Feat: data updated

xirui-li · Jun 17, 2024 · e8d434c · e8d434c
1 parent 818b6c0
commit e8d434c
Show file tree

Hide file tree

Showing 110 changed files with 4,531 additions and 411,154 deletions.
diff --git a/data/results/Claude 3 Haiku/data.js b/data/results/Claude 3 Haiku/data.js
diff --git a/data/results/Claude 3 Opus (web)/data.js b/data/results/Claude 3 Opus (web)/data.js
@@ -0,0 +1 @@
+cache['Claude 3 Opus'] = {}
diff --git a/data/results/Claude 3 Opus (web)/meta.json b/data/results/Claude 3 Opus (web)/meta.json
@@ -0,0 +1,6 @@
+{
+    "Method": "LMM 🖼️",
+    "Source": "https://www-cdn.anthropic.com/de8ba9b01c9ab7cbabf5c33b80b7bbc618857627/Model_Card_Claude_3.pdf",
+    "Date": "2024-03-04",
+    "Comment": ""
+}
diff --git a/.../Grok-1.5V/scores_grok_1.5v_testmini.json → ...s (web)/scores_claude3_opus_testmini.json b/.../Grok-1.5V/scores_grok_1.5v_testmini.json → ...s (web)/scores_claude3_opus_testmini.json
@@ -1,6 +1,6 @@
 {
   "average": {
-    "accuracy": "52.8"
+    "accuracy": "50.5"
   },
   "task": {
     "figure question answering": {

diff --git a/data/results/Claude 3 Sonnet/data.js b/data/results/Claude 3 Sonnet/data.js
diff --git a/data/results/CoT ChatGPT (Caption+OCR)/meta.json b/data/results/CoT ChatGPT (Caption+OCR)/meta.json
diff --git a/data/results/CoT ChatGPT (Caption+OCR)/output_chatgpt_2shot_solution_use_caption_ocr.json b/data/results/CoT ChatGPT (Caption+OCR)/output_chatgpt_2shot_solution_use_caption_ocr.json
diff --git a/data/results/CoT ChatGPT (Caption+OCR)/scores_chatgpt_2shot_solution_use_caption_ocr.json b/data/results/CoT ChatGPT (Caption+OCR)/scores_chatgpt_2shot_solution_use_caption_ocr.json
diff --git a/data/results/CoT Claude (Caption+OCR)/meta.json b/data/results/CoT Claude (Caption+OCR)/meta.json
diff --git a/data/results/CoT Claude (Caption+OCR)/output_claude_2shot_solution_use_caption_ocr.json b/data/results/CoT Claude (Caption+OCR)/output_claude_2shot_solution_use_caption_ocr.json
diff --git a/data/results/CoT Claude (Caption+OCR)/scores_claude_2shot_solution_use_caption_ocr.json b/data/results/CoT Claude (Caption+OCR)/scores_claude_2shot_solution_use_caption_ocr.json
diff --git a/data/results/CoT GPT4 (Caption+OCR)/meta.json b/data/results/CoT GPT4 (Caption+OCR)/meta.json
diff --git a/data/results/CoT GPT4 (Caption+OCR)/output_gpt4_2shot_solution_use_caption_ocr.json b/data/results/CoT GPT4 (Caption+OCR)/output_gpt4_2shot_solution_use_caption_ocr.json
diff --git a/data/results/CoT GPT4 (Caption+OCR)/scores_gpt4_2shot_solution_use_caption_ocr.json b/data/results/CoT GPT4 (Caption+OCR)/scores_gpt4_2shot_solution_use_caption_ocr.json
diff --git a/data/results/GPT-4V (Playground)/meta.json → data/results/GPT-4V (web)/meta.json b/data/results/GPT-4V (Playground)/meta.json → data/results/GPT-4V (web)/meta.json
diff --git a/...Playground)/scores_gpt4v_manual_1015.json → ...PT-4V (web)/scores_gpt4v_manual_1015.json b/...Playground)/scores_gpt4v_manual_1015.json → ...PT-4V (web)/scores_gpt4v_manual_1015.json
diff --git a/data/results/GPT-4V/data.js b/data/results/GPT-4V/data.js
diff --git a/data/results/GPT-4V (IIL)/meta.json → data/results/GPT-4V/meta.json b/data/results/GPT-4V (IIL)/meta.json → data/results/GPT-4V/meta.json
diff --git a/...ts/GPT-4V (IIL)/scores_mathvista_iil.json → .../results/GPT-4V/scores_mathvista_iil.json b/...ts/GPT-4V (IIL)/scores_mathvista_iil.json → .../results/GPT-4V/scores_mathvista_iil.json
diff --git a/data/results/GPT-4o/data.js b/data/results/GPT-4o/data.js
diff --git a/...esults/InstructBLIP (Vicuna-7B)/meta.json → data/results/GPT-4o/meta.json b/...esults/InstructBLIP (Vicuna-7B)/meta.json → data/results/GPT-4o/meta.json
@@ -1,6 +1,6 @@
 {
     "Method": "LMM 🖼️",
     "Source": "https://arxiv.org/abs/2310.02255",
-    "Date": "2023-10-03",
+    "Date": "2023-10-15",
     "Comment": ""
 }
diff --git a/...0 Ultra/scores_gemini_ultra_testmini.json → ...ults/GPT-4o/scores_gpt4v_manual_1015.json b/...0 Ultra/scores_gemini_ultra_testmini.json → ...ults/GPT-4o/scores_gpt4v_manual_1015.json
@@ -1,45 +1,45 @@
 {
   "average": {
-    "accuracy": "53.0"
+    "accuracy": "49.9"
   },
   "task": {
     "figure question answering": {
-      "accuracy": "49.07"
+      "accuracy": "43.12"
     },
     "geometry problem solving": {
-      "accuracy": "56.25"
+      "accuracy": "50.48"
     },
     "math word problem": {
-      "accuracy": "53.76"
+      "accuracy": "57.53"
     },
     "textbook question answering": {
-      "accuracy": "68.99"
+      "accuracy": "65.19"
     },
     "visual question answering": {
-      "accuracy": "40.22"
+      "accuracy": "37.99"
     }
   },
   "skills": {
     "algebraic reasoning": {
-      "accuracy": "58.36"
+      "accuracy": "53.02"
     },
     "arithmetic reasoning": {
-      "accuracy": "45.89"
+      "accuracy": "49.01"
     },
     "geometry reasoning": {
-      "accuracy": "55.65"
+      "accuracy": "51.05"
     },
     "logical reasoning": {
       "accuracy": "21.62"
     },
     "numeric commonsense": {
-      "accuracy": "38.89"
+      "accuracy": "20.14"
     },
     "scientific reasoning": {
-      "accuracy": "62.30"
+      "accuracy": "63.11"
     },
     "statistical reasoning": {
-      "accuracy": "59.47"
+      "accuracy": "55.81"
     }
   }
 }
diff --git a/data/results/Gemini 1.0 Nano 1/scores_gemini_nano1_testmini.json b/data/results/Gemini 1.0 Nano 1/scores_gemini_nano1_testmini.json
diff --git a/data/results/Gemini 1.0 Nano 2/meta.json b/data/results/Gemini 1.0 Nano 2/meta.json
diff --git a/data/results/Gemini 1.0 Nano 2/scores_gemini_nano2_testmini.json b/data/results/Gemini 1.0 Nano 2/scores_gemini_nano2_testmini.json
diff --git a/data/results/Gemini 1.0 Pro/meta.json b/data/results/Gemini 1.0 Pro/meta.json
diff --git a/data/results/Gemini 1.0 Ultra/meta.json b/data/results/Gemini 1.0 Ultra/meta.json
diff --git a/data/results/Gemini Advanced (web)/data.js b/data/results/Gemini Advanced (web)/data.js