Upload eval_results/abacaj/phi-2-super/main/arc/results_2024-03-02T15-37-33.889816.json with huggingface_hub

eval_results/abacaj/phi-2-super/main/arc/results_2024-03-02T15-37-33.889816.json

@@ -0,0 +1,90 @@
+{
+  "config_general": {
+    "lighteval_sha": "?",
+    "num_fewshot_seeds": 1,
+    "override_batch_size": 1,
+    "max_samples": null,
+    "job_id": "",
+    "start_time": 3195090.162292721,
+    "end_time": 3195189.65956409,
+    "total_evaluation_time_secondes": "99.49727136874571",
+    "model_name": "abacaj/phi-2-super",
+    "model_sha": "f1e578c868e6cc20fb1ea8eeee427ddf6e0e2ee4",
+    "model_dtype": "torch.bfloat16",
+    "model_size": "5.19 GB",
+    "config": null
+  },
+  "results": {
+    "lighteval|arc:challenge|25": {
+      "acc": 0.5656996587030717,
+      "acc_stderr": 0.01448470304885736,
+      "acc_norm": 0.5938566552901023,
+      "acc_norm_stderr": 0.014351656690097863
+    }
+  },
+  "versions": {
+    "lighteval|arc:challenge|25": 0
+  },
+  "config_tasks": {
+    "lighteval|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": [
+        "lighteval",
+        "arc"
+      ],
+      "original_num_docs": 1172,
+      "effective_num_docs": 1172
+    }
+  },
+  "summary_tasks": {
+    "lighteval|arc:challenge|25": {
+      "hashes": {
+        "hash_examples": "17b0cae357c0259e",
+        "hash_full_prompts": "ae8b8051979d1c10",
+        "hash_input_tokens": "5944dc8232b678f9",
+        "hash_cont_tokens": "ed17e576dbafa5da"
+      },
+      "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": "bb1789810e3352b1",
+      "hash_input_tokens": "4f859a6e8fead725",
+      "hash_cont_tokens": "5a0ec6ad4b0b170d"
+    },
+    "truncated": 0,
+    "non_truncated": 1172,
+    "padded": 4660,
+    "non_padded": 27,
+    "num_truncated_few_shots": 0
+  }
+}