{ "config_general": { "lighteval_sha": "?", "num_fewshot_seeds": 1, "override_batch_size": 1, "max_samples": null, "job_id": "", "start_time": 2528263.841499473, "end_time": 2528617.22314353, "total_evaluation_time_secondes": "353.38164405710995", "model_name": "NousResearch/Nous-Hermes-2-Yi-34B", "model_sha": "fcb0a8847e76aea14aba9aa44009d4418ad7c18f", "model_dtype": "torch.bfloat16", "model_size": "64.17 GB", "config": null }, "results": { "lighteval|winogrande|5": { "acc": 0.8058405682715075, "acc_stderr": 0.01111698339239267 } }, "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": "29e044bcf40d6a6d", "hash_input_tokens": "40a1a38e49b4bbf6", "hash_cont_tokens": "ce554afdcfed3691" }, "truncated": 0, "non_truncated": 1267, "padded": 2289, "non_padded": 245, "effective_few_shots": 5.0, "num_truncated_few_shots": 0 } }, "summary_general": { "hashes": { "hash_examples": "b9a49975cc41fab7", "hash_full_prompts": "2f908b2b9b5ec583", "hash_input_tokens": "47df941b8db9c47a", "hash_cont_tokens": "8582fd788bb8f289" }, "truncated": 0, "non_truncated": 1267, "padded": 2289, "non_padded": 245, "num_truncated_few_shots": 0 } }