generation_kwargs:
-  max_new_tokens: 120
-  temperature: 0
-  top_p: 0
-  num_beams: 1
-  do_sample: false
-# The return value of process_results will be used by metrics
-process_results: !function utils.vcr_zh_process_results
-# Note that the metric name can be either a registed metric function (such as the case for GQA) or a key name returned by process_results
-metric_list:
-  - metric: mme_percetion_score
-    aggregation: !function utils.vcr_zh_process_results
-    higher_is_better: true
-  - metric: mme_cognition_score
-    aggregation: !function utils.vcr_zh_process_results
-    higher_is_better: true
-model_specific_prompt_kwargs:
-  default:
-    pre_prompt: ""
-    post_prompt: "图像中被覆盖的文本是什么？请在不输出解释的情况下还原被覆盖的文本。"
-metadata:
-  - version: 0.0.1
diff --git a/lmms_eval/tasks/vcr_wiki/_default_template_vcr_yaml b/lmms_eval/tasks/vcr_wiki/_default_template_vcr_yaml
new file mode 100644
index 00000000..37ab5e74
--- /dev/null
+++ b/lmms_eval/tasks/vcr_wiki/_default_template_vcr_yaml
@@ -0,0 +1,17 @@
+
+dataset_kwargs:
+  token: True
+output_type: generate_until
+doc_to_visual: !function utils.vcr_doc_to_visual
+doc_to_text: !function utils.vcr_doc_to_text
+doc_to_target: "answer"
+generation_kwargs:
+  max_new_tokens: 120
+  temperature: 0
+  top_p: 0
+  num_beams: 1
+  do_sample: false
+# The return value of process_results will be used by metrics
+# Note that the metric name can be either a registed metric function (such as the case for GQA) or a key name returned by process_results
+metadata:
+  - version: 0.0.1
\ No newline at end of file
diff --git a/lmms_eval/tasks/vcr/utils.py b/lmms_eval/tasks/vcr_wiki/utils.py
similarity index 63%
rename from lmms_eval/tasks/vcr/utils.py
rename to lmms_eval/tasks/vcr_wiki/utils.py
index 0079c8fd..8f3aca04 100644
--- a/lmms_eval/tasks/vcr/utils.py
+++ b/lmms_eval/tasks/vcr_wiki/utils.py
@@ -1,9 +1,5 @@
-from collections import defaultdict
 import os
 from difflib import SequenceMatcher as SM
-import datetime
-import json
-from lmms_eval.tasks._task_utils.file_utils import generate_submission_file
 import evaluate
 import logging
 import spacy
@@ -34,6 +30,21 @@
 }
 
 
+def fast_filter(answer_text):
+    if "I can't" in answer_text:
+        return True
+    elif "I cannot" in answer_text:
+        return True
+    elif "sorry" in answer_text.lower():
+        return True
+    if "无法" in answer_text:
+        return True
+    elif "抱歉" in answer_text:
+        return True
+    else:
+        return False
+
+
 def vcr_doc_to_visual(doc):
     return [doc["stacked_image"].convert("RGB"), doc["only_it_image"].convert("RGB")]
 
@@ -63,7 +74,7 @@ def tokenize(text, language):
     return [token.text for token in processed_text]
 
 
-def vcr_process_results_single(doc, result, language):
+def vcr_process_results_single(crossed_text, result, language):
     """
     Args:
         doc: a instance of the eval dataset
@@ -71,8 +82,21 @@ def vcr_process_results_single(doc, result, language):
     Returns:
         a dictionary with key: metric name (in this case mme score), value: metric value
     """
+
     assert language in ["en", "zh"], f"Language {language} is not supported."
-    crossed_text = doc["crossed_text"]
+
+    if fast_filter(result):
+        return {
+            "crossed_text": crossed_text,
+            "max_sim_val": 0,
+            "max_sim_string": "",
+            "precision": 0,
+            "recall": 0,
+            "f1": 0,
+            "jaccard": 0,
+            "rouge1": 0,
+            "exact_match": 0,
+        }
     tokens_result = tokenize(result, language)
     tokens_crossed_text = tokenize(crossed_text, language)
 
@@ -150,10 +174,26 @@ def vcr_en_process_results(doc, results):
         a dictionary with key: metric name (in this case mme score), value: metric value
     """
     assert len(results) == 2, f"Expected 2 results, got {len(results)}"
-    output = {
-        "res_stacked_image": vcr_process_results_single(doc, results[0], "en"),
-        "res_only_it_image": vcr_process_results_single(doc, results[1], "en"),
-    }
+    output = {}
+    for i in range(len(doc["crossed_text"])):
+        res_stacked_image_results = vcr_process_results_single(
+            doc["crossed_text"][i], results[0], "en"
+        )
+        res_only_image_results = vcr_process_results_single(
+            doc["crossed_text"][i], results[1], "en"
+        )
+        output.update(
+            {
+                f"res_stacked_image__{k}___{i}": v
+                for k, v in res_stacked_image_results.items()
+            }
+        )
+        output.update(
+            {
+                f"res_only_it_image__{k}___{i}": v
+                for k, v in res_only_image_results.items()
+            }
+        )
     return output
 
 
@@ -166,10 +206,26 @@ def vcr_zh_process_results(doc, results):
         a dictionary with key: metric name (in this case mme score), value: metric value
     """
     assert len(results) == 2, f"Expected 2 results, got {len(results)}"
-    output = {
-        "res_stacked_image": vcr_process_results_single(doc, results[0], "zh"),
-        "res_only_it_image": vcr_process_results_single(doc, results[1], "zh"),
-    }
+    output = {}
+    for i in range(len(doc["crossed_text"])):
+        res_stacked_image_results = vcr_process_results_single(
+            doc["crossed_text"][i], results[0], "zh"
+        )
+        res_only_image_results = vcr_process_results_single(
+            doc["crossed_text"][i], results[1], "zh"
+        )
+        output.update(
+            {
+                f"res_stacked_image__{k}___{i}": v
+                for k, v in res_stacked_image_results.items()
+            }
+        )
+        output.update(
+            {
+                f"res_only_it_image__{k}___{i}": v
+                for k, v in res_only_image_results.items()
+            }
+        )
     return output
 
 
@@ -180,36 +236,29 @@ def vcr_aggregate_results(results):
     Returns:
         A dictionary of dictionary of float, where the outer dictionary has keys "res_stacked_image" and "res_only_it_image"
     """
-
     output = {
-        "res_stacked_image": {
-            "max_sim_val": 0,
-            "precision": 0,
-            "recall": 0,
-            "f1": 0,
-            "jaccard": 0,
-            "rouge1": 0,
-        },
-        "res_only_it_image": {
-            "max_sim_val": 0,
-            "precision": 0,
-            "recall": 0,
-            "f1": 0,
-            "jaccard": 0,
-            "rouge1": 0,
-        },
+        "res_stacked_image__precision": 0,
+        "res_stacked_image__recall": 0,
+        "res_stacked_image__f1": 0,
+        "res_stacked_image__jaccard": 0,
+        "res_stacked_image__rouge1": 0,
+        "res_stacked_image__exact_match": 0,
+        "res_only_it_image__precision": 0,
+        "res_only_it_image__recall": 0,
+        "res_only_it_image__f1": 0,
+        "res_only_it_image__jaccard": 0,
+        "res_only_it_image__rouge1": 0,
+        "res_only_it_image__exact_match": 0,
     }
-    for target_domain in output.keys():
-        for target_metric_name in output[target_domain].keys():
-            score = 0
-            count = 0
-            for inner_dict in results:
-                for inner_key, inner_value in inner_dict.items():
-                    if inner_key == target_domain:
-                        for blank_id, blank_metrics in inner_value.items():
-                            for metric_name, metric_value in blank_metrics.items():
-                                if metric_name == target_metric_name:
-                                    score += metric_value
-                                    count += 1
-            output[target_domain][target_metric_name] = score / count
+
+    for output_key in output.keys():
+        count = 0
+        query_domain, query_metric_name = output_key.split("__")
+        for inner_dict in results:
+            for inner_key, inner_value in inner_dict.items():
+                key_domain, key_metric_name, _ = inner_key.split("__")
+                if query_domain == key_domain and query_metric_name == key_metric_name:
+                    output[output_key] += inner_value
+                    count += 1
+        output[output_key] /= count
     return output
diff --git a/lmms_eval/tasks/vcr_wiki/vcr_wiki_en_easy_100.yaml b/lmms_eval/tasks/vcr_wiki/vcr_wiki_en_easy_100.yaml
new file mode 100644
index 00000000..7df8eb38
--- /dev/null
+++ b/lmms_eval/tasks/vcr_wiki/vcr_wiki_en_easy_100.yaml
@@ -0,0 +1,16 @@
+"include": "_default_template_vcr_yaml"
+dataset_path: vcr-org/VCR-wiki-en-easy-test
+task: "vcr_wiki_en_easy"
+test_split: train[:100]
+process_results: !function utils.vcr_en_process_results
+metric_list:
+  - metric: vcr_percetion_score
+    aggregation: !function utils.vcr_en_process_results
+    higher_is_better: true
+  - metric: vcr_cognition_score
+    aggregation: !function utils.vcr_en_process_results
+    higher_is_better: true
+model_specific_prompt_kwargs:
+  default:
+    pre_prompt: ""
+    post_prompt: "What is the covered texts in the image? Please restore the covered texts without outputting the explanations."
\ No newline at end of file
diff --git a/lmms_eval/tasks/vcr_wiki/vcr_wiki_en_easy_500.yaml b/lmms_eval/tasks/vcr_wiki/vcr_wiki_en_easy_500.yaml
new file mode 100644
index 00000000..9079d2cf
--- /dev/null
+++ b/lmms_eval/tasks/vcr_wiki/vcr_wiki_en_easy_500.yaml
@@ -0,0 +1,16 @@
+"include": "_default_template_vcr_yaml"
+dataset_path: vcr-org/VCR-wiki-en-easy-test
+task: "vcr_wiki_en_easy"
+test_split: train[:500]
+process_results: !function utils.vcr_en_process_results
+metric_list:
+  - metric: vcr_percetion_score
+    aggregation: !function utils.vcr_en_process_results
+    higher_is_better: true
+  - metric: vcr_cognition_score
+    aggregation: !function utils.vcr_en_process_results
+    higher_is_better: true
+model_specific_prompt_kwargs:
+  default:
+    pre_prompt: ""
+    post_prompt: "What is the covered texts in the image? Please restore the covered texts without outputting the explanations."
\ No newline at end of file
diff --git a/lmms_eval/tasks/vcr_wiki/vcr_wiki_en_easy_5000.yaml b/lmms_eval/tasks/vcr_wiki/vcr_wiki_en_easy_5000.yaml
new file mode 100644
index 00000000..7a5b09e2
--- /dev/null
+++ b/lmms_eval/tasks/vcr_wiki/vcr_wiki_en_easy_5000.yaml
@@ -0,0 +1,16 @@
+"include": "_default_template_vcr_yaml"
+dataset_path: vcr-org/VCR-wiki-en-easy-test
+task: "vcr_wiki_en_easy"
+test_split: train
+process_results: !function utils.vcr_en_process_results
+metric_list:
+  - metric: vcr_percetion_score
+    aggregation: !function utils.vcr_en_process_results
+    higher_is_better: true
+  - metric: vcr_cognition_score
+    aggregation: !function utils.vcr_en_process_results
+    higher_is_better: true
+model_specific_prompt_kwargs:
+  default:
+    pre_prompt: ""
+    post_prompt: "What is the covered texts in the image? Please restore the covered texts without outputting the explanations."
\ No newline at end of file
diff --git a/lmms_eval/tasks/vcr_wiki/vcr_wiki_en_hard_100.yaml b/lmms_eval/tasks/vcr_wiki/vcr_wiki_en_hard_100.yaml
new file mode 100644
index 00000000..3f24be85
--- /dev/null
+++ b/lmms_eval/tasks/vcr_wiki/vcr_wiki_en_hard_100.yaml
@@ -0,0 +1,16 @@
+"include": "_default_template_vcr_yaml"
+dataset_path: vcr-org/VCR-wiki-en-hard-test
+task: "vcr_wiki_en_hard"
+test_split: train[:100]
+process_results: !function utils.vcr_en_process_results
+metric_list:
+  - metric: vcr_percetion_score
+    aggregation: !function utils.vcr_en_process_results
+    higher_is_better: true
+  - metric: vcr_cognition_score
+    aggregation: !function utils.vcr_en_process_results
+    higher_is_better: true
+model_specific_prompt_kwargs:
+  default:
+    pre_prompt: ""
+    post_prompt: "What is the covered texts in the image? Please restore the covered texts without outputting the explanations."
\ No newline at end of file
diff --git a/lmms_eval/tasks/vcr_wiki/vcr_wiki_en_hard_500.yaml b/lmms_eval/tasks/vcr_wiki/vcr_wiki_en_hard_500.yaml
new file mode 100644
index 00000000..c88c91fe
--- /dev/null
+++ b/lmms_eval/tasks/vcr_wiki/vcr_wiki_en_hard_500.yaml
@@ -0,0 +1,16 @@
+"include": "_default_template_vcr_yaml"
+dataset_path: vcr-org/VCR-wiki-en-hard-test
+task: "vcr_wiki_en_hard"
+test_split: train[:500]
+process_results: !function utils.vcr_en_process_results
+metric_list:
+  - metric: vcr_percetion_score
+    aggregation: !function utils.vcr_en_process_results
+    higher_is_better: true
+  - metric: vcr_cognition_score
+    aggregation: !function utils.vcr_en_process_results
+    higher_is_better: true
+model_specific_prompt_kwargs:
+  default:
+    pre_prompt: ""
+    post_prompt: "What is the covered texts in the image? Please restore the covered texts without outputting the explanations."
\ No newline at end of file
diff --git a/lmms_eval/tasks/vcr_wiki/vcr_wiki_en_hard_5000.yaml b/lmms_eval/tasks/vcr_wiki/vcr_wiki_en_hard_5000.yaml
new file mode 100644
index 00000000..aa5e3634
--- /dev/null
+++ b/lmms_eval/tasks/vcr_wiki/vcr_wiki_en_hard_5000.yaml
@@ -0,0 +1,16 @@
+"include": "_default_template_vcr_yaml"
+dataset_path: vcr-org/VCR-wiki-en-hard-test
+task: "vcr_wiki_en_hard"
+test_split: train
+process_results: !function utils.vcr_en_process_results
+metric_list:
+  - metric: vcr_percetion_score
+    aggregation: !function utils.vcr_en_process_results
+    higher_is_better: true
+  - metric: vcr_cognition_score
+    aggregation: !function utils.vcr_en_process_results
+    higher_is_better: true
+model_specific_prompt_kwargs:
+  default:
+    pre_prompt: ""
+    post_prompt: "What is the covered texts in the image? Please restore the covered texts without outputting the explanations."
\ No newline at end of file
diff --git a/lmms_eval/tasks/vcr_wiki/vcr_wiki_zh_easy_100.yaml b/lmms_eval/tasks/vcr_wiki/vcr_wiki_zh_easy_100.yaml
new file mode 100644
index 00000000..9ce02a93
--- /dev/null
+++ b/lmms_eval/tasks/vcr_wiki/vcr_wiki_zh_easy_100.yaml
@@ -0,0 +1,16 @@
+"include": "_default_template_vcr_yaml"
+dataset_path: vcr-org/VCR-wiki-zh-easy-test
+task: "vcr_wiki_zh_easy"
+test_split: train[:100]
+process_results: !function utils.vcr_zh_process_results
+metric_list:
+  - metric: vcr_percetion_score
+    aggregation: !function utils.vcr_zh_process_results
+    higher_is_better: true
+  - metric: vcr_cognition_score
+    aggregation: !function utils.vcr_zh_process_results
+    higher_is_better: true
+model_specific_prompt_kwargs:
+  default:
+    pre_prompt: ""
+    post_prompt: "图像中被覆盖的文本是什么？请在不输出解释的情况下还原被覆盖的文本。"
\ No newline at end of file
diff --git a/lmms_eval/tasks/vcr_wiki/vcr_wiki_zh_easy_500.yaml b/lmms_eval/tasks/vcr_wiki/vcr_wiki_zh_easy_500.yaml
new file mode 100644
index 00000000..da210bb8
--- /dev/null
+++ b/lmms_eval/tasks/vcr_wiki/vcr_wiki_zh_easy_500.yaml
@@ -0,0 +1,16 @@
+"include": "_default_template_vcr_yaml"
+dataset_path: vcr-org/VCR-wiki-zh-easy-test
+task: "vcr_wiki_zh_easy"
+test_split: train[:500]
+process_results: !function utils.vcr_zh_process_results
+metric_list:
+  - metric: vcr_percetion_score
+    aggregation: !function utils.vcr_zh_process_results
+    higher_is_better: true
+  - metric: vcr_cognition_score
+    aggregation: !function utils.vcr_zh_process_results
+    higher_is_better: true
+model_specific_prompt_kwargs:
+  default:
+    pre_prompt: ""
+    post_prompt: "图像中被覆盖的文本是什么？请在不输出解释的情况下还原被覆盖的文本。"
\ No newline at end of file
diff --git a/lmms_eval/tasks/vcr_wiki/vcr_wiki_zh_easy_5000.yaml b/lmms_eval/tasks/vcr_wiki/vcr_wiki_zh_easy_5000.yaml
new file mode 100644
index 00000000..8807db4a
--- /dev/null
+++ b/lmms_eval/tasks/vcr_wiki/vcr_wiki_zh_easy_5000.yaml
@@ -0,0 +1,16 @@
+"include": "_default_template_vcr_yaml"
+dataset_path: vcr-org/VCR-wiki-zh-easy-test
+task: "vcr_wiki_zh_easy"
+test_split: train
+process_results: !function utils.vcr_zh_process_results
+metric_list:
+  - metric: vcr_percetion_score
+    aggregation: !function utils.vcr_zh_process_results
+    higher_is_better: true
+  - metric: vcr_cognition_score
+    aggregation: !function utils.vcr_zh_process_results
+    higher_is_better: true
+model_specific_prompt_kwargs:
+  default:
+    pre_prompt: ""
+    post_prompt: "图像中被覆盖的文本是什么？请在不输出解释的情况下还原被覆盖的文本。"
\ No newline at end of file
diff --git a/lmms_eval/tasks/vcr_wiki/vcr_wiki_zh_hard_100.yaml b/lmms_eval/tasks/vcr_wiki/vcr_wiki_zh_hard_100.yaml
new file mode 100644
index 00000000..c15c8059
--- /dev/null
+++ b/lmms_eval/tasks/vcr_wiki/vcr_wiki_zh_hard_100.yaml
@@ -0,0 +1,16 @@
+"include": "_default_template_vcr_yaml"
+dataset_path: vcr-org/VCR-wiki-zh-hard-test
+task: "vcr_wiki_zh_hard"
+test_split: train[:100]
+process_results: !function utils.vcr_zh_process_results
+metric_list:
+  - metric: vcr_percetion_score
+    aggregation: !function utils.vcr_zh_process_results
+    higher_is_better: true
+  - metric: vcr_cognition_score
+    aggregation: !function utils.vcr_zh_process_results
+    higher_is_better: true
+model_specific_prompt_kwargs:
+  default:
+    pre_prompt: ""
+    post_prompt: "图像中被覆盖的文本是什么？请在不输出解释的情况下还原被覆盖的文本。"
\ No newline at end of file
diff --git a/lmms_eval/tasks/vcr_wiki/vcr_wiki_zh_hard_500.yaml b/lmms_eval/tasks/vcr_wiki/vcr_wiki_zh_hard_500.yaml
new file mode 100644
index 00000000..90bba34e
--- /dev/null
+++ b/lmms_eval/tasks/vcr_wiki/vcr_wiki_zh_hard_500.yaml
@@ -0,0 +1,16 @@
+"include": "_default_template_vcr_yaml"
+dataset_path: vcr-org/VCR-wiki-zh-hard-test
+task: "vcr_wiki_zh_hard"
+test_split: train[:500]
+process_results: !function utils.vcr_zh_process_results
+metric_list:
+  - metric: vcr_percetion_score
+    aggregation: !function utils.vcr_zh_process_results
+    higher_is_better: true
+  - metric: vcr_cognition_score
+    aggregation: !function utils.vcr_zh_process_results
+    higher_is_better: true
+model_specific_prompt_kwargs:
+  default:
+    pre_prompt: ""
+    post_prompt: "图像中被覆盖的文本是什么？请在不输出解释的情况下还原被覆盖的文本。"
\ No newline at end of file
diff --git a/lmms_eval/tasks/vcr_wiki/vcr_wiki_zh_hard_5000.yaml b/lmms_eval/tasks/vcr_wiki/vcr_wiki_zh_hard_5000.yaml
new file mode 100644
index 00000000..48893f73
--- /dev/null
+++ b/lmms_eval/tasks/vcr_wiki/vcr_wiki_zh_hard_5000.yaml
@@ -0,0 +1,16 @@
+"include": "_default_template_vcr_yaml"
+dataset_path: vcr-org/VCR-wiki-zh-hard-test
+task: "vcr_wiki_zh_hard"
+test_split: train
+process_results: !function utils.vcr_zh_process_results
+metric_list:
+  - metric: vcr_percetion_score
+    aggregation: !function utils.vcr_zh_process_results
+    higher_is_better: true
+  - metric: vcr_cognition_score
+    aggregation: !function utils.vcr_zh_process_results
+    higher_is_better: true
+model_specific_prompt_kwargs:
+  default:
+    pre_prompt: ""
+    post_prompt: "图像中被覆盖的文本是什么？请在不输出解释的情况下还原被覆盖的文本。"
\ No newline at end of file