Upload organize_model_results.json with huggingface_hub

organize_model_results.json

@@ -17,20 +17,28 @@
             "Qwen2-Audio-7B-Instruct": 79.0
         }
     },
-    "mmau_mini": {
-        "string_match": {
-            "gpt-4o-audio": 0.0,
-            "Qwen-Audio-Chat": 38.5,
-            "MERaLiON-AudioLLM-Whisper-SEA-LION": 60.5,
-            "Qwen2-Audio-7B-Instruct": 44.4,
-            "cascade_whisper_large_v2_gemma2_9b_cpt_sea_lionv3_instruct": 46.7,
-            "gemini-1.5-flash": 31.4,
-            "phi_4_multimodal_instruct": 54.50000000000001,
-            "seallms_audio_7b": 51.5,
-            "SALMONN_7B": 40.5,
-            "cascade_whisper_large_v3_llama_3_8b_instruct": 51.0
+    "slue_p2_sqa5_test": {
+        "llama3_70b_judge": {
+            "gpt-4o-audio": 89.41176470588235,
+            "Qwen-Audio-Chat": 79.36274509803921,
+            "MERaLiON-AudioLLM-Whisper-SEA-LION": 86.76470588235293,
+            "Qwen2-Audio-7B-Instruct": 80.04901960784315,
+            "cascade_whisper_large_v2_gemma2_9b_cpt_sea_lionv3_instruct": 88.57843137254902,
+            "phi_4_multimodal_instruct": 88.33333333333334,
+            "seallms_audio_7b": 83.52941176470588,
+            "WavLLM_fairseq": 83.92156862745098,
+            "SALMONN_7B": 83.48039215686273,
+            "cascade_whisper_large_v3_llama_3_8b_instruct": 82.99019607843137
         },
+        "gpt4o_judge": {
+            "MERaLiON-AudioLLM-Whisper-SEA-LION": 88.23529411764707,
+            "Qwen2-Audio-7B-Instruct": 84.86666666666666,
+            "cascade_whisper_large_v3_llama_3_8b_instruct": 87.79411764705883
+        }
+    },
+    "mmau_mini": {
         "llama3_70b_judge": {
+            "gpt-4o-audio": 40.6,
             "Qwen-Audio-Chat": 53.6,
             "MERaLiON-AudioLLM-Whisper-SEA-LION": 64.60000000000001,
             "Qwen2-Audio-7B-Instruct": 58.9,
@@ -41,25 +49,25 @@
             "SALMONN_7B": 48.4,
             "cascade_whisper_large_v3_llama_3_8b_instruct": 55.60000000000001
         },
+        "string_match": {
+            "Qwen-Audio-Chat": 38.5,
+            "MERaLiON-AudioLLM-Whisper-SEA-LION": 60.5,
+            "Qwen2-Audio-7B-Instruct": 44.4,
+            "cascade_whisper_large_v2_gemma2_9b_cpt_sea_lionv3_instruct": 46.7,
+            "gemini-1.5-flash": 31.4,
+            "phi_4_multimodal_instruct": 54.50000000000001,
+            "seallms_audio_7b": 51.5,
+            "SALMONN_7B": 40.5,
+            "cascade_whisper_large_v3_llama_3_8b_instruct": 51.0
+        },
         "gpt4o_judge": {
             "MERaLiON-AudioLLM-Whisper-SEA-LION": 63.9,
             "Qwen2-Audio-7B-Instruct": 53.0
         }
     },
     "mmau_mini_music": {
-        "string_match": {
-            "gpt-4o-audio": 0.0,
-            "Qwen-Audio-Chat": 0.4311377245508982,
-            "MERaLiON-AudioLLM-Whisper-SEA-LION": 0.6077844311377245,
-            "Qwen2-Audio-7B-Instruct": 0.45808383233532934,
-            "cascade_whisper_large_v2_gemma2_9b_cpt_sea_lionv3_instruct": 0.39520958083832336,
-            "gemini-1.5-flash": 0.2904191616766467,
-            "phi_4_multimodal_instruct": 0.6377245508982036,
-            "seallms_audio_7b": 0.6047904191616766,
-            "SALMONN_7B": 0.4820359281437126,
-            "cascade_whisper_large_v3_llama_3_8b_instruct": 0.5
-        },
         "llama3_70b_judge": {
+            "gpt-4o-audio": 0.2904191616766467,
             "Qwen-Audio-Chat": 0.5958083832335329,
             "MERaLiON-AudioLLM-Whisper-SEA-LION": 0.6437125748502994,
             "Qwen2-Audio-7B-Instruct": 0.6017964071856288,
@@ -70,25 +78,25 @@
             "SALMONN_7B": 0.5598802395209581,
             "cascade_whisper_large_v3_llama_3_8b_instruct": 0.5359281437125748
         },
+        "string_match": {
+            "Qwen-Audio-Chat": 0.4311377245508982,
+            "MERaLiON-AudioLLM-Whisper-SEA-LION": 0.6077844311377245,
+            "Qwen2-Audio-7B-Instruct": 0.45808383233532934,
+            "cascade_whisper_large_v2_gemma2_9b_cpt_sea_lionv3_instruct": 0.39520958083832336,
+            "gemini-1.5-flash": 0.2904191616766467,
+            "phi_4_multimodal_instruct": 0.6377245508982036,
+            "seallms_audio_7b": 0.6047904191616766,
+            "SALMONN_7B": 0.4820359281437126,
+            "cascade_whisper_large_v3_llama_3_8b_instruct": 0.5
+        },
         "gpt4o_judge": {
             "MERaLiON-AudioLLM-Whisper-SEA-LION": 0.6347305389221557,
             "Qwen2-Audio-7B-Instruct": 0.5473684210526316
         }
     },
     "mmau_mini_sound": {
-        "string_match": {
-            "gpt-4o-audio": 0.0,
-            "Qwen-Audio-Chat": 0.43543543543543545,
-            "MERaLiON-AudioLLM-Whisper-SEA-LION": 0.6606606606606606,
-            "Qwen2-Audio-7B-Instruct": 0.4744744744744745,
-            "cascade_whisper_large_v2_gemma2_9b_cpt_sea_lionv3_instruct": 0.45045045045045046,
-            "gemini-1.5-flash": 0.3483483483483483,
-            "phi_4_multimodal_instruct": 0.5975975975975976,
-            "seallms_audio_7b": 0.5165165165165165,
-            "SALMONN_7B": 0.4594594594594595,
-            "cascade_whisper_large_v3_llama_3_8b_instruct": 0.46546546546546547
-        },
         "llama3_70b_judge": {
+            "gpt-4o-audio": 0.3843843843843844,
             "Qwen-Audio-Chat": 0.5945945945945946,
             "MERaLiON-AudioLLM-Whisper-SEA-LION": 0.7027027027027027,
             "Qwen2-Audio-7B-Instruct": 0.6306306306306306,
@@ -99,25 +107,25 @@
             "SALMONN_7B": 0.5105105105105106,
             "cascade_whisper_large_v3_llama_3_8b_instruct": 0.5105105105105106
         },
+        "string_match": {
+            "Qwen-Audio-Chat": 0.43543543543543545,
+            "MERaLiON-AudioLLM-Whisper-SEA-LION": 0.6606606606606606,
+            "Qwen2-Audio-7B-Instruct": 0.4744744744744745,
+            "cascade_whisper_large_v2_gemma2_9b_cpt_sea_lionv3_instruct": 0.45045045045045046,
+            "gemini-1.5-flash": 0.3483483483483483,
+            "phi_4_multimodal_instruct": 0.5975975975975976,
+            "seallms_audio_7b": 0.5165165165165165,
+            "SALMONN_7B": 0.4594594594594595,
+            "cascade_whisper_large_v3_llama_3_8b_instruct": 0.46546546546546547
+        },
         "gpt4o_judge": {
             "MERaLiON-AudioLLM-Whisper-SEA-LION": 0.6996996996996997,
             "Qwen2-Audio-7B-Instruct": 0.5980392156862745
         }
     },
     "mmau_mini_speech": {
-        "string_match": {
-            "gpt-4o-audio": 0.0,
-            "Qwen-Audio-Chat": 0.2882882882882883,
-            "MERaLiON-AudioLLM-Whisper-SEA-LION": 0.5465465465465466,
-            "Qwen2-Audio-7B-Instruct": 0.3993993993993994,
-            "cascade_whisper_large_v2_gemma2_9b_cpt_sea_lionv3_instruct": 0.5555555555555556,
-            "gemini-1.5-flash": 0.3033033033033033,
-            "phi_4_multimodal_instruct": 0.3993993993993994,
-            "seallms_audio_7b": 0.42342342342342343,
-            "SALMONN_7B": 0.2732732732732733,
-            "cascade_whisper_large_v3_llama_3_8b_instruct": 0.5645645645645646
-        },
         "llama3_70b_judge": {
+            "gpt-4o-audio": 0.5435435435435435,
             "Qwen-Audio-Chat": 0.4174174174174174,
             "MERaLiON-AudioLLM-Whisper-SEA-LION": 0.5915915915915916,
             "Qwen2-Audio-7B-Instruct": 0.5345345345345346,
@@ -128,30 +136,22 @@
             "SALMONN_7B": 0.3813813813813814,
             "cascade_whisper_large_v3_llama_3_8b_instruct": 0.6216216216216216
         },
+        "string_match": {
+            "Qwen-Audio-Chat": 0.2882882882882883,
+            "MERaLiON-AudioLLM-Whisper-SEA-LION": 0.5465465465465466,
+            "Qwen2-Audio-7B-Instruct": 0.3993993993993994,
+            "cascade_whisper_large_v2_gemma2_9b_cpt_sea_lionv3_instruct": 0.5555555555555556,
+            "gemini-1.5-flash": 0.3033033033033033,
+            "phi_4_multimodal_instruct": 0.3993993993993994,
+            "seallms_audio_7b": 0.42342342342342343,
+            "SALMONN_7B": 0.2732732732732733,
+            "cascade_whisper_large_v3_llama_3_8b_instruct": 0.5645645645645646
+        },
         "gpt4o_judge": {
             "MERaLiON-AudioLLM-Whisper-SEA-LION": 0.5825825825825826,
             "Qwen2-Audio-7B-Instruct": 0.44660194174757284
         }
     },
-    "slue_p2_sqa5_test": {
-        "llama3_70b_judge": {
-            "gpt-4o-audio": 89.41176470588235,
-            "Qwen-Audio-Chat": 79.36274509803921,
-            "MERaLiON-AudioLLM-Whisper-SEA-LION": 86.76470588235293,
-            "Qwen2-Audio-7B-Instruct": 80.04901960784315,
-            "cascade_whisper_large_v2_gemma2_9b_cpt_sea_lionv3_instruct": 88.57843137254902,
-            "phi_4_multimodal_instruct": 88.33333333333334,
-            "seallms_audio_7b": 83.52941176470588,
-            "WavLLM_fairseq": 83.92156862745098,
-            "SALMONN_7B": 83.48039215686273,
-            "cascade_whisper_large_v3_llama_3_8b_instruct": 82.99019607843137
-        },
-        "gpt4o_judge": {
-            "MERaLiON-AudioLLM-Whisper-SEA-LION": 88.23529411764707,
-            "Qwen2-Audio-7B-Instruct": 84.86666666666666,
-            "cascade_whisper_large_v3_llama_3_8b_instruct": 87.79411764705883
-        }
-    },
     "voxceleb_accent_test": {
         "llama3_70b_judge": {
             "Qwen-Audio-Chat": 48.05088223225277,