Upload eval_results/stabilityai/stablelm-zephyr-3b/main/truthfulqa/results_2024-03-05T21-31-29.717562.json with huggingface_hub

eval_results/stabilityai/stablelm-zephyr-3b/main/truthfulqa/results_2024-03-05T21-31-29.717562.json

@@ -0,0 +1,87 @@
+{
+  "config_general": {
+    "lighteval_sha": "?",
+    "num_fewshot_seeds": 1,
+    "override_batch_size": 1,
+    "max_samples": null,
+    "job_id": "",
+    "start_time": 665331.546510972,
+    "end_time": 665421.424897198,
+    "total_evaluation_time_secondes": "89.87838622601703",
+    "model_name": "stabilityai/stablelm-zephyr-3b",
+    "model_sha": "8b471c751c0e78cb46cf9f47738dd0eb45392071",
+    "model_dtype": "torch.bfloat16",
+    "model_size": "5.22 GB",
+    "config": null
+  },
+  "results": {
+    "lighteval|truthfulqa:mc|0": {
+      "truthfulqa_mc1": 0.31456548347613217,
+      "truthfulqa_mc1_stderr": 0.01625524199317919,
+      "truthfulqa_mc2": 0.46191847644392925,
+      "truthfulqa_mc2_stderr": 0.01652926903240728
+    }
+  },
+  "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": "d9e7f73beed694af",
+        "hash_input_tokens": "9e25a46a476ba894",
+        "hash_cont_tokens": "a8c5f18d15513ba3"
+      },
+      "truncated": 0,
+      "non_truncated": 817,
+      "padded": 9403,
+      "non_padded": 593,
+      "effective_few_shots": 0.0,
+      "num_truncated_few_shots": 0
+    }
+  },
+  "summary_general": {
+    "hashes": {
+      "hash_examples": "aed1dfc67e53d0f2",
+      "hash_full_prompts": "2227ee7c9ae23bfe",
+      "hash_input_tokens": "6fffab8cbe6f765f",
+      "hash_cont_tokens": "f03ca374e4e7e4a8"
+    },
+    "truncated": 0,
+    "non_truncated": 817,
+    "padded": 9403,
+    "non_padded": 593,
+    "num_truncated_few_shots": 0
+  }
+}