Upload eval_results/openchat/openchat-3.5-0106/main/hellaswag/results_2024-03-02T19-04-45.004463.json with huggingface_hub

eval_results/openchat/openchat-3.5-0106/main/hellaswag/results_2024-03-02T19-04-45.004463.json

@@ -0,0 +1,90 @@
+{
+  "config_general": {
+    "lighteval_sha": "?",
+    "num_fewshot_seeds": 1,
+    "override_batch_size": 1,
+    "max_samples": null,
+    "job_id": "",
+    "start_time": 1859322.703428361,
+    "end_time": 1859856.591927352,
+    "total_evaluation_time_secondes": "533.8884989907965",
+    "model_name": "openchat/openchat-3.5-0106",
+    "model_sha": "9619fb7d2a8e25fa6b0633c0f57f7f4aa79b45c4",
+    "model_dtype": "torch.bfloat16",
+    "model_size": "13.61 GB",
+    "config": null
+  },
+  "results": {
+    "lighteval|hellaswag|10": {
+      "acc": 0.635929097789285,
+      "acc_stderr": 0.004801852881329748,
+      "acc_norm": 0.8010356502688707,
+      "acc_norm_stderr": 0.003984052679887765
+    }
+  },
+  "versions": {
+    "lighteval|hellaswag|10": 0
+  },
+  "config_tasks": {
+    "lighteval|hellaswag": {
+      "name": "hellaswag",
+      "prompt_function": "hellaswag_harness",
+      "hf_repo": "hellaswag",
+      "hf_subset": "default",
+      "metric": [
+        "loglikelihood_acc",
+        "loglikelihood_acc_norm"
+      ],
+      "hf_avail_splits": [
+        "train",
+        "test",
+        "validation"
+      ],
+      "evaluation_splits": [
+        "validation"
+      ],
+      "few_shots_split": null,
+      "few_shots_select": "random_sampling_from_train",
+      "generation_size": -1,
+      "stop_sequence": [
+        "\n"
+      ],
+      "output_regex": null,
+      "frozen": false,
+      "suite": [
+        "lighteval"
+      ],
+      "original_num_docs": 10042,
+      "effective_num_docs": 10042
+    }
+  },
+  "summary_tasks": {
+    "lighteval|hellaswag|10": {
+      "hashes": {
+        "hash_examples": "31985c805c3a737e",
+        "hash_full_prompts": "08e85a95c77ef5d7",
+        "hash_input_tokens": "e9d46aa127260cdf",
+        "hash_cont_tokens": "c6f3c2e813309123"
+      },
+      "truncated": 0,
+      "non_truncated": 10042,
+      "padded": 39989,
+      "non_padded": 179,
+      "effective_few_shots": 10.0,
+      "num_truncated_few_shots": 0
+    }
+  },
+  "summary_general": {
+    "hashes": {
+      "hash_examples": "63bc2cf8bae03fbc",
+      "hash_full_prompts": "40a8192c4ff85272",
+      "hash_input_tokens": "c4862cdd1f33f0cf",
+      "hash_cont_tokens": "521f5aa8ef929783"
+    },
+    "truncated": 0,
+    "non_truncated": 10042,
+    "padded": 39989,
+    "non_padded": 179,
+    "num_truncated_few_shots": 0
+  }
+}