Upload eval_results/open-r1/Qwen2.5-Coder-7B-Instruct-Codeforces-GRPO-100/v01.00-step-000000010/lcb_v4/results_2025-04-29T00-26-02.026827.json with huggingface_hub

eval_results/open-r1/Qwen2.5-Coder-7B-Instruct-Codeforces-GRPO-100/v01.00-step-000000010/lcb_v4/results_2025-04-29T00-26-02.026827.json

@@ -0,0 +1,115 @@
+{
+  "config_general": {
+    "lighteval_sha": "?",
+    "num_fewshot_seeds": 1,
+    "override_batch_size": -1,
+    "max_samples": null,
+    "job_id": 0,
+    "start_time": 3319645.951731248,
+    "end_time": 3320962.31882039,
+    "total_evaluation_time_secondes": "1316.3670891420916",
+    "model_name": "open-r1/Qwen2.5-Coder-7B-Instruct-Codeforces-GRPO-100",
+    "model_sha": "",
+    "model_dtype": null,
+    "model_size": null,
+    "generation_parameters": {
+      "early_stopping": null,
+      "repetition_penalty": null,
+      "frequency_penalty": null,
+      "length_penalty": null,
+      "presence_penalty": null,
+      "max_new_tokens": 32768,
+      "min_new_tokens": null,
+      "seed": null,
+      "stop_tokens": null,
+      "temperature": 0.6,
+      "top_k": null,
+      "min_p": null,
+      "top_p": 0.95,
+      "truncate_prompt": null,
+      "response_format": null
+    }
+  },
+  "results": {
+    "extended|lcb:codegeneration_v4|0": {
+      "codegen_pass@1:16": 0.0922029702970297,
+      "codegen_pass@1:16_stderr": 0.021945062118775654
+    },
+    "all": {
+      "codegen_pass@1:16": 0.0922029702970297,
+      "codegen_pass@1:16_stderr": 0.021945062118775654
+    }
+  },
+  "versions": {
+    "extended|lcb:codegeneration_v4|0": 0
+  },
+  "config_tasks": {
+    "extended|lcb:codegeneration_v4": {
+      "name": "lcb:codegeneration_v4",
+      "prompt_function": "lcb_codegeneration_prompt_fn",
+      "hf_repo": "livecodebench/code_generation_lite",
+      "hf_subset": "v4",
+      "metric": [
+        {
+          "metric_name": "codegen_pass@1:16",
+          "higher_is_better": true,
+          "category": "5",
+          "use_case": "6",
+          "sample_level_fn": "codegen_metric",
+          "corpus_level_fn": "mean"
+        }
+      ],
+      "hf_revision": null,
+      "hf_filter": null,
+      "hf_avail_splits": [
+        "test"
+      ],
+      "trust_dataset": true,
+      "evaluation_splits": [
+        "test"
+      ],
+      "few_shots_split": null,
+      "few_shots_select": null,
+      "generation_size": 32768,
+      "generation_grammar": null,
+      "stop_sequence": [],
+      "num_samples": null,
+      "suite": [
+        "extended"
+      ],
+      "original_num_docs": 101,
+      "effective_num_docs": 101,
+      "must_remove_duplicate_docs": false,
+      "version": 0
+    }
+  },
+  "summary_tasks": {
+    "extended|lcb:codegeneration_v4|0": {
+      "hashes": {
+        "hash_examples": "ee74af48a322e28a",
+        "hash_full_prompts": "c6c49c955018e254",
+        "hash_input_tokens": "7e95092a7a37b598",
+        "hash_cont_tokens": "60e1a2ac2f9b8f81"
+      },
+      "truncated": 0,
+      "non_truncated": 101,
+      "padded": 0,
+      "non_padded": 101,
+      "effective_few_shots": 0.0,
+      "num_truncated_few_shots": 0
+    }
+  },
+  "summary_general": {
+    "hashes": {
+      "hash_examples": "3c08862e3cfc7d00",
+      "hash_full_prompts": "71a643eaec05e8bc",
+      "hash_input_tokens": "261b27bced938589",
+      "hash_cont_tokens": "91ee9485d6cba06b"
+    },
+    "truncated": 0,
+    "non_truncated": 101,
+    "padded": 0,
+    "non_padded": 101,
+    "num_truncated_few_shots": 0
+  }
+}