Upload eval_results/meta-llama/Llama-2-13b-chat-hf/main/truthfulqa/results_2024-03-04T21-46-13.700082.json with huggingface_hub

eval_results/meta-llama/Llama-2-13b-chat-hf/main/truthfulqa/results_2024-03-04T21-46-13.700082.json

@@ -0,0 +1,87 @@
+{
+  "config_general": {
+    "lighteval_sha": "?",
+    "num_fewshot_seeds": 1,
+    "override_batch_size": 1,
+    "max_samples": null,
+    "job_id": "",
+    "start_time": 1534588.573615086,
+    "end_time": 1534753.53926857,
+    "total_evaluation_time_secondes": "164.96565348398872",
+    "model_name": "meta-llama/Llama-2-13b-chat-hf",
+    "model_sha": "c2f3ec81aac798ae26dcc57799a994dfbf521496",
+    "model_dtype": "torch.bfloat16",
+    "model_size": "24.32 GB",
+    "config": null
+  },
+  "results": {
+    "lighteval|truthfulqa:mc|0": {
+      "truthfulqa_mc1": 0.2692778457772338,
+      "truthfulqa_mc1_stderr": 0.01552856663708729,
+      "truthfulqa_mc2": 0.42762212617097595,
+      "truthfulqa_mc2_stderr": 0.015391942715143636
+    }
+  },
+  "versions": {
+    "lighteval|truthfulqa:mc|0": 0
+  },
+  "config_tasks": {
+    "lighteval|truthfulqa:mc": {
+      "name": "truthfulqa:mc",
+      "prompt_function": "truthful_qa_multiple_choice",
+      "hf_repo": "truthful_qa",
+      "hf_subset": "multiple_choice",
+      "metric": [
+        "truthfulqa_mc_metrics"
+      ],
+      "hf_avail_splits": [
+        "validation"
+      ],
+      "evaluation_splits": [
+        "validation"
+      ],
+      "few_shots_split": null,
+      "few_shots_select": null,
+      "generation_size": -1,
+      "stop_sequence": [
+        "\n"
+      ],
+      "output_regex": null,
+      "frozen": false,
+      "suite": [
+        "lighteval"
+      ],
+      "original_num_docs": 817,
+      "effective_num_docs": 817
+    }
+  },
+  "summary_tasks": {
+    "lighteval|truthfulqa:mc|0": {
+      "hashes": {
+        "hash_examples": "36a6d90e75d92d4a",
+        "hash_full_prompts": "bc19700dcc192702",
+        "hash_input_tokens": "691231bfc79d7533",
+        "hash_cont_tokens": "f5da56a132aab151"
+      },
+      "truncated": 0,
+      "non_truncated": 817,
+      "padded": 9623,
+      "non_padded": 373,
+      "effective_few_shots": 0.0,
+      "num_truncated_few_shots": 0
+    }
+  },
+  "summary_general": {
+    "hashes": {
+      "hash_examples": "aed1dfc67e53d0f2",
+      "hash_full_prompts": "046a77ce3ce1d5d3",
+      "hash_input_tokens": "19c9780a6766a7b6",
+      "hash_cont_tokens": "52845ca5a27c2b40"
+    },
+    "truncated": 0,
+    "non_truncated": 817,
+    "padded": 9623,
+    "non_padded": 373,
+    "num_truncated_few_shots": 0
+  }
+}