{ "config_general": { "model_name": "meta-llama/Llama-2-13b-hf", "model_sha": "db6b8eb1feabb38985fdf785a89895959e944936", "model_size": "24.32 GB", "model_dtype": "torch.float16", "lighteval_sha": "457ac5672c5fdebfd6bc95bb94bda825c148eccf", "num_few_shot_default": 0, "num_fewshot_seeds": 1, "override_batch_size": 1, "max_samples": null, "job_id": "" }, "results": { "harness|winogrande|5": { "acc": 0.7663772691397001, "acc_stderr": 0.011892194477183524 }, "all": { "acc": 0.7663772691397001, "acc_stderr": 0.011892194477183524 } }, "versions": { "harness|winogrande|5": 0, "all": 0 }, "config_tasks": { "harness|winogrande": "LM Harness task" }, "summary_tasks": { "harness|winogrande|5": { "hashes": { "hash_examples": "aada0a176fd81218", "hash_full_prompts": "c8655cbd12de8409", "hash_input_tokens": "c0bedf98cb040854", "hash_cont_tokens": "f08975ad6f2d5864" }, "truncated": 0, "non-truncated": 2534, "padded": 2432, "non-padded": 102, "effective_few_shots": 5.0, "num_truncated_few_shots": 0 } }, "summary_general": { "hashes": { "hash_examples": "42f54c7ae3f28ef3", "hash_full_prompts": "897c968b27a8c59a", "hash_input_tokens": "ee5c3cb253d643d1", "hash_cont_tokens": "273a70958f734c00" }, "total_evaluation_time_secondes": "172.65713024139404", "truncated": 0, "non-truncated": 2534, "padded": 2432, "non-padded": 102, "num_truncated_few_shots": 0 } }