Upload eval_results/HuggingFaceH4/mistral-7b-odpo/v1.4/hellaswag/results_2024-03-28T08-56-00.633403.json with huggingface_hub

eval_results/HuggingFaceH4/mistral-7b-odpo/v1.4/hellaswag/results_2024-03-28T08-56-00.633403.json

@@ -0,0 +1,91 @@
+{
+  "config_general": {
+    "lighteval_sha": "?",
+    "num_fewshot_seeds": 1,
+    "override_batch_size": 1,
+    "max_samples": null,
+    "job_id": "",
+    "start_time": 1730567.046278846,
+    "end_time": 1731115.074954551,
+    "total_evaluation_time_secondes": "548.0286757049616",
+    "model_name": "HuggingFaceH4/mistral-7b-odpo",
+    "model_sha": "ef1ce65ffa99036cb9282be83e3c8104dbfb78f7",
+    "model_dtype": "torch.bfloat16",
+    "model_size": "13.99 GB",
+    "config": null
+  },
+  "results": {
+    "leaderboard|hellaswag|10": {
+      "acc": 0.6419040031866162,
+      "acc_stderr": 0.004784607222774644,
+      "acc_norm": 0.7972515435172276,
+      "acc_norm_stderr": 0.004012249939174918
+    }
+  },
+  "versions": {
+    "leaderboard|hellaswag|10": 0
+  },
+  "config_tasks": {
+    "leaderboard|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": [
+        "leaderboard"
+      ],
+      "original_num_docs": 10042,
+      "effective_num_docs": 10042,
+      "trust_dataset": true
+    }
+  },
+  "summary_tasks": {
+    "leaderboard|hellaswag|10": {
+      "hashes": {
+        "hash_examples": "31985c805c3a737e",
+        "hash_full_prompts": "4dd72470ed555bb2",
+        "hash_input_tokens": "e20ae2f1591baa00",
+        "hash_cont_tokens": "b230c669e22af139"
+      },
+      "truncated": 0,
+      "non_truncated": 10042,
+      "padded": 40095,
+      "non_padded": 73,
+      "effective_few_shots": 10.0,
+      "num_truncated_few_shots": 0
+    }
+  },
+  "summary_general": {
+    "hashes": {
+      "hash_examples": "63bc2cf8bae03fbc",
+      "hash_full_prompts": "5ef9c23a43597cc5",
+      "hash_input_tokens": "f0d792266afec368",
+      "hash_cont_tokens": "1578bdcfedc79874"
+    },
+    "truncated": 0,
+    "non_truncated": 10042,
+    "padded": 40095,
+    "non_padded": 73,
+    "num_truncated_few_shots": 0
+  }
+}