{ "config_general": { "lighteval_sha": "?", "num_fewshot_seeds": 1, "override_batch_size": 1, "max_samples": null, "job_id": "", "start_time": 2990279.65571055, "end_time": 2990935.353819561, "total_evaluation_time_secondes": "655.6981090111658", "model_name": "meta-llama/Llama-2-70b-chat-hf", "model_sha": "e1ce257bd76895e0864f3b4d6c7ed3c4cdec93e2", "model_dtype": "torch.bfloat16", "model_size": "128.64 GB", "config": null }, "results": { "lighteval|winogrande|5": { "acc": 0.6732438831886346, "acc_stderr": 0.013181997302131368 } }, "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": "3f64226a6875ba74", "hash_input_tokens": "01aad46645262d5b", "hash_cont_tokens": "3d7ba882ca59844b" }, "truncated": 0, "non_truncated": 1267, "padded": 2377, "non_padded": 157, "effective_few_shots": 5.0, "num_truncated_few_shots": 0 } }, "summary_general": { "hashes": { "hash_examples": "b9a49975cc41fab7", "hash_full_prompts": "6d2f137a99291ce9", "hash_input_tokens": "0b392ba3a310fb51", "hash_cont_tokens": "0a779f58fe02fb16" }, "truncated": 0, "non_truncated": 1267, "padded": 2377, "non_padded": 157, "num_truncated_few_shots": 0 } }