Upload eval_results/alignment-handbook/zephyr-2b-gemma-dpo-hermes-beta-0.4-epoch-2/main/hellaswag/results_2024-03-05T20-32-29.847481.json with huggingface_hub

eval_results/alignment-handbook/zephyr-2b-gemma-dpo-hermes-beta-0.4-epoch-2/main/hellaswag/results_2024-03-05T20-32-29.847481.json

@@ -0,0 +1,90 @@
+{
+  "config_general": {
+    "lighteval_sha": "?",
+    "num_fewshot_seeds": 1,
+    "override_batch_size": 1,
+    "max_samples": null,
+    "job_id": "",
+    "start_time": 892694.861568923,
+    "end_time": 893127.512257681,
+    "total_evaluation_time_secondes": "432.6506887580035",
+    "model_name": "alignment-handbook/zephyr-2b-gemma-dpo-hermes-beta-0.4-epoch-2",
+    "model_sha": "0863bbaced9b876130dc2e449cf151230926ca7e",
+    "model_dtype": "torch.bfloat16",
+    "model_size": "4.68 GB",
+    "config": null
+  },
+  "results": {
+    "lighteval|hellaswag|10": {
+      "acc": 0.5408285202150966,
+      "acc_stderr": 0.004973117975062489,
+      "acc_norm": 0.7136028679545907,
+      "acc_norm_stderr": 0.004511533039406227
+    }
+  },
+  "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": "451e22e8c3aaeeb6",
+        "hash_input_tokens": "6588539749c9d962",
+        "hash_cont_tokens": "6232829aec74d42d"
+      },
+      "truncated": 0,
+      "non_truncated": 10042,
+      "padded": 39848,
+      "non_padded": 320,
+      "effective_few_shots": 10.0,
+      "num_truncated_few_shots": 0
+    }
+  },
+  "summary_general": {
+    "hashes": {
+      "hash_examples": "63bc2cf8bae03fbc",
+      "hash_full_prompts": "b2452b7405d0326a",
+      "hash_input_tokens": "a6c61cc900d3365f",
+      "hash_cont_tokens": "c2b17126697982c3"
+    },
+    "truncated": 0,
+    "non_truncated": 10042,
+    "padded": 39848,
+    "non_padded": 320,
+    "num_truncated_few_shots": 0
+  }
+}