feat: update new table

app.py

@@ -9,13 +9,13 @@ from assets.text import INTRODUCTION_TEXT, METRICS_TEXT, EVALUTION_TEXT, ACKNOWL
 ORIGINAL_DF = pd.read_csv("./data/chinese_benchmark_gen.csv", encoding='utf-8') # space separated values
 ORIGINAL_DF_PER = pd.read_csv("./data/chinese_benchmark_per.csv", encoding='utf-8') #
 
-
 ORIGINAL_DF_SUB_GEN = pd.read_csv("./data/subclass_gen.csv", encoding='utf-8') #
 ORIGINAL_DF_SUB_PER = pd.read_csv("./data/subclass_per.csv", encoding='utf-8')
 
+ORIGINAL_DF_NEW = pd.read_csv("./data/ChineseGuardBench.csv", encoding='utf-8') # new table
 
 METRICS = ["Accuracy", "Precision_Unsafe", "Recall_Unsafe", "Precision_Safe", "Recall_Safe", "None"]
-
+# METRICS_GuardBench = ["Accuracy", "Precision_Unsafe", "Recall_Unsafe", "Precision_Safe", "Recall_Safe", "None"]
 
 SUBCLASS = ["Discrimination", "Variant", "Psychology", "Politics", "Eroticism", "Vulgarity", "Property", "Injury", "Criminality", "Ethics"]
 
@@ -70,6 +70,19 @@ def format_number(x):
     return float(f"{x:.3}")
 
 
+def get_dataset_new_csv(
+    model_size: List[str],
+):
+    df = ORIGINAL_DF_NEW[ORIGINAL_DF_NEW['Size'].isin(model_size)]
+    df = df.drop(columns="Size")
+
+    leaderboard_table = gr.components.Dataframe(
+        value=df,
+        interactive=False,
+        visible=True,
+    )
+    return leaderboard_table
+
 def get_dataset_csv(
     model_size: List[str],
 ):
@@ -146,6 +159,17 @@ def get_dataset_classfier_gen(
         leaderboard_table = get_dataset_csv_sub_gen(model_size, subclass_choice)
     return leaderboard_table
 
+def get_ChineseGuardBench(
+    model_size: List[str],
+    main_choice: List[str],
+):
+    leaderboard_table = get_dataset_new_csv(model_size)
+    # elif main_choice != "Subclass":
+    #     subclass_choice = main_choice
+    #     leaderboard_table = get_dataset_csv_sub_gen(model_size, subclass_choice)
+    return leaderboard_table
+
+
 def get_dataset_classfier_per(
     model_size: List[str],
     main_choice: List[str],
@@ -200,6 +224,11 @@ with gr.Blocks() as demo:
             dataframe_all_per = gr.components.Dataframe(
                 elem_id="leaderboard-table",
             )
+            
+        with gr.TabItem("🏅 NEW", elem_id="od-benchmark-tab-table", id=7):
+            dataframe_all_guardbench = gr.components.Dataframe(
+                elem_id="leaderboard-table",
+            )
 
     # ----------------- modify text -----------------
     with gr.Row():
@@ -261,6 +290,24 @@ with gr.Blocks() as demo:
         outputs=dataframe_all_gen,
     )
 
+    # this is new results for ChineseGuardBench
+    # main_choice.change(
+    #     get_ChineseGuardBench,
+    #     inputs=[model_choice, main_choice],
+    #     outputs=dataframe_all_guardbench,
+    # )
+
+    model_choice.change(
+        get_ChineseGuardBench,
+        inputs=[model_choice, main_choice],
+        outputs=dataframe_all_guardbench,
+    )
+
+    demo.load(
+        fn=get_ChineseGuardBench,
+        inputs=[model_choice, main_choice],
+        outputs=dataframe_all_guardbench,
+    )
 
 demo.launch(share=True)
 

assets/text.py

@@ -35,7 +35,8 @@ EVALUTION_TEXT= """
 We evaluate the models using two methods: perplexity(multiple choice) and generation. 
 For perplexity, we select the label which is the lowest perplexity as the predicted results. 
 For generation, we use the content generated by the model to make prediction.
-The following are the results of the evaluation. 👇👇👇    
+The following are the results of the evaluation. 
+In the "New" table, we present evaluation results for generation on a meticulously curated new benchmark, with details of its processing to be introduced later.👇👇👇    
 </span> <br><br>
 
 

data/ChineseGuardBench.csv

@@ -0,0 +1,33 @@
+Model,Size,F1,Accuracy,Precision,Recall,FPR,FNR
+Deepexi-Guard-3B,1B~5B,89.63 ,89.72 ,85.53 ,94.15 ,14.24 ,5.85 
+Qwen3-32B,~30B,88.54 ,89.25 ,89.08 ,88.02 ,9.64 ,11.98 
+Qwen3-235B-A22B,>65B,87.92 ,88.96 ,90.86 ,85.17 ,7.66 ,14.83 
+Qwen3-235B-A22B-Instruct-2507,>65B,87.81 ,89.13 ,93.27 ,82.96 ,5.35 ,17.04 
+GLM-Z1-9B-0414,5B~10B,87.36 ,88.03 ,87.11 ,87.61 ,11.59 ,12.39 
+Qwen2.5-72B-Instruct,>65B,86.81 ,88.27 ,92.50 ,81.79 ,5.93 ,18.21 
+QwQ-32B,~30B,86.80 ,88.35 ,93.33 ,81.12 ,5.18 ,18.88 
+Phi-4,10B~20B,85.95 ,86.88 ,86.90 ,85.02 ,11.45 ,14.98 
+Gemma-3-27B-it,~30B,85.29 ,86.78 ,89.83 ,81.19 ,8.22 ,18.81 
+DeepSeek-R1-0528,>65B,85.24 ,87.47 ,96.02 ,76.63 ,2.84 ,23.37 
+Mistral-Small-3.2-24B-Instruct,~30B,85.07 ,87.03 ,93.14 ,78.29 ,5.15 ,21.71 
+GLM-4-9B-chat,5B~10B,84.85 ,86.27 ,88.47 ,81.52 ,9.49 ,18.48 
+MD-Judge-v0_2-internlm2_7B,5B~10B,84.63 ,85.88 ,87.03 ,82.37 ,10.98 ,17.63 
+DeepSeek-R1-Distill-Qwen-32B,~30B,84.55 ,86.64 ,93.05 ,77.47 ,5.17 ,22.53 
+Hunyuan-A13B-Instruct,>65B,84.32 ,86.21 ,90.97 ,78.58 ,6.98 ,21.42 
+Moonlight-16B-A3B-Instruct,10B~20B,84.21 ,84.35 ,80.41 ,88.38 ,19.25 ,11.62 
+GLM-Z1-32B-0414,~30B,83.40 ,85.75 ,92.63 ,75.85 ,5.40 ,24.15 
+Qwen3-8B,5B~10B,83.05 ,85.51 ,92.69 ,75.23 ,5.30 ,24.77 
+Qwen2.5-7B-Instruct,5B~10B,82.96 ,84.99 ,89.41 ,77.37 ,8.20 ,22.63 
+Qwen2.5-1.5B-Instruct,1B~5B,79.48 ,77.08 ,68.83 ,94.03 ,38.07 ,5.97 
+shieldgemma-2B,1B~5B,79.19 ,79.63 ,76.50 ,82.06 ,22.54 ,17.94 
+Qwen2.5-3B-Instruct,1B~5B,79.05 ,77.57 ,70.69 ,89.66 ,33.25 ,10.34 
+SHTEC_safety_fence_model_7B,5B~10B,78.44 ,82.48 ,93.54 ,67.54 ,4.17 ,32.46 
+Qwen3-4B,1B~5B,78.16 ,82.50 ,95.12 ,66.33 ,3.04 ,33.67 
+SmolLM3-3B,1B~5B,76.10 ,79.19 ,83.09 ,70.19 ,12.77 ,29.81 
+ERNIE-4.5-21B-A3B-Paddle,~20B,75.21 ,80.58 ,94.58 ,62.42 ,3.20 ,37.58 
+Qwen3-1.7B,1B~5B,74.46 ,79.34 ,89.36 ,63.82 ,6.79 ,36.18 
+internlm2_5-7B-chat,5B~10B,71.52 ,78.49 ,95.34 ,57.22 ,2.50 ,42.78 
+Llama-Guard-4-12B,10B~20B,65.66 ,74.64 ,90.99 ,51.36 ,4.54 ,48.64 
+Llama-Guard-3-8B,5B~10B,59.33 ,72.44 ,97.80 ,42.58 ,0.86 ,57.42 
+DeepSeek-R1-Distill-Qwen-7B,5B~10B,45.27 ,65.53 ,90.36 ,30.20 ,2.88 ,69.80 
+Gemma-3n-E4B-it,5B~10B,44.05 ,64.88 ,88.80 ,29.29 ,3.30 ,70.71