Upload eval_results/deepseek-ai/deepseek-math-7b-instruct/main/math_deepseek_cot/results_2024-07-05T04-48-14.313421.json with huggingface_hub

eval_results/deepseek-ai/deepseek-math-7b-instruct/main/math_deepseek_cot/results_2024-07-05T04-48-14.313421.json

@@ -0,0 +1,93 @@
+{
+  "config_general": {
+    "lighteval_sha": "?",
+    "num_fewshot_seeds": 1,
+    "override_batch_size": 4,
+    "max_samples": null,
+    "job_id": "",
+    "start_time": 559097.243001276,
+    "end_time": 591610.628096366,
+    "total_evaluation_time_secondes": "32513.385095089907",
+    "model_name": "deepseek-ai/deepseek-math-7b-instruct",
+    "model_sha": "0a5828f800a36df0fd7f0ed581b983246c0677ff",
+    "model_dtype": "torch.bfloat16",
+    "model_size": "12.87 GB",
+    "config": null
+  },
+  "results": {
+    "custom|math_deepseek_cot|0": {
+      "qem": 0.4572,
+      "qem_stderr": 0.007045818668695732
+    },
+    "all": {
+      "qem": 0.4572,
+      "qem_stderr": 0.007045818668695732
+    }
+  },
+  "versions": {
+    "custom|math_deepseek_cot|0": 0
+  },
+  "config_tasks": {
+    "custom|math_deepseek_cot": {
+      "name": "math_deepseek_cot",
+      "prompt_function": "math_deepseek_prompt_fn",
+      "hf_repo": "lighteval/MATH",
+      "hf_subset": "all",
+      "metric": [
+        "quasi_exact_match_math"
+      ],
+      "hf_avail_splits": [
+        "train",
+        "test",
+        "validation"
+      ],
+      "evaluation_splits": [
+        "test"
+      ],
+      "few_shots_split": null,
+      "few_shots_select": null,
+      "generation_size": 4096,
+      "stop_sequence": null,
+      "output_regex": null,
+      "num_samples": null,
+      "frozen": false,
+      "suite": [
+        "custom"
+      ],
+      "original_num_docs": 5000,
+      "effective_num_docs": 5000,
+      "trust_dataset": null,
+      "must_remove_duplicate_docs": null,
+      "version": 0
+    }
+  },
+  "summary_tasks": {
+    "custom|math_deepseek_cot|0": {
+      "hashes": {
+        "hash_examples": "cf631482d6050744",
+        "hash_full_prompts": "9f97e041434639f8",
+        "hash_input_tokens": "a8167289c1361a82",
+        "hash_cont_tokens": "30c11f33ba480b04"
+      },
+      "truncated": 4997,
+      "non_truncated": 3,
+      "padded": 441,
+      "non_padded": 4559,
+      "effective_few_shots": 0.0,
+      "num_truncated_few_shots": 0
+    }
+  },
+  "summary_general": {
+    "hashes": {
+      "hash_examples": "f608c5a647279eca",
+      "hash_full_prompts": "ced6a5464ab37f30",
+      "hash_input_tokens": "4fa94cee2e1952ab",
+      "hash_cont_tokens": "2a22ced33dfc180c"
+    },
+    "truncated": 4997,
+    "non_truncated": 3,
+    "padded": 441,
+    "non_padded": 4559,
+    "num_truncated_few_shots": 0
+  }
+}