{ "config_general": { "lighteval_sha": "?", "num_fewshot_seeds": 1, "override_batch_size": 1, "max_samples": null, "job_id": "", "start_time": 5788572.599641486, "end_time": 5790051.169883963, "total_evaluation_time_secondes": "1478.5702424766496", "model_name": "deepseek-ai/deepseek-coder-33b-instruct", "model_sha": "6f09197224af9638c32c01a9060e78b0cf5a4479", "model_dtype": "torch.bfloat16", "model_size": "62.62 GB", "config": null }, "results": { "lighteval|winogrande|5": { "acc": 0.584846093133386, "acc_stderr": 0.013848684086658585 } }, "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": "007153b0b92f539b", "hash_input_tokens": "45de1668c4fb2f0a", "hash_cont_tokens": "3521e103601ec455" }, "truncated": 0, "non_truncated": 1267, "padded": 2407, "non_padded": 127, "effective_few_shots": 5.0, "num_truncated_few_shots": 0 } }, "summary_general": { "hashes": { "hash_examples": "b9a49975cc41fab7", "hash_full_prompts": "4d88626268a64dd9", "hash_input_tokens": "42ab8f0d5adcbba0", "hash_cont_tokens": "f8aef976ee8aa806" }, "truncated": 0, "non_truncated": 1267, "padded": 2407, "non_padded": 127, "num_truncated_few_shots": 0 } }