Upload eval_results/alignment-handbook/zephyr-2b-gemma-sft-v1/main/arc/results_2024-03-05T08-21-37.856919.json with huggingface_hub

eval_results/alignment-handbook/zephyr-2b-gemma-sft-v1/main/arc/results_2024-03-05T08-21-37.856919.json

@@ -0,0 +1,90 @@
+{
+  "config_general": {
+    "lighteval_sha": "?",
+    "num_fewshot_seeds": 1,
+    "override_batch_size": 1,
+    "max_samples": null,
+    "job_id": "",
+    "start_time": 209236.457369216,
+    "end_time": 209309.791886093,
+    "total_evaluation_time_secondes": "73.33451687701745",
+    "model_name": "alignment-handbook/zephyr-2b-gemma-sft-v1",
+    "model_sha": "6607539d04bc66d5856772cc04b3b1ce17ebe644",
+    "model_dtype": "torch.bfloat16",
+    "model_size": "4.68 GB",
+    "config": null
+  },
+  "results": {
+    "lighteval|arc:challenge|25": {
+      "acc": 0.4667235494880546,
+      "acc_stderr": 0.014578995859605797,
+      "acc_norm": 0.4906143344709898,
+      "acc_norm_stderr": 0.014608816322065003
+    }
+  },
+  "versions": {
+    "lighteval|arc:challenge|25": 0
+  },
+  "config_tasks": {
+    "lighteval|arc:challenge": {
+      "name": "arc:challenge",
+      "prompt_function": "arc",
+      "hf_repo": "ai2_arc",
+      "hf_subset": "ARC-Challenge",
+      "metric": [
+        "loglikelihood_acc",
+        "loglikelihood_acc_norm_nospace"
+      ],
+      "hf_avail_splits": [
+        "train",
+        "test"
+      ],
+      "evaluation_splits": [
+        "test"
+      ],
+      "few_shots_split": null,
+      "few_shots_select": "random_sampling_from_train",
+      "generation_size": 1,
+      "stop_sequence": [
+        "\n"
+      ],
+      "output_regex": null,
+      "frozen": false,
+      "suite": [
+        "lighteval",
+        "arc"
+      ],
+      "original_num_docs": 1172,
+      "effective_num_docs": 1172
+    }
+  },
+  "summary_tasks": {
+    "lighteval|arc:challenge|25": {
+      "hashes": {
+        "hash_examples": "17b0cae357c0259e",
+        "hash_full_prompts": "b6811efb46396b91",
+        "hash_input_tokens": "43252b0f9e27d08c",
+        "hash_cont_tokens": "3bb85634c18ea13e"
+      },
+      "truncated": 0,
+      "non_truncated": 1172,
+      "padded": 4664,
+      "non_padded": 23,
+      "effective_few_shots": 25.0,
+      "num_truncated_few_shots": 0
+    }
+  },
+  "summary_general": {
+    "hashes": {
+      "hash_examples": "aaa6929c6d3771fb",
+      "hash_full_prompts": "5d0ecf65b2506346",
+      "hash_input_tokens": "a0f5b6983839c6bd",
+      "hash_cont_tokens": "274bfa3146be24a9"
+    },
+    "truncated": 0,
+    "non_truncated": 1172,
+    "padded": 4664,
+    "non_padded": 23,
+    "num_truncated_few_shots": 0
+  }
+}