Upload eval_results/Qwen/Qwen1.5-1.8B-Chat/main/winogrande/results_2024-03-02T15-24-35.629848.json with huggingface_hub

eval_results/Qwen/Qwen1.5-1.8B-Chat/main/winogrande/results_2024-03-02T15-24-35.629848.json

@@ -0,0 +1,87 @@
+{
+  "config_general": {
+    "lighteval_sha": "?",
+    "num_fewshot_seeds": 1,
+    "override_batch_size": 1,
+    "max_samples": null,
+    "job_id": "",
+    "start_time": 6395475.721588789,
+    "end_time": 6395533.051459764,
+    "total_evaluation_time_secondes": "57.32987097557634",
+    "model_name": "Qwen/Qwen1.5-1.8B-Chat",
+    "model_sha": "3aede71902ad578aac72678f9f8b6199ca6ab53b",
+    "model_dtype": "torch.bfloat16",
+    "model_size": "3.8 GB",
+    "config": null
+  },
+  "results": {
+    "lighteval|winogrande|5": {
+      "acc": 0.5556432517758485,
+      "acc_stderr": 0.013965196769083558
+    }
+  },
+  "versions": {
+    "lighteval|winogrande|5": 0
+  },
+  "config_tasks": {
+    "lighteval|winogrande": {
+      "name": "winogrande",
+      "prompt_function": "winogrande",
+      "hf_repo": "winogrande",
+      "hf_subset": "winogrande_xl",
+      "metric": [
+        "loglikelihood_acc"
+      ],
+      "hf_avail_splits": [
+        "train",
+        "test",
+        "validation"
+      ],
+      "evaluation_splits": [
+        "validation"
+      ],
+      "few_shots_split": null,
+      "few_shots_select": "random_sampling",
+      "generation_size": -1,
+      "stop_sequence": [
+        "\n"
+      ],
+      "output_regex": null,
+      "frozen": false,
+      "suite": [
+        "lighteval"
+      ],
+      "original_num_docs": 1267,
+      "effective_num_docs": 1267
+    }
+  },
+  "summary_tasks": {
+    "lighteval|winogrande|5": {
+      "hashes": {
+        "hash_examples": "087d5d1a1afd4c7b",
+        "hash_full_prompts": "a62b027af32f1fe4",
+        "hash_input_tokens": "2d5322eff4ad78c0",
+        "hash_cont_tokens": "af83e7e371fcf4d4"
+      },
+      "truncated": 0,
+      "non_truncated": 1267,
+      "padded": 2356,
+      "non_padded": 178,
+      "effective_few_shots": 5.0,
+      "num_truncated_few_shots": 0
+    }
+  },
+  "summary_general": {
+    "hashes": {
+      "hash_examples": "b9a49975cc41fab7",
+      "hash_full_prompts": "1d1bae16e80ffe07",
+      "hash_input_tokens": "4228ab07fdbbb16c",
+      "hash_cont_tokens": "cf29ee9ce86b26c4"
+    },
+    "truncated": 0,
+    "non_truncated": 1267,
+    "padded": 2356,
+    "non_padded": 178,
+    "num_truncated_few_shots": 0
+  }
+}