diff --git a/ais_bench/benchmark/configs/datasets/json_mode_eval/json_mode_eval_gen.py b/ais_bench/benchmark/configs/datasets/json_mode_eval/json_mode_eval_gen.py
new file mode 100644
index 0000000000000000000000000000000000000000..a0483bb95c6a58e9132d76dc744c1a1b0b179ea5
--- /dev/null
+++ b/ais_bench/benchmark/configs/datasets/json_mode_eval/json_mode_eval_gen.py
@@ -0,0 +1,37 @@
+from ais_bench.benchmark.openicl.icl_prompt_template import PromptTemplate
+from ais_bench.benchmark.openicl.icl_retriever import ZeroRetriever
+from ais_bench.benchmark.openicl.icl_inferencer import GenInferencer
+from ais_bench.benchmark.datasets import JsonModeEvalDataset, JsonModeEvalEvaluator
+
+json_mode_eval_reader_cfg = dict(
+    input_columns=['prompt'], 
+    output_column='completion', 
+    train_split='train', 
+    test_split='test'
+)
+
+JSON_MODE_EVAL_TEMPLATE = ''
+
+json_mode_eval_infer_cfg = dict(
+    prompt_template=dict(
+        type=PromptTemplate, 
+        template=JSON_MODE_EVAL_TEMPLATE
+    ),
+    retriever=dict(type=ZeroRetriever),
+    inferencer=dict(type=GenInferencer), 
+)
+
+json_mode_eval_eval_cfg = dict(
+    evaluator=dict(type=JsonModeEvalEvaluator), 
+)
+
+json_mode_eval_datasets = [
+    dict(
+        abbr='json_mode_eval',
+        type=JsonModeEvalDataset,
+        path='/data/dataset/json-mode-eval',  # 数据集路径
+        reader_cfg=json_mode_eval_reader_cfg,
+        infer_cfg=json_mode_eval_infer_cfg,
+        eval_cfg=json_mode_eval_eval_cfg,
+    )
+]
diff --git a/ais_bench/benchmark/datasets/json_mode_eval.py b/ais_bench/benchmark/datasets/json_mode_eval.py
new file mode 100644
index 0000000000000000000000000000000000000000..a0608990f2bf074d5d38fbd5d3632bc5755178b0
--- /dev/null
+++ b/ais_bench/benchmark/datasets/json_mode_eval.py
@@ -0,0 +1,64 @@
+import json
+import os
+from os import environ
+import pandas as pd
+from datasets import Dataset, DatasetDict
+from ais_bench.benchmark.openicl import BaseEvaluator
+from ais_bench.benchmark.registry import LOAD_DATASET
+
+@LOAD_DATASET.register_module()
+class JsonModeEvalDataset:
+    @staticmethod
+    def load(path):
+        path = os.path.abspath(path)
+        
+        datasets = {}
+        for split in ['train', 'test']:
+            file_path = os.path.join(path, f'{split}.parquet')
+            
+            if not os.path.exists(file_path):
+                continue
+                
+            df = pd.read_parquet(file_path)
+            
+            required_columns = {'prompt', 'completion', 'schema'}
+            if not required_columns.issubset(df.columns):
+                missing = required_columns - set(df.columns)
+                raise ValueError(f"Parquet Missing: {missing}")
+            
+            dataset = Dataset.from_pandas(df)
+            datasets[split] = dataset
+        
+        return DatasetDict(datasets)
+
+class JsonModeEvalEvaluator(BaseEvaluator):
+    def is_json(self, s):
+        try:
+            json.loads(s)
+            return True
+        except (json.JSONDecodeError, TypeError):
+            return False
+
+    def score(self, predictions, references):  
+        correct = 0
+        count = len(predictions)
+        details = []
+        
+        for pred, ref in zip(predictions, references):
+            pred_valid = self.is_json(pred)
+            ref_valid = self.is_json(ref)
+            both_valid = pred_valid and ref_valid
+            
+            if both_valid:
+                correct += 1
+                
+            details.append({
+                'pred': pred,
+                'answer': ref,
+                'pred_valid': pred_valid,
+                'ref_valid': ref_valid,
+                'correct': both_valid
+            })
+        
+        accuracy = 100 * correct / count if count > 0 else 0
+        return {'accuracy': accuracy, 'details': details}