Upload eval_results/HuggingFaceH4/mistral-7b-odpo/v1.10/winogrande/results_2024-03-28T13-34-40.564021.json with huggingface_hub

eval_results/HuggingFaceH4/mistral-7b-odpo/v1.10/winogrande/results_2024-03-28T13-34-40.564021.json

@@ -0,0 +1,88 @@
+{
+  "config_general": {
+    "lighteval_sha": "?",
+    "num_fewshot_seeds": 1,
+    "override_batch_size": 1,
+    "max_samples": null,
+    "job_id": "",
+    "start_time": 2067692.147020078,
+    "end_time": 2067769.970545838,
+    "total_evaluation_time_secondes": "77.82352575985715",
+    "model_name": "HuggingFaceH4/mistral-7b-odpo",
+    "model_sha": "7a160cee82577cdaa7868c96120aa7af05b0ab13",
+    "model_dtype": "torch.bfloat16",
+    "model_size": "13.99 GB",
+    "config": null
+  },
+  "results": {
+    "leaderboard|winogrande|5": {
+      "acc": 0.6795580110497238,
+      "acc_stderr": 0.013115085457681717
+    }
+  },
+  "versions": {
+    "leaderboard|winogrande|5": 0
+  },
+  "config_tasks": {
+    "leaderboard|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": [
+        "leaderboard"
+      ],
+      "original_num_docs": 1267,
+      "effective_num_docs": 1267,
+      "trust_dataset": true
+    }
+  },
+  "summary_tasks": {
+    "leaderboard|winogrande|5": {
+      "hashes": {
+        "hash_examples": "087d5d1a1afd4c7b",
+        "hash_full_prompts": "68a2321c6687b7fb",
+        "hash_input_tokens": "4ee3b4a10c000d4e",
+        "hash_cont_tokens": "b62ff56a2b8ad70e"
+      },
+      "truncated": 0,
+      "non_truncated": 1267,
+      "padded": 2403,
+      "non_padded": 131,
+      "effective_few_shots": 5.0,
+      "num_truncated_few_shots": 0
+    }
+  },
+  "summary_general": {
+    "hashes": {
+      "hash_examples": "b9a49975cc41fab7",
+      "hash_full_prompts": "61d22a7151c380da",
+      "hash_input_tokens": "be615631b41f228f",
+      "hash_cont_tokens": "02fbe23ec37d1582"
+    },
+    "truncated": 0,
+    "non_truncated": 1267,
+    "padded": 2403,
+    "non_padded": 131,
+    "num_truncated_few_shots": 0
+  }
+}