update results

ZeroEval-main/result_dirs/zebra-grid.summary.json

@@ -120,6 +120,17 @@
     "Total Puzzles": 1000,
     "Reason Lens": "1324.55"
   },
+  {
+    "Model": "gpt-4o-mini-2024-07-18",
+    "Mode": "greedy",
+    "Puzzle Acc": "20.10",
+    "Cell Acc": "41.26",
+    "No answer": "0.10",
+    "Easy Puzzle Acc": "62.50",
+    "Hard Puzzle Acc": "3.61",
+    "Total Puzzles": 1000,
+    "Reason Lens": "943.52"
+  },
   {
     "Model": "gemini-1.5-pro",
     "Mode": "sampling",
@@ -361,5 +372,16 @@
     "Hard Puzzle Acc": "0.00",
     "Total Puzzles": 1000,
     "Reason Lens": "1592.60"
+  },
+  {
+    "Model": "gemma-2-27b-it@vllm",
+    "Mode": "greedy",
+    "Puzzle Acc": "0.47",
+    "Cell Acc": "0.31",
+    "No answer": "96.23",
+    "Easy Puzzle Acc": "2.08",
+    "Hard Puzzle Acc": "0.00",
+    "Total Puzzles": 212,
+    "Reason Lens": "1280.62"
   }
 ]

model_info.json

@@ -32,6 +32,7 @@
     "gpt-3.5-turbo-0125": {"pretty_name": "gpt-3.5-turbo-0125", "hf_model_id": "https://platform.openai.com/"},
     "gpt-4-0125-preview": {"pretty_name": "gpt-4-0125-preview", "hf_model_id": "https://platform.openai.com/"},
     "gpt-4o-2024-05-13": {"pretty_name": "gpt-4o-2024-05-13", "hf_model_id": "https://platform.openai.com/"},
+    "gpt-4o-mini-2024-07-18": {"pretty_name": "gpt-4o-mini-2024-07-18", "hf_model_id": "https://platform.openai.com/"},
     "gpt-4-turbo-2024-04-09": {"pretty_name": "gpt-4-turbo-2024-04-09", "hf_model_id": "https://platform.openai.com/"},
     "gpt-4-0314": {"pretty_name": "gpt-4-0314", "hf_model_id": "https://platform.openai.com/"},
     "tulu-2-dpo-70b": {"pretty_name": "Tulu-2-dpo-70b", "hf_model_id": "allenai/tulu-2-dpo-70b"},