Upload eval_results/HuggingFaceTB/SmolLM2-1.7B-Instruct/main/gsm8k/results_2025-02-12T15-44-04.626650.json with huggingface_hub

eval_results/HuggingFaceTB/SmolLM2-1.7B-Instruct/main/gsm8k/results_2025-02-12T15-44-04.626650.json

@@ -0,0 +1,99 @@
+{
+  "config_general": {
+    "lighteval_sha": "?",
+    "num_fewshot_seeds": 1,
+    "override_batch_size": -1,
+    "max_samples": null,
+    "job_id": 0,
+    "start_time": 665894.934985964,
+    "end_time": 666012.292897578,
+    "total_evaluation_time_secondes": "117.35791161400266",
+    "model_name": "HuggingFaceTB/SmolLM2-1.7B-Instruct",
+    "model_sha": "",
+    "model_dtype": null,
+    "model_size": null
+  },
+  "results": {
+    "custom|gsm8k|0": {
+      "extractive_match": 0.3169067475360121,
+      "extractive_match_stderr": 0.012815868296721378
+    },
+    "all": {
+      "extractive_match": 0.3169067475360121,
+      "extractive_match_stderr": 0.012815868296721378
+    }
+  },
+  "versions": {
+    "custom|gsm8k|0": 0
+  },
+  "config_tasks": {
+    "custom|gsm8k": {
+      "name": "gsm8k",
+      "prompt_function": "gsm8k_prompt_fn",
+      "hf_repo": "gsm8k",
+      "hf_subset": "main",
+      "metric": [
+        {
+          "metric_name": "extractive_match",
+          "higher_is_better": true,
+          "category": "3",
+          "use_case": "1",
+          "sample_level_fn": "sample_level_fn",
+          "corpus_level_fn": "mean"
+        }
+      ],
+      "hf_revision": null,
+      "hf_filter": null,
+      "hf_avail_splits": [
+        "train",
+        "test"
+      ],
+      "trust_dataset": false,
+      "evaluation_splits": [
+        "test"
+      ],
+      "few_shots_split": null,
+      "few_shots_select": null,
+      "generation_size": 8192,
+      "generation_grammar": null,
+      "stop_sequence": [],
+      "num_samples": null,
+      "suite": [
+        "custom"
+      ],
+      "original_num_docs": 1319,
+      "effective_num_docs": 1319,
+      "must_remove_duplicate_docs": false,
+      "version": 0
+    }
+  },
+  "summary_tasks": {
+    "custom|gsm8k|0": {
+      "hashes": {
+        "hash_examples": "4c0843a5d99bcfdc",
+        "hash_full_prompts": "c53d40c3dafc71d9",
+        "hash_input_tokens": "df9416030feff162",
+        "hash_cont_tokens": "569fe4ad83cf9673"
+      },
+      "truncated": 0,
+      "non_truncated": 1319,
+      "padded": 0,
+      "non_padded": 1319,
+      "effective_few_shots": 0.0,
+      "num_truncated_few_shots": 0
+    }
+  },
+  "summary_general": {
+    "hashes": {
+      "hash_examples": "18b756b7813d1bdf",
+      "hash_full_prompts": "bbae0e5faff2dc56",
+      "hash_input_tokens": "6d816743a921fc1e",
+      "hash_cont_tokens": "4fe00ed950244f12"
+    },
+    "truncated": 0,
+    "non_truncated": 1319,
+    "padded": 0,
+    "non_padded": 1319,
+    "num_truncated_few_shots": 0
+  }
+}