Upload eval_results/AI-MO/deepseek-math-7b-sft/aimo_v03.00/aimo_kaggle/results_2024-05-07T12-15-37.241134.json with huggingface_hub

eval_results/AI-MO/deepseek-math-7b-sft/aimo_v03.00/aimo_kaggle/results_2024-05-07T12-15-37.241134.json

@@ -0,0 +1,144 @@
+{
+  "config_general": {
+    "lighteval_sha": "?",
+    "num_fewshot_seeds": 1,
+    "override_batch_size": 4,
+    "max_samples": null,
+    "job_id": "",
+    "start_time": 437176.178939583,
+    "end_time": 437485.420318817,
+    "total_evaluation_time_secondes": "309.2413792340085",
+    "model_name": "AI-MO/deepseek-math-7b-sft",
+    "model_sha": "f500a523198914c8fa81f759aed450551064ab3a",
+    "model_dtype": "torch.bfloat16",
+    "model_size": "12.93 GB",
+    "config": null
+  },
+  "results": {
+    "custom|aimo_kaggle:train|0": {
+      "qem": 0.0,
+      "qem_stderr": 0.0
+    },
+    "custom|aimo_kaggle:valid|0": {
+      "qem": 0.28,
+      "qem_stderr": 0.06414269805898185
+    },
+    "custom|aimo_kaggle:_average|0": {
+      "qem": 0.14,
+      "qem_stderr": 0.03207134902949092
+    },
+    "all": {
+      "qem": 0.14,
+      "qem_stderr": 0.03207134902949092
+    }
+  },
+  "versions": {
+    "custom|aimo_kaggle:train|0": 0,
+    "custom|aimo_kaggle:valid|0": 0
+  },
+  "config_tasks": {
+    "custom|aimo_kaggle:train": {
+      "name": "aimo_kaggle:train",
+      "prompt_function": "kaggle_prompt_fn",
+      "hf_repo": "AI-MO/kaggle-train-set",
+      "hf_subset": "",
+      "metric": [
+        "quasi_exact_match_math"
+      ],
+      "hf_avail_splits": [
+        "train"
+      ],
+      "evaluation_splits": [
+        "train"
+      ],
+      "few_shots_split": null,
+      "few_shots_select": null,
+      "generation_size": 2048,
+      "stop_sequence": null,
+      "output_regex": null,
+      "num_samples": null,
+      "frozen": false,
+      "suite": [
+        "custom"
+      ],
+      "original_num_docs": 10,
+      "effective_num_docs": 10,
+      "trust_dataset": null,
+      "must_remove_duplicate_docs": null,
+      "version": 0
+    },
+    "custom|aimo_kaggle:valid": {
+      "name": "aimo_kaggle:valid",
+      "prompt_function": "kaggle_prompt_fn",
+      "hf_repo": "AI-MO/kaggle-validation-set",
+      "hf_subset": "v0",
+      "metric": [
+        "quasi_exact_match_math"
+      ],
+      "hf_avail_splits": [
+        "train"
+      ],
+      "evaluation_splits": [
+        "train"
+      ],
+      "few_shots_split": null,
+      "few_shots_select": null,
+      "generation_size": 2048,
+      "stop_sequence": null,
+      "output_regex": null,
+      "num_samples": null,
+      "frozen": false,
+      "suite": [
+        "custom"
+      ],
+      "original_num_docs": 50,
+      "effective_num_docs": 50,
+      "trust_dataset": null,
+      "must_remove_duplicate_docs": null,
+      "version": 0
+    }
+  },
+  "summary_tasks": {
+    "custom|aimo_kaggle:train|0": {
+      "hashes": {
+        "hash_examples": "3ef01feac22949ce",
+        "hash_full_prompts": "74c9baf89033f76a",
+        "hash_input_tokens": "2f2b429412d88494",
+        "hash_cont_tokens": "28f1d3fdc9ea3afd"
+      },
+      "truncated": 10,
+      "non_truncated": 0,
+      "padded": 6,
+      "non_padded": 4,
+      "effective_few_shots": 0.0,
+      "num_truncated_few_shots": 0
+    },
+    "custom|aimo_kaggle:valid|0": {
+      "hashes": {
+        "hash_examples": "2a919e6b839e921a",
+        "hash_full_prompts": "e0424ade9e31a0fe",
+        "hash_input_tokens": "7da1857b6831dd82",
+        "hash_cont_tokens": "9139787617b042e2"
+      },
+      "truncated": 48,
+      "non_truncated": 2,
+      "padded": 34,
+      "non_padded": 16,
+      "effective_few_shots": 0.0,
+      "num_truncated_few_shots": 0
+    }
+  },
+  "summary_general": {
+    "hashes": {
+      "hash_examples": "42ad458c3603ccb6",
+      "hash_full_prompts": "51040dc54a721b0c",
+      "hash_input_tokens": "1337f29495dfd28d",
+      "hash_cont_tokens": "a4f1caa708b78677"
+    },
+    "truncated": 58,
+    "non_truncated": 2,
+    "padded": 40,
+    "non_padded": 20,
+    "num_truncated_few_shots": 0
+  }
+}