Upload eval_results/abhishek/autotrain-llama3-oh-sft-v0-2/main/arc/results_2024-05-03T11-20-23.854387.json with huggingface_hub

eval_results/abhishek/autotrain-llama3-oh-sft-v0-2/main/arc/results_2024-05-03T11-20-23.854387.json

@@ -0,0 +1,98 @@
+{
+  "config_general": {
+    "lighteval_sha": "?",
+    "num_fewshot_seeds": 1,
+    "override_batch_size": 4,
+    "max_samples": null,
+    "job_id": "",
+    "start_time": 288388.31520138,
+    "end_time": 289343.786492847,
+    "total_evaluation_time_secondes": "955.4712914670235",
+    "model_name": "abhishek/autotrain-llama3-oh-sft-v0-2",
+    "model_sha": "14e07850d3e1ee35f5788270ab514c2e3b3821bf",
+    "model_dtype": "torch.float16",
+    "model_size": "131.73 GB",
+    "config": null
+  },
+  "results": {
+    "leaderboard|arc:challenge|25": {
+      "acc": 0.6271331058020477,
+      "acc_stderr": 0.014131176760131167,
+      "acc_norm": 0.6629692832764505,
+      "acc_norm_stderr": 0.013813476652902276
+    },
+    "all": {
+      "acc": 0.6271331058020477,
+      "acc_stderr": 0.014131176760131167,
+      "acc_norm": 0.6629692832764505,
+      "acc_norm_stderr": 0.013813476652902276
+    }
+  },
+  "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": "e2882d78a54f8ba2",
+        "hash_cont_tokens": "365680daf495ecc9"
+      },
+      "truncated": 0,
+      "non_truncated": 1172,
+      "padded": 4626,
+      "non_padded": 61,
+      "effective_few_shots": 25.0,
+      "num_truncated_few_shots": 0
+    }
+  },
+  "summary_general": {
+    "hashes": {
+      "hash_examples": "aaa6929c6d3771fb",
+      "hash_full_prompts": "848db321e3b0d4cd",
+      "hash_input_tokens": "f7217796b53eed3c",
+      "hash_cont_tokens": "a52b10551e3d42ae"
+    },
+    "truncated": 0,
+    "non_truncated": 1172,
+    "padded": 4626,
+    "non_padded": 61,
+    "num_truncated_few_shots": 0
+  }
+}