Upload eval_results/openchat/openchat-3.5-0106/main/winogrande/results_2024-03-02T18-57-22.906806.json with huggingface_hub

eval_results/openchat/openchat-3.5-0106/main/winogrande/results_2024-03-02T18-57-22.906806.json

@@ -0,0 +1,87 @@
+{
+  "config_general": {
+    "lighteval_sha": "?",
+    "num_fewshot_seeds": 1,
+    "override_batch_size": 1,
+    "max_samples": null,
+    "job_id": "",
+    "start_time": 633550.952513975,
+    "end_time": 633625.842360194,
+    "total_evaluation_time_secondes": "74.88984621898271",
+    "model_name": "openchat/openchat-3.5-0106",
+    "model_sha": "9619fb7d2a8e25fa6b0633c0f57f7f4aa79b45c4",
+    "model_dtype": "torch.bfloat16",
+    "model_size": "13.61 GB",
+    "config": null
+  },
+  "results": {
+    "lighteval|winogrande|5": {
+      "acc": 0.7253354380426204,
+      "acc_stderr": 0.012544516005117185
+    }
+  },
+  "versions": {
+    "lighteval|winogrande|5": 0
+  },
+  "config_tasks": {
+    "lighteval|winogrande": {
+      "name": "winogrande",
+      "prompt_function": "winogrande",
+      "hf_repo": "winogrande",
+      "hf_subset": "winogrande_xl",
+      "metric": [
+        "loglikelihood_acc"
+      ],
+      "hf_avail_splits": [
+        "train",
+        "test",
+        "validation"
+      ],
+      "evaluation_splits": [
+        "validation"
+      ],
+      "few_shots_split": null,
+      "few_shots_select": "random_sampling",
+      "generation_size": -1,
+      "stop_sequence": [
+        "\n"
+      ],
+      "output_regex": null,
+      "frozen": false,
+      "suite": [
+        "lighteval"
+      ],
+      "original_num_docs": 1267,
+      "effective_num_docs": 1267
+    }
+  },
+  "summary_tasks": {
+    "lighteval|winogrande|5": {
+      "hashes": {
+        "hash_examples": "087d5d1a1afd4c7b",
+        "hash_full_prompts": "83cea48ac98b7c4d",
+        "hash_input_tokens": "15ae25043c657375",
+        "hash_cont_tokens": "0dbce81de89a82c0"
+      },
+      "truncated": 0,
+      "non_truncated": 1267,
+      "padded": 2323,
+      "non_padded": 211,
+      "effective_few_shots": 5.0,
+      "num_truncated_few_shots": 0
+    }
+  },
+  "summary_general": {
+    "hashes": {
+      "hash_examples": "b9a49975cc41fab7",
+      "hash_full_prompts": "78ebe7f871366f57",
+      "hash_input_tokens": "9ae245775df1c848",
+      "hash_cont_tokens": "366469c3d3fbabf2"
+    },
+    "truncated": 0,
+    "non_truncated": 1267,
+    "padded": 2323,
+    "non_padded": 211,
+    "num_truncated_few_shots": 0
+  }
+}