Upload eval_results/abhishek/autotrain-mixtral-8x7b-orpo-v1/main/arc/results_2024-05-01T17-43-45.291244.json with huggingface_hub

eval_results/abhishek/autotrain-mixtral-8x7b-orpo-v1/main/arc/results_2024-05-01T17-43-45.291244.json

@@ -0,0 +1,98 @@
+{
+  "config_general": {
+    "lighteval_sha": "?",
+    "num_fewshot_seeds": 1,
+    "override_batch_size": 4,
+    "max_samples": null,
+    "job_id": "",
+    "start_time": 416557.511929079,
+    "end_time": 419015.501884126,
+    "total_evaluation_time_secondes": "2457.989955046971",
+    "model_name": "abhishek/autotrain-mixtral-8x7b-orpo-v1",
+    "model_sha": "a8be37cf01ad767a0c71e0ba3af29c0b3ebcb559",
+    "model_dtype": "torch.bfloat16",
+    "model_size": "87.49 GB",
+    "config": null
+  },
+  "results": {
+    "leaderboard|arc:challenge|25": {
+      "acc": 0.6228668941979523,
+      "acc_stderr": 0.014163366896192603,
+      "acc_norm": 0.6578498293515358,
+      "acc_norm_stderr": 0.013864152159177278
+    },
+    "all": {
+      "acc": 0.6228668941979523,
+      "acc_stderr": 0.014163366896192603,
+      "acc_norm": 0.6578498293515358,
+      "acc_norm_stderr": 0.013864152159177278
+    }
+  },
+  "versions": {
+    "leaderboard|arc:challenge|25": 0
+  },
+  "config_tasks": {
+    "leaderboard|arc:challenge": {
+      "name": "arc:challenge",
+      "prompt_function": "arc",
+      "hf_repo": "ai2_arc",
+      "hf_subset": "ARC-Challenge",
+      "metric": [
+        "loglikelihood_acc",
+        "loglikelihood_acc_norm_nospace"
+      ],
+      "hf_avail_splits": [
+        "train",
+        "test"
+      ],
+      "evaluation_splits": [
+        "test"
+      ],
+      "few_shots_split": null,
+      "few_shots_select": "random_sampling_from_train",
+      "generation_size": 1,
+      "stop_sequence": [
+        "\n"
+      ],
+      "output_regex": null,
+      "frozen": false,
+      "suite": [
+        "leaderboard",
+        "arc"
+      ],
+      "original_num_docs": 1172,
+      "effective_num_docs": 1172,
+      "trust_dataset": true,
+      "must_remove_duplicate_docs": null
+    }
+  },
+  "summary_tasks": {
+    "leaderboard|arc:challenge|25": {
+      "hashes": {
+        "hash_examples": "17b0cae357c0259e",
+        "hash_full_prompts": "66498cebc864f4ca",
+        "hash_input_tokens": "c85051f999cc10e8",
+        "hash_cont_tokens": "d2a3d367d3b1fd98"
+      },
+      "truncated": 0,
+      "non_truncated": 1172,
+      "padded": 4660,
+      "non_padded": 27,
+      "effective_few_shots": 25.0,
+      "num_truncated_few_shots": 0
+    }
+  },
+  "summary_general": {
+    "hashes": {
+      "hash_examples": "aaa6929c6d3771fb",
+      "hash_full_prompts": "848db321e3b0d4cd",
+      "hash_input_tokens": "9c913c0b3ad5bc8d",
+      "hash_cont_tokens": "7291c2ed2d10d5cb"
+    },
+    "truncated": 0,
+    "non_truncated": 1172,
+    "padded": 4660,
+    "non_padded": 27,
+    "num_truncated_few_shots": 0
+  }
+}