Upload eval_results/HuggingFaceH4/qwen-1.5-1.8b-dpo/v0.2/hellaswag/results_2024-03-22T22-33-31.415683.json with huggingface_hub

eval_results/HuggingFaceH4/qwen-1.5-1.8b-dpo/v0.2/hellaswag/results_2024-03-22T22-33-31.415683.json

@@ -0,0 +1,91 @@
+{
+  "config_general": {
+    "lighteval_sha": "?",
+    "num_fewshot_seeds": 1,
+    "override_batch_size": 1,
+    "max_samples": null,
+    "job_id": "",
+    "start_time": 935996.160063051,
+    "end_time": 936387.931085874,
+    "total_evaluation_time_secondes": "391.7710228229407",
+    "model_name": "HuggingFaceH4/qwen-1.5-1.8b-dpo",
+    "model_sha": "d8c1d30de4a597868755e6cb3e48d2e5abef047f",
+    "model_dtype": "torch.bfloat16",
+    "model_size": "3.79 GB",
+    "config": null
+  },
+  "results": {
+    "leaderboard|hellaswag|10": {
+      "acc": 0.4643497311292571,
+      "acc_stderr": 0.004977081808179421,
+      "acc_norm": 0.6184027086237801,
+      "acc_norm_stderr": 0.004847857546957477
+    }
+  },
+  "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": "451e22e8c3aaeeb6",
+        "hash_input_tokens": "821cee38b4253cb3",
+        "hash_cont_tokens": "3a21ed80a4a7585d"
+      },
+      "truncated": 0,
+      "non_truncated": 10042,
+      "padded": 39979,
+      "non_padded": 189,
+      "effective_few_shots": 10.0,
+      "num_truncated_few_shots": 0
+    }
+  },
+  "summary_general": {
+    "hashes": {
+      "hash_examples": "63bc2cf8bae03fbc",
+      "hash_full_prompts": "b2452b7405d0326a",
+      "hash_input_tokens": "eb87e19e7255e13b",
+      "hash_cont_tokens": "b7bfbb64b00158d2"
+    },
+    "truncated": 0,
+    "non_truncated": 10042,
+    "padded": 39979,
+    "non_padded": 189,
+    "num_truncated_few_shots": 0
+  }
+}