add Grok-2-vision & a huge refactoring to clean up code

constants.py

@@ -116,6 +116,7 @@ MODEL_NAME_MAP = {
     "InternVL2_5_78B": "InternVL2.5-78B",
     "InternVL2_5_2B": "InternVL2.5-2B",
     "InternVL2_5_8B": "InternVL2.5-8B",
+    "Grok-2-vision-1212": "Grok-2-vision-1212",
 }
 
 DIMENSION_NAME_MAP = {
@@ -203,14 +204,15 @@ MODEL_URLS = {
     "InternVL2_5_78B": "https://huggingface.co/OpenGVLab/InternVL2_5-78B",
     "InternVL2_5_2B": "https://huggingface.co/OpenGVLab/InternVL2_5-2B",
     "InternVL2_5_8B": "https://huggingface.co/OpenGVLab/InternVL2_5-8B",
+    "Grok-2-vision-1212": "https://x.ai/blog/grok-1212",
 }
 
 # Define the base MODEL_GROUPS structure
 BASE_MODEL_GROUPS = {
     "All": list(MODEL_NAME_MAP.keys()),
-    "Flagship Models": ['Claude_3.5_new', 'GPT_4o', 'Claude_3.5', 'Gemini_1.5_pro_002', 'Qwen2_VL_72B', 'InternVL2_76B', 'llava_onevision_72B', 'NVLM', 'Molmo_72B', 'InternVL2_5_78B'],
+    "Flagship Models": ['Claude_3.5_new', 'GPT_4o', 'Claude_3.5', 'Gemini_1.5_pro_002', 'Qwen2_VL_72B', 'InternVL2_76B', 'llava_onevision_72B', 'NVLM', 'Molmo_72B', 'InternVL2_5_78B', 'Grok-2-vision-1212'],
     "Efficiency Models": ['Gemini_1.5_flash_002', 'GPT_4o_mini', 'Qwen2_VL_7B', 'Pixtral_12B', 'Aria', 'InternVL2_8B', 'Phi-3.5-vision', 'MiniCPM_v2.6', 'llava_onevision_7B', 'Llama_3_2_11B', 'Idefics3', 'Molmo_7B_D', "Aquila_VL_2B", "POINTS_7B", "Qwen2_VL_2B", "InternVL2_2B", "InternVL2_5_2B", "InternVL2_5_8B"],
-    "Proprietary Flagship models": ['Claude_3.5_new', 'GPT_4o', 'Claude_3.5', 'Gemini_1.5_pro_002'],
+    "Proprietary Flagship models": ['Claude_3.5_new', 'GPT_4o', 'Claude_3.5', 'Gemini_1.5_pro_002', 'Grok-2-vision-1212'],
     "Proprietary Efficiency Models": ['Gemini_1.5_flash_002', 'GPT_4o_mini'],
     "Open-source Flagship Models": ['Qwen2_VL_72B', 'InternVL2_76B', 'llava_onevision_72B', 'NVLM', "Molmo_72B", "InternVL2_5_78B"],
     "Open-source Efficiency Models": ['Qwen2_VL_7B', 'Pixtral_12B', 'Aria', 'InternVL2_8B', 'Phi-3.5-vision', 'MiniCPM_v2.6', 'llava_onevision_7B', 'Llama_3_2_11B', 'Idefics3', 'Molmo_7B_D', "Aquila_VL_2B", "POINTS_7B", "Qwen2_VL_2B", "InternVL2_2B", "InternVL2_5_2B", "InternVL2_5_8B"]

static/eval_results/SI/Grok-2-vision-1212/summary_and_keyword_stats.json

@@ -0,0 +1,213 @@
+{
+    "model_summary": {
+        "core": {
+            "num_eval_tasks": 273,
+            "num_eval_samples": 4108,
+            "macro_mean_score": 0.4120738315897316
+        },
+        "open": {
+            "num_eval_tasks": 42,
+            "num_eval_samples": 808,
+            "macro_mean_score": 0.5369427320775519
+        },
+        "overall_score": 0.42872301832144094
+    },
+    "keyword_stats": {
+        "skills": {
+            "Object Recognition and Classification": {
+                "count": 172,
+                "num_samples": 2704,
+                "tasks": [],
+                "average_score": 0.466222487838683
+            },
+            "Language Understanding and Generation": {
+                "count": 102,
+                "num_samples": 1707,
+                "tasks": [],
+                "average_score": 0.49260409084481493
+            },
+            "Commonsense and Social Reasoning": {
+                "count": 38,
+                "num_samples": 652,
+                "tasks": [],
+                "average_score": 0.5513856107049714
+            },
+            "Scene and Event Understanding": {
+                "count": 60,
+                "num_samples": 1004,
+                "tasks": [],
+                "average_score": 0.5869208042949662
+            },
+            "Domain-Specific Knowledge and Skills": {
+                "count": 46,
+                "num_samples": 896,
+                "tasks": [],
+                "average_score": 0.4815724520339999
+            },
+            "Ethical and Safety Reasoning": {
+                "count": 10,
+                "num_samples": 170,
+                "tasks": [],
+                "average_score": 0.6636804511278196
+            },
+            "Text Recognition (OCR)": {
+                "count": 101,
+                "num_samples": 1680,
+                "tasks": [],
+                "average_score": 0.3702735127125422
+            },
+            "Spatial and Temporal Reasoning": {
+                "count": 78,
+                "num_samples": 1270,
+                "tasks": [],
+                "average_score": 0.33515724252578744
+            },
+            "Mathematical and Logical Reasoning": {
+                "count": 91,
+                "num_samples": 1628,
+                "tasks": [],
+                "average_score": 0.38305225927176256
+            },
+            "Planning and Decision Making": {
+                "count": 23,
+                "num_samples": 355,
+                "tasks": [],
+                "average_score": 0.10535044936296332
+            }
+        },
+        "input_format": {
+            "Photographs": {
+                "count": 83,
+                "num_samples": 1310,
+                "tasks": [],
+                "average_score": 0.5618064274876843
+            },
+            "Artistic and Creative Content": {
+                "count": 22,
+                "num_samples": 388,
+                "tasks": [],
+                "average_score": 0.5826343022222362
+            },
+            "Diagrams and Data Visualizations": {
+                "count": 88,
+                "num_samples": 1523,
+                "tasks": [],
+                "average_score": 0.46202217635966664
+            },
+            "Text-Based Images and Documents": {
+                "count": 53,
+                "num_samples": 847,
+                "tasks": [],
+                "average_score": 0.30345879283667027
+            },
+            "User Interface Screenshots": {
+                "count": 67,
+                "num_samples": 1117,
+                "tasks": [],
+                "average_score": 0.2751054353728693
+            },
+            "3D Models and Aerial Imagery": {
+                "count": 2,
+                "num_samples": 30,
+                "tasks": [],
+                "average_score": 0.21326546545524588
+            }
+        },
+        "output_format": {
+            "contextual_formatted_text": {
+                "count": 63,
+                "num_samples": 972,
+                "tasks": [],
+                "average_score": 0.36955979847719544
+            },
+            "open_ended_output": {
+                "count": 51,
+                "num_samples": 986,
+                "tasks": [],
+                "average_score": 0.531045897627514
+            },
+            "structured_output": {
+                "count": 72,
+                "num_samples": 1120,
+                "tasks": [],
+                "average_score": 0.39618293480240524
+            },
+            "numerical_data": {
+                "count": 39,
+                "num_samples": 694,
+                "tasks": [],
+                "average_score": 0.4022896767902012
+            },
+            "multiple_choice": {
+                "count": 33,
+                "num_samples": 567,
+                "tasks": [],
+                "average_score": 0.5637194455376273
+            },
+            "exact_text": {
+                "count": 57,
+                "num_samples": 876,
+                "tasks": [],
+                "average_score": 0.3835953032430645
+            }
+        },
+        "input_num": {
+            "1-image": {
+                "count": 315,
+                "num_samples": 5215,
+                "tasks": [],
+                "average_score": 0.4287230183214409
+            }
+        },
+        "app": {
+            "Knowledge": {
+                "count": 77,
+                "num_samples": 1291,
+                "tasks": [],
+                "average_score": 0.5298084871907297
+            },
+            "Perception": {
+                "count": 82,
+                "num_samples": 1318,
+                "tasks": [],
+                "average_score": 0.5357263973810524
+            },
+            "Coding": {
+                "count": 16,
+                "num_samples": 244,
+                "tasks": [],
+                "average_score": 0.4783708274976657
+            },
+            "Science": {
+                "count": 22,
+                "num_samples": 469,
+                "tasks": [],
+                "average_score": 0.4448688427088975
+            },
+            "Information_Extraction": {
+                "count": 41,
+                "num_samples": 639,
+                "tasks": [],
+                "average_score": 0.312597090907984
+            },
+            "Planning": {
+                "count": 44,
+                "num_samples": 712,
+                "tasks": [],
+                "average_score": 0.18803058075452733
+            },
+            "Mathematics": {
+                "count": 30,
+                "num_samples": 497,
+                "tasks": [],
+                "average_score": 0.35624322358581967
+            },
+            "Metrics": {
+                "count": 3,
+                "num_samples": 45,
+                "tasks": [],
+                "average_score": 0.3682539682539683
+            }
+        }
+    }
+}

static/eval_results/SI/Grok-2-vision-1212/task_results.json

@@ -0,0 +1,2207 @@
+[
+    {
+        "name": "ascii_art_30",
+        "score": 0.14285714285714285,
+        "eval_type": "llm",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "humor_explanation",
+        "score": 0.8600000000000001,
+        "eval_type": "llm",
+        "num_demo": 1,
+        "num_query": 15
+    },
+    {
+        "name": "science_figure_explanation",
+        "score": 0.8448275862068965,
+        "eval_type": "llm",
+        "num_demo": 1,
+        "num_query": 29
+    },
+    {
+        "name": "vibe_eval_phrase",
+        "score": 0.6357142857142858,
+        "eval_type": "llm",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "traffic_accident_analysis",
+        "score": 0.39999999999999997,
+        "eval_type": "llm",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "figurative_speech_explanation",
+        "score": 0.8517241379310343,
+        "eval_type": "llm",
+        "num_demo": 1,
+        "num_query": 29
+    },
+    {
+        "name": "table2latex_complex",
+        "score": 0.611111111111111,
+        "eval_type": "llm",
+        "num_demo": 1,
+        "num_query": 9
+    },
+    {
+        "name": "unusual_images",
+        "score": 0.8275862068965517,
+        "eval_type": "llm",
+        "num_demo": 1,
+        "num_query": 29
+    },
+    {
+        "name": "art_explanation",
+        "score": 0.7689655172413796,
+        "eval_type": "llm",
+        "num_demo": 1,
+        "num_query": 29
+    },
+    {
+        "name": "ocr_open_ended_qa",
+        "score": 0.8172413793103449,
+        "eval_type": "llm",
+        "num_demo": 1,
+        "num_query": 29
+    },
+    {
+        "name": "bar_chart_interpretation",
+        "score": 0.5310344827586206,
+        "eval_type": "llm",
+        "num_demo": 1,
+        "num_query": 29
+    },
+    {
+        "name": "scibench_w_solution_open_ended",
+        "score": 0.39,
+        "eval_type": "llm",
+        "num_demo": 1,
+        "num_query": 25
+    },
+    {
+        "name": "GUI_Chat_Hard",
+        "score": 0.4538461538461539,
+        "eval_type": "llm",
+        "num_demo": 1,
+        "num_query": 26
+    },
+    {
+        "name": "image_humor_understanding",
+        "score": 0.886206896551724,
+        "eval_type": "llm",
+        "num_demo": 1,
+        "num_query": 29
+    },
+    {
+        "name": "defeasible_reasoning",
+        "score": 0.8655172413793104,
+        "eval_type": "llm",
+        "num_demo": 1,
+        "num_query": 29
+    },
+    {
+        "name": "funny_image_title",
+        "score": 0.6357142857142858,
+        "eval_type": "llm",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "tweets_captioning",
+        "score": 0.6214285714285716,
+        "eval_type": "llm",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "graph_interpretation",
+        "score": 0.7896551724137929,
+        "eval_type": "llm",
+        "num_demo": 1,
+        "num_query": 29
+    },
+    {
+        "name": "meme_explain",
+        "score": 0.892857142857143,
+        "eval_type": "llm",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "guess_image_generation_prompt",
+        "score": 0.7368421052631581,
+        "eval_type": "llm",
+        "num_demo": 1,
+        "num_query": 19
+    },
+    {
+        "name": "visualization_with_code",
+        "score": 0.5142857142857142,
+        "eval_type": "llm",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "iq_test_open_ended",
+        "score": 0.42758620689655175,
+        "eval_type": "llm",
+        "num_demo": 1,
+        "num_query": 29
+    },
+    {
+        "name": "electrocardiogram",
+        "score": 0.2785714285714285,
+        "eval_type": "llm",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "image_captioning_with_additional_requirements",
+        "score": 0.8428571428571432,
+        "eval_type": "llm",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "docci_image_description_long",
+        "score": 0.7142857142857143,
+        "eval_type": "llm",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "GUI_Chat_Easy",
+        "score": 0.5923076923076924,
+        "eval_type": "llm",
+        "num_demo": 1,
+        "num_query": 26
+    },
+    {
+        "name": "bridge_strategies_advanced",
+        "score": 0.20714285714285716,
+        "eval_type": "llm",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "bridge_strategies_worldclass",
+        "score": 0.1928571428571429,
+        "eval_type": "llm",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "bridge_strategies_expert",
+        "score": 0.3642857142857142,
+        "eval_type": "llm",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "multi_lingual_Ruozhiba_expalnation_Spanish",
+        "score": 0.12857142857142856,
+        "eval_type": "llm",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "multi_lingual_Ruozhiba_expalnation_Japanese",
+        "score": 0.09285714285714286,
+        "eval_type": "llm",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "multi_lingual_Ruozhiba_expalnation_French",
+        "score": 0.09999999999999999,
+        "eval_type": "llm",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "multi_lingual_Ruozhiba_expalnation_Arabic",
+        "score": 0.13571428571428573,
+        "eval_type": "llm",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "multi_lingual_Ruozhiba_expalnation_Russian",
+        "score": 0.07142857142857142,
+        "eval_type": "llm",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "multi_lingual_Ruozhiba_expalnation_English",
+        "score": 0.07142857142857142,
+        "eval_type": "llm",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "table_understanding_fetaqa",
+        "score": 0.4142857142857143,
+        "eval_type": "llm",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "red_teaming_celebrity",
+        "score": 0.8250000000000002,
+        "eval_type": "llm",
+        "num_demo": 0,
+        "num_query": 20
+    },
+    {
+        "name": "red_teaming_captcha",
+        "score": 0.12105263157894738,
+        "eval_type": "llm",
+        "num_demo": 1,
+        "num_query": 19
+    },
+    {
+        "name": "red_teaming_jailbreak",
+        "score": 0.575,
+        "eval_type": "llm",
+        "num_demo": 0,
+        "num_query": 20
+    },
+    {
+        "name": "red_teaming_visualmisleading",
+        "score": 0.8789473684210528,
+        "eval_type": "llm",
+        "num_demo": 1,
+        "num_query": 19
+    },
+    {
+        "name": "red_teaming_racial",
+        "score": 0.7250000000000001,
+        "eval_type": "llm",
+        "num_demo": 0,
+        "num_query": 20
+    },
+    {
+        "name": "red_teaming_politics",
+        "score": 0.7150000000000001,
+        "eval_type": "llm",
+        "num_demo": 0,
+        "num_query": 20
+    },
+    {
+        "name": "brand_logo_recognition_and_elaboration",
+        "score": 0.78,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 25
+    },
+    {
+        "name": "exchange_rate_estimate_plot",
+        "score": 0.9693285714285712,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "math_parity",
+        "score": 0.8,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 15
+    },
+    {
+        "name": "traffic_future_prediction_from_line_plot",
+        "score": 0.6339473684210527,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 19
+    },
+    {
+        "name": "graph_chordless_cycle",
+        "score": 0.21428571428571427,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "youtube_video_info_parsing",
+        "score": 0.1547619047619048,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "super_clevr_scene_understanding",
+        "score": 0.5,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "figureqa",
+        "score": 0.6428571428571429,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "face_keypoint_detection",
+        "score": 0.5273791704902433,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "widerface_face_count_and_event_classification",
+        "score": 0.6071428571428571,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "average_humidity_estimate_plot",
+        "score": 0.764,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 15
+    },
+    {
+        "name": "weather_info_parsing",
+        "score": 0.5277777777777779,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "egocentric_analysis_single_image",
+        "score": 0.4444444444444444,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 9
+    },
+    {
+        "name": "waybill_number_sequence_extraction",
+        "score": 0.2857142857142857,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "graph_maxflow",
+        "score": 0.2,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 15
+    },
+    {
+        "name": "TV_show_info_parsing",
+        "score": 0.3412698412698412,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "insect_order_classification",
+        "score": 0.3333333333333333,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 15
+    },
+    {
+        "name": "electricity_plot_future_prediction",
+        "score": 0.806521052631579,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 19
+    },
+    {
+        "name": "chemistry_exams_v",
+        "score": 0.14285714285714285,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "finance_table_understanding",
+        "score": 0.42857142857142855,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "funsd_document_qa",
+        "score": 0.21428571428571427,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "vibe_eval_open",
+        "score": 0.14285714285714285,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "question_solution_solving",
+        "score": 0.07142857142857142,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "graph_theory",
+        "score": 0.35714285714285715,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "geometry_analytic",
+        "score": 0.21428571428571427,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "geometry_length",
+        "score": 0.35714285714285715,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "algebra",
+        "score": 0.2857142857142857,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "chess_puzzle_single_step",
+        "score": 0.0,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 15
+    },
+    {
+        "name": "chess_winner_identification",
+        "score": 0.3333333333333333,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 15
+    },
+    {
+        "name": "physical_property_reasoning",
+        "score": 0.9285714285714286,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "humor_understand_caption_match",
+        "score": 0.6666666666666666,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 15
+    },
+    {
+        "name": "coco_object_detection_by_query_property",
+        "score": 0.6660750158426613,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "multilingual_game_info_parsing",
+        "score": 0.26785714285714285,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "mnist_pattern",
+        "score": 0.0,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "dvqa",
+        "score": 0.9473684210526315,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 19
+    },
+    {
+        "name": "physics_exams_v",
+        "score": 0.5,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "snli_ve_visual_entailment",
+        "score": 0.7333333333333333,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 15
+    },
+    {
+        "name": "3d_indoor_scene_text_bbox_selection",
+        "score": 0.2857142857142857,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "geometry_descriptive",
+        "score": 0.14285714285714285,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "top_rated_hotel_identification",
+        "score": 0.0,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "science_molecule_chemistry",
+        "score": 0.8,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 15
+    },
+    {
+        "name": "game_info_parsing",
+        "score": 0.5064935064935063,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "deciphering_oracle_bone",
+        "score": 0.0,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "signboard_identification",
+        "score": 0.5,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "image_style_recognition",
+        "score": 1.0,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "math_convexity_value_estimation",
+        "score": 0.5147224146995476,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 15
+    },
+    {
+        "name": "3d_indoor_scene_text_bbox_prediction",
+        "score": 0.14081664519620604,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "movie_info_parsing",
+        "score": 0.3125,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "human_relationship_reasoning",
+        "score": 1.0,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "graph_shortest_path_kamada_kawai",
+        "score": 0.07142857142857142,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "coco_person_detection",
+        "score": 0.6271347226619672,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "chart_vqa",
+        "score": 0.5714285714285714,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "nlvr2_two_image_compare_qa",
+        "score": 0.7857142857142857,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "math_exams_v",
+        "score": 0.6428571428571429,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "newspaper_ocr_in_query_box",
+        "score": 0.13333333333333333,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 15
+    },
+    {
+        "name": "mvsa_sentiment_classification",
+        "score": 0.6428571428571429,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "egocentric_spatial_reasoning",
+        "score": 0.4444444444444444,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 9
+    },
+    {
+        "name": "graph_isomorphism",
+        "score": 0.6,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 15
+    },
+    {
+        "name": "code_programming_test_easy",
+        "score": 0.2708333333333333,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 24
+    },
+    {
+        "name": "biology_exams_v",
+        "score": 0.2857142857142857,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "long_string_number_recognition",
+        "score": 0.07142857142857142,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "kvqa_knowledge_aware_qa",
+        "score": 0.47368421052631576,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 19
+    },
+    {
+        "name": "math_breakpoint",
+        "score": 0.5333333333333333,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 15
+    },
+    {
+        "name": "landmark_recognition_and_qa",
+        "score": 0.5555555555555555,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 15
+    },
+    {
+        "name": "map_diagram_qa",
+        "score": 0.5714285714285714,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "pmc_vqa_medical_image_qa",
+        "score": 0.8421052631578947,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 19
+    },
+    {
+        "name": "newspaper_page_parse_and_count",
+        "score": 0.5777777777777776,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 15
+    },
+    {
+        "name": "science_basic_physics",
+        "score": 0.8,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 15
+    },
+    {
+        "name": "electricity_future_prediction_from_table",
+        "score": 0.568157894736842,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 19
+    },
+    {
+        "name": "license_plate_recognition",
+        "score": 0.6428571428571429,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "places365_scene_type_classification",
+        "score": 0.9285714285714286,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "music_info_parsing",
+        "score": 0.3482142857142857,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "multilingual_movie_info_parsing",
+        "score": 0.2040816326530612,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "iconqa_count_and_reasoning",
+        "score": 0.5263157894736842,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 19
+    },
+    {
+        "name": "graph_connectivity",
+        "score": 0.5833333333333334,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 15
+    },
+    {
+        "name": "graph_shortest_path_planar",
+        "score": 0.14285714285714285,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "famous_building_recognition",
+        "score": 0.84375,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 16
+    },
+    {
+        "name": "geometry_transformation",
+        "score": 0.07142857142857142,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "long_string_letter_recognition",
+        "score": 0.0,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "handwritten_math_expression_extraction",
+        "score": 0.5,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "geometry_solid",
+        "score": 0.07142857142857142,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "animal_pose_estimation",
+        "score": 0.29270584066940136,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "single_person_pose_estimation",
+        "score": 0.36770882803106975,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "geometry_area",
+        "score": 0.21428571428571427,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "hotel_booking_confirmation_parsing",
+        "score": 0.1,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "ili_ratio_future_prediction",
+        "score": 0.0,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "electricity_load_estimate_plot",
+        "score": 0.6239285714285715,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "tqa_textbook_qa",
+        "score": 0.8571428571428571,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "stock_info_parsing",
+        "score": 0.21428571428571433,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "quizlet_question_solving",
+        "score": 0.35714285714285715,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "stock_price_future_prediction",
+        "score": 0.46871428571428575,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "Ad_count_detection",
+        "score": 0.35714285714285715,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "recover_masked_word_in_figure",
+        "score": 0.07142857142857142,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "polygon_interior_angles",
+        "score": 0.0,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "web_action_grounding",
+        "score": 0.07142857142857142,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "latex_complex_formula_convertion",
+        "score": 0.35294117647058826,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 17
+    },
+    {
+        "name": "transit_map_intersection_points",
+        "score": 0.2261904761904762,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "arxiv_vqa",
+        "score": 0.6428571428571429,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "medical_image_artifacts_indentification",
+        "score": 0.14285714285714285,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "song_title_identification_from_lyrics",
+        "score": 0.6071428571428571,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "actor_recognition_in_Movie",
+        "score": 0.7142857142857143,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "bongard_problem",
+        "score": 0.21052631578947367,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 19
+    },
+    {
+        "name": "ascii_art_understanding",
+        "score": 0.5,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "calendar_schedule_suggestion",
+        "score": 0.14285714285714285,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "geometry_reasoning_overlapped_circle",
+        "score": 0.4642857142857143,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "planning_screenshot_barman",
+        "score": 0.0,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 15
+    },
+    {
+        "name": "planning_screenshot_floortile",
+        "score": 0.0,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 15
+    },
+    {
+        "name": "medical_blood_vessels_recognition",
+        "score": 0.7142857142857143,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "location_vqa",
+        "score": 0.6428571428571429,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "mindmap_elements_parsing",
+        "score": 0.2857142857142857,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "mensa_iq_test",
+        "score": 0.3803921568627451,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 17
+    },
+    {
+        "name": "flowchart_code_generation",
+        "score": 0.7777777777777778,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 9
+    },
+    {
+        "name": "stackoverflow_debug_QA",
+        "score": 0.5714285714285714,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "logical_reasoning_find_odd_one_out",
+        "score": 0.03571428571428571,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "web_action_prediction",
+        "score": 0.42857142857142855,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "code_execution",
+        "score": 0.5,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 16
+    },
+    {
+        "name": "music_sheet_format_QA",
+        "score": 0.5714285714285714,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "annoying_word_search",
+        "score": 0.0,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "interpret_force_perspective_illusion",
+        "score": 0.7333333333333333,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 15
+    },
+    {
+        "name": "healthcare_info_judgement",
+        "score": 0.7857142857142857,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "geometry_plot_position_relationship",
+        "score": 0.7142857142857143,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "relative_depth_of_different_points",
+        "score": 0.5714285714285714,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "topological_sort",
+        "score": 0.0,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "scibench_fundamental_wo_solution",
+        "score": 0.3673469387755102,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 49
+    },
+    {
+        "name": "geometry_reasoning_nested_squares",
+        "score": 0.42857142857142855,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "font_recognition",
+        "score": 0.07142857142857142,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "geometry_reasoning_count_line_intersections",
+        "score": 0.4642857142857143,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "circuit_diagram_understanding",
+        "score": 0.06666666666666667,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 15
+    },
+    {
+        "name": "go_capture_stone",
+        "score": 0.0,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 15
+    },
+    {
+        "name": "monthly_weather_days_count",
+        "score": 0.2857142857142857,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "weather_map_climate_type_temperature_parsing",
+        "score": 0.5714285714285714,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "top_video_creator_identification",
+        "score": 0.0,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "rebus",
+        "score": 0.391304347826087,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 23
+    },
+    {
+        "name": "ishihara_test",
+        "score": 0.4857142857142857,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "paper_vqa",
+        "score": 0.21428571428571427,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "signage_navigation",
+        "score": 0.6428571428571429,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "webpage_code_understanding",
+        "score": 0.7777777777777778,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 9
+    },
+    {
+        "name": "medical_counting_lymphocytes",
+        "score": 0.0,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "game_platform_support_identification",
+        "score": 0.6428571428571429,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "GUI_Act_Mobile_swipe",
+        "score": 0.519652560877361,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 13
+    },
+    {
+        "name": "mahjong",
+        "score": 0.0,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "scibench_calculus_wo_solution",
+        "score": 0.2653061224489796,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 49
+    },
+    {
+        "name": "knowledge_graph_understanding",
+        "score": 0.4666666666666667,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 15
+    },
+    {
+        "name": "image_translation_en2cn",
+        "score": 0.42609161120798006,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 9
+    },
+    {
+        "name": "realworld_qa_en2cn",
+        "score": 0.8571428571428571,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "planning_visual_storage",
+        "score": 0.0,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 15
+    },
+    {
+        "name": "GUI_Act_Web_Multi",
+        "score": 0.4711837794298108,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "chinese_idiom_recognition",
+        "score": 0.35714285714285715,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "number_comparison",
+        "score": 0.9285714285714286,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "planning_screenshot_blocksworld",
+        "score": 0.13333333333333333,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 15
+    },
+    {
+        "name": "product_ocr_qa",
+        "score": 0.2857142857142857,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "geometry_reasoning_circled_letter",
+        "score": 0.7142857142857143,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "GUI_Act_Web_Single",
+        "score": 0.049588137529181904,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "extract_webpage_headline",
+        "score": 0.35714285714285715,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "planning_screenshot_storage",
+        "score": 0.0,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 15
+    },
+    {
+        "name": "soccer_offside",
+        "score": 0.3333333333333333,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 9
+    },
+    {
+        "name": "geometry_reasoning_grid",
+        "score": 0.6785714285714286,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "relative_reflectance_of_different_regions",
+        "score": 0.21428571428571427,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "entertainment_web_game_style",
+        "score": 0.75,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "orchestra_score_recognition",
+        "score": 0.10714285714285714,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "icon_arithmetic_puzzle",
+        "score": 0.5714285714285714,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "planning_screenshot_grippers",
+        "score": 0.13333333333333333,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 15
+    },
+    {
+        "name": "MMMU_pro_exam_screenshot",
+        "score": 0.24242424242424243,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 99
+    },
+    {
+        "name": "clevrer_physics",
+        "score": 0.55,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 20
+    },
+    {
+        "name": "MMMU_physics_chemistry_selected",
+        "score": 0.7857142857142857,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "planning_screenshot_tyreworld",
+        "score": 0.06666666666666667,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 15
+    },
+    {
+        "name": "music_sheet_note_count",
+        "score": 0.17647058823529413,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 17
+    },
+    {
+        "name": "hashtag_recommendation",
+        "score": 0.9226190476190476,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "llavaguard",
+        "score": 0.25,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "medical_multi_organ_segmentation_rater",
+        "score": 0.35714285714285715,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "cultural_vqa",
+        "score": 0.4666666666666667,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 15
+    },
+    {
+        "name": "logical_reasoning_fit_pattern",
+        "score": 0.07142857142857142,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "character_recognition_in_TV_shows",
+        "score": 0.7142857142857143,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "highest_discount_game_price_identification",
+        "score": 0.0,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "remaining_playback_time_calculation",
+        "score": 0.0,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "medical_cell_recognition",
+        "score": 0.35714285714285715,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "chess_find_legal_moves",
+        "score": 0.04744105231699729,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "distinguish_ai_generated_image",
+        "score": 0.6842105263157895,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 19
+    },
+    {
+        "name": "autonomous_driving_scene_analysis",
+        "score": 0.9285714285714286,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "counting_single_image",
+        "score": 0.8571428571428571,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "GUI_Act_Mobile_tap",
+        "score": 0.2857142857142857,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "road_map_find_highway_between_two_place",
+        "score": 0.5882352941176471,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 17
+    },
+    {
+        "name": "chess_sygyzy_endgames",
+        "score": 0.09471861471861472,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "planning_screenshot_termes",
+        "score": 0.0,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 15
+    },
+    {
+        "name": "multiple_states_identify_asia",
+        "score": 0.5428571428571428,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "multiple_states_identify_africa",
+        "score": 0.6428571428571429,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "multiple_states_identify_europe",
+        "score": 0.5857142857142857,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "multiple_states_identify_americas",
+        "score": 0.6428571428571429,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "adapted_cvbench_distance",
+        "score": 0.6428571428571429,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "adapted_cvbench_count",
+        "score": 0.42857142857142855,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "adapted_cvbench_depth",
+        "score": 1.0,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "adapted_cvbench_relation",
+        "score": 0.42857142857142855,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "symbolic_graphics_programs_computer_aided_design",
+        "score": 0.35714285714285715,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "symbolic_graphics_programs_scalable_vector_graphics",
+        "score": 0.2777777777777778,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 18
+    },
+    {
+        "name": "table_understanding_complex_question_answering",
+        "score": 0.21428571428571427,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "table_understanding_fact_verification",
+        "score": 0.7619047619047618,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "panel_images_multi_question",
+        "score": 0.8095238095238094,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "panel_images_single_question",
+        "score": 0.9285714285714286,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "MMSoc_Misinformation_GossipCop",
+        "score": 0.5714285714285714,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "MMSoc_HatefulMemes",
+        "score": 0.7857142857142857,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "MMSoc_Memotion",
+        "score": 0.6470588235294119,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 17
+    },
+    {
+        "name": "MMSoc_Misinformation_PolitiFact",
+        "score": 0.7142857142857143,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "poetry_acrostic_alliteration",
+        "score": 0.0,
+        "eval_type": "rule",
+        "num_demo": 0,
+        "num_query": 15
+    },
+    {
+        "name": "poetry_acrostic",
+        "score": 0.3333333333333333,
+        "eval_type": "rule",
+        "num_demo": 0,
+        "num_query": 15
+    },
+    {
+        "name": "poetry_limerick",
+        "score": 0.7333333333333333,
+        "eval_type": "rule",
+        "num_demo": 0,
+        "num_query": 15
+    },
+    {
+        "name": "poetry_custom_rhyming_scheme",
+        "score": 0.13333333333333333,
+        "eval_type": "rule",
+        "num_demo": 0,
+        "num_query": 15
+    },
+    {
+        "name": "poetry_petrarchian_sonnet_optional_meter",
+        "score": 0.0,
+        "eval_type": "rule",
+        "num_demo": 0,
+        "num_query": 15
+    },
+    {
+        "name": "poetry_haiku",
+        "score": 1.0,
+        "eval_type": "rule",
+        "num_demo": 0,
+        "num_query": 15
+    },
+    {
+        "name": "poetry_shakespearean_sonnet",
+        "score": 0.3333333333333333,
+        "eval_type": "rule",
+        "num_demo": 0,
+        "num_query": 15
+    },
+    {
+        "name": "screenshot_lighteval_math",
+        "score": 0.2,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 15
+    },
+    {
+        "name": "screenshot_theoremqa",
+        "score": 0.07142857142857142,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "number_puzzle_sudoku",
+        "score": 0.0,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 15
+    },
+    {
+        "name": "number_puzzle_kakuro_5x5",
+        "score": 0.0,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 15
+    },
+    {
+        "name": "text_entity_replace",
+        "score": 0.7142857142857143,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "background_change",
+        "score": 0.7857142857142857,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "face_attribute_edit",
+        "score": 0.5,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "face_swap",
+        "score": 0.5714285714285714,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "text_style",
+        "score": 0.6428571428571429,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "out_of_context",
+        "score": 0.9285714285714286,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "clip_stable_diffusion_generate",
+        "score": 0.2857142857142857,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "veracity",
+        "score": 0.8571428571428571,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "counterfactual_arithmetic",
+        "score": 0.6428571428571429,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "maze_2d_8x8",
+        "score": 0.0,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "shape_composition_shapes",
+        "score": 0.40136054421768713,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "shape_composition_colours",
+        "score": 0.4562641723356009,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "autorater_artifact",
+        "score": 0.21428571428571427,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "autorater_artifact_reason",
+        "score": 0.5333333333333333,
+        "eval_type": "rule",
+        "num_demo": 0,
+        "num_query": 15
+    },
+    {
+        "name": "chess_puzzles_crushing",
+        "score": 0.0,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "chess_puzzles_checkmate",
+        "score": 0.0,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "chess_puzzles_equality",
+        "score": 0.0,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 15
+    },
+    {
+        "name": "app_layout_understanding_notes",
+        "score": 0.21428571428571427,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "app_layout_understanding_twitter",
+        "score": 0.14285714285714285,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "app_layout_understanding_youtube",
+        "score": 0.14285714285714285,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "app_layout_understanding_tiktok",
+        "score": 0.42857142857142855,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "app_layout_understanding_excel",
+        "score": 0.07142857142857142,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "app_layout_understanding_amazon",
+        "score": 0.07142857142857142,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "app_layout_understanding_instagram",
+        "score": 0.5,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "app_layout_understanding_zoom",
+        "score": 0.3333333333333333,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 15
+    },
+    {
+        "name": "app_layout_understanding_word",
+        "score": 0.07142857142857142,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "app_layout_understanding_iphone_settings",
+        "score": 0.6428571428571429,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "app_layout_understanding_leetcode",
+        "score": 0.35714285714285715,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "app_layout_understanding_ppt",
+        "score": 0.14285714285714285,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "app_layout_understanding_alipay",
+        "score": 0.7058823529411765,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 17
+    },
+    {
+        "name": "ocr_table_to_markdown",
+        "score": 0.35714285714285715,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "ocr_table_to_latex",
+        "score": 0.6428571428571429,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "ocr_resume_employer_plain",
+        "score": 0.14285714285714285,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "ocr_article_journal",
+        "score": 0.07142857142857142,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "ocr_resume_experience_plain",
+        "score": 0.42857142857142855,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "ocr_math_text_latex",
+        "score": 0.21428571428571427,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "ocr_article_authors",
+        "score": 0.21428571428571427,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "ocr_table_to_csv",
+        "score": 0.5714285714285714,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "ocr_math_equation",
+        "score": 0.35714285714285715,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "ocr_resume_school_plain",
+        "score": 0.2857142857142857,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "ocr_table_to_html",
+        "score": 0.5,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "ocr_resume_skill_plain",
+        "score": 0.21428571428571427,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "crossword_mini_5x5",
+        "score": 0.4285714285714285,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "contain_position_length",
+        "score": 0.6,
+        "eval_type": "rule",
+        "num_demo": 0,
+        "num_query": 15
+    },
+    {
+        "name": "contain_repeat_length",
+        "score": 0.26666666666666666,
+        "eval_type": "rule",
+        "num_demo": 0,
+        "num_query": 15
+    },
+    {
+        "name": "contain_length",
+        "score": 0.8666666666666667,
+        "eval_type": "rule",
+        "num_demo": 0,
+        "num_query": 15
+    },
+    {
+        "name": "contain_contain_length",
+        "score": 1.0,
+        "eval_type": "rule",
+        "num_demo": 0,
+        "num_query": 15
+    },
+    {
+        "name": "pictionary_skribbl_io",
+        "score": 0.2,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 20
+    },
+    {
+        "name": "pictionary_doodle_guess",
+        "score": 0.8,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 15
+    },
+    {
+        "name": "pictionary_genai_output_chinese",
+        "score": 0.21428571428571427,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "pictionary_cartoon_drawing_guess",
+        "score": 0.8571428571428571,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "pictionary_chinese_food_img2en",
+        "score": 0.5714285714285714,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "reward_models_i2t_reward",
+        "score": 0.35714285714285715,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "memorization_chinese_celebrity",
+        "score": 0.21428571428571427,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "memorization_papers",
+        "score": 0.4666666666666667,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 15
+    },
+    {
+        "name": "memorization_famous_treaty",
+        "score": 0.5357142857142857,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "memorization_indian_celebrity",
+        "score": 0.5,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "research_website_parsing_blogpost",
+        "score": 0.0,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "research_website_parsing_publication",
+        "score": 0.14285714285714285,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    },
+    {
+        "name": "research_website_parsing_homepage",
+        "score": 0.21428571428571427,
+        "eval_type": "rule",
+        "num_demo": 1,
+        "num_query": 14
+    }
+]

static/eval_results/SI/Molmo_72B/{summary_results.json → summary_and_keyword_stats.json}

@@ -7,8 +7,8 @@
             "macro_mean_score": 0.36480000609384927,
             "missing_tasks": [
                 "planning_screenshot_termes",
-                "table_understanding",
-                "MMSoc_Misinformation_PolitiFact"
+                "MMSoc_Misinformation_PolitiFact",
+                "table_understanding"
             ]
         },
         "open": {

static/eval_results/SI/Molmo_72B/task_results.json

@@ -1897,14 +1897,14 @@
         "num_query": 0
     },
     {
-        "name": "table_understanding",
+        "name": "MMSoc_Misinformation_PolitiFact",
         "score": 0.0,
         "eval_type": "rule",
         "num_demo": 0,
         "num_query": 0
     },
     {
-        "name": "MMSoc_Misinformation_PolitiFact",
+        "name": "table_understanding",
         "score": 0.0,
         "eval_type": "rule",
         "num_demo": 0,