{
  "config_general": {
    "lighteval_sha": "?",
    "num_fewshot_seeds": 1,
    "override_batch_size": 1,
    "max_samples": null,
    "job_id": "",
    "start_time": 1769919.296175537,
    "end_time": 1770205.166829818,
    "total_evaluation_time_secondes": "285.87065428099595",
    "model_name": "Nexusflow/Starling-LM-7B-beta",
    "model_sha": "aa21e7f117d41f9463b6d48d2e127bbf2e93256d",
    "model_dtype": "torch.bfloat16",
    "model_size": "13.61 GB",
    "config": null
  },
  "results": {
    "lighteval|agieval:aqua-rat|0": {
      "acc": 0.2677165354330709,
      "acc_stderr": 0.027836648866445348,
      "acc_norm": 0.28346456692913385,
      "acc_norm_stderr": 0.02833400492130763
    },
    "lighteval|agieval:gaokao-biology|0": {
      "acc": 0.3476190476190476,
      "acc_stderr": 0.032940430891650836,
      "acc_norm": 0.38095238095238093,
      "acc_norm_stderr": 0.03359110046749991
    },
    "lighteval|agieval:gaokao-chemistry|0": {
      "acc": 0.3188405797101449,
      "acc_stderr": 0.03246964709878484,
      "acc_norm": 0.2995169082125604,
      "acc_norm_stderr": 0.031913606824066625
    },
    "lighteval|agieval:gaokao-chinese|0": {
      "acc": 0.2682926829268293,
      "acc_stderr": 0.028306754023121848,
      "acc_norm": 0.2601626016260163,
      "acc_norm_stderr": 0.028028995361669366
    },
    "lighteval|agieval:gaokao-english|0": {
      "acc": 0.6437908496732027,
      "acc_stderr": 0.027420477662629252,
      "acc_norm": 0.6339869281045751,
      "acc_norm_stderr": 0.027582811415159614
    },
    "lighteval|agieval:gaokao-geography|0": {
      "acc": 0.47738693467336685,
      "acc_stderr": 0.03549709401084567,
      "acc_norm": 0.4723618090452261,
      "acc_norm_stderr": 0.035479125346565575
    },
    "lighteval|agieval:gaokao-history|0": {
      "acc": 0.46382978723404256,
      "acc_stderr": 0.03260038511835771,
      "acc_norm": 0.4595744680851064,
      "acc_norm_stderr": 0.03257901482099834
    },
    "lighteval|agieval:gaokao-mathqa|0": {
      "acc": 0.29914529914529914,
      "acc_stderr": 0.02447490780047233,
      "acc_norm": 0.2905982905982906,
      "acc_norm_stderr": 0.024269376594480006
    },
    "lighteval|agieval:gaokao-physics|0": {
      "acc": 0.355,
      "acc_stderr": 0.033920910080708536,
      "acc_norm": 0.365,
      "acc_norm_stderr": 0.03412767927155775
    },
    "lighteval|agieval:logiqa-en|0": {
      "acc": 0.3486943164362519,
      "acc_stderr": 0.018692104055797926,
      "acc_norm": 0.3655913978494624,
      "acc_norm_stderr": 0.01888973378759108
    },
    "lighteval|agieval:logiqa-zh|0": {
      "acc": 0.3778801843317972,
      "acc_stderr": 0.01901767399112105,
      "acc_norm": 0.3901689708141321,
      "acc_norm_stderr": 0.019132619951195383
    },
    "lighteval|agieval:lsat-ar|0": {
      "acc": 0.1956521739130435,
      "acc_stderr": 0.026214799709819596,
      "acc_norm": 0.19130434782608696,
      "acc_norm_stderr": 0.025991852462828483
    },
    "lighteval|agieval:lsat-lr|0": {
      "acc": 0.48823529411764705,
      "acc_stderr": 0.02215597466931114,
      "acc_norm": 0.4568627450980392,
      "acc_norm_stderr": 0.022079476786510514
    },
    "lighteval|agieval:lsat-rc|0": {
      "acc": 0.5278810408921933,
      "acc_stderr": 0.03049483976158836,
      "acc_norm": 0.483271375464684,
      "acc_norm_stderr": 0.030525261933744594
    },
    "lighteval|agieval:sat-en|0": {
      "acc": 0.7330097087378641,
      "acc_stderr": 0.030897665523458157,
      "acc_norm": 0.6747572815533981,
      "acc_norm_stderr": 0.0327190473759639
    },
    "lighteval|agieval:sat-en-without-passage|0": {
      "acc": 0.4077669902912621,
      "acc_stderr": 0.03432222290260264,
      "acc_norm": 0.3737864077669903,
      "acc_norm_stderr": 0.03379061271978838
    },
    "lighteval|agieval:sat-math|0": {
      "acc": 0.33636363636363636,
      "acc_stderr": 0.03192622349349312,
      "acc_norm": 0.3,
      "acc_norm_stderr": 0.030966176864266674
    },
    "lighteval|agieval:_average|0": {
      "acc": 0.40335912126462936,
      "acc_stderr": 0.028775809391776958,
      "acc_norm": 0.39302120470153423,
      "acc_norm_stderr": 0.028823558641481985
    },
    "all": {
      "acc": 0.40335912126462936,
      "acc_stderr": 0.028775809391776958,
      "acc_norm": 0.39302120470153423,
      "acc_norm_stderr": 0.028823558641481985
    }
  },
  "versions": {
    "lighteval|agieval:aqua-rat|0": 0,
    "lighteval|agieval:gaokao-biology|0": 0,
    "lighteval|agieval:gaokao-chemistry|0": 0,
    "lighteval|agieval:gaokao-chinese|0": 0,
    "lighteval|agieval:gaokao-english|0": 0,
    "lighteval|agieval:gaokao-geography|0": 0,
    "lighteval|agieval:gaokao-history|0": 0,
    "lighteval|agieval:gaokao-mathqa|0": 0,
    "lighteval|agieval:gaokao-physics|0": 0,
    "lighteval|agieval:logiqa-en|0": 0,
    "lighteval|agieval:logiqa-zh|0": 0,
    "lighteval|agieval:lsat-ar|0": 0,
    "lighteval|agieval:lsat-lr|0": 0,
    "lighteval|agieval:lsat-rc|0": 0,
    "lighteval|agieval:sat-en|0": 0,
    "lighteval|agieval:sat-en-without-passage|0": 0,
    "lighteval|agieval:sat-math|0": 0
  },
  "config_tasks": {
    "lighteval|agieval:aqua-rat": {
      "name": "agieval:aqua-rat",
      "prompt_function": "agieval",
      "hf_repo": "dmayhem93/agieval-aqua-rat",
      "hf_subset": "default",
      "metric": [
        "loglikelihood_acc",
        "loglikelihood_acc_norm_nospace"
      ],
      "hf_avail_splits": [
        "test"
      ],
      "evaluation_splits": [
        "test"
      ],
      "few_shots_split": null,
      "few_shots_select": "random_sampling",
      "generation_size": 1,
      "stop_sequence": null,
      "output_regex": null,
      "frozen": false,
      "suite": [
        "lighteval"
      ],
      "original_num_docs": 254,
      "effective_num_docs": 254,
      "trust_dataset": true,
      "must_remove_duplicate_docs": null
    },
    "lighteval|agieval:gaokao-biology": {
      "name": "agieval:gaokao-biology",
      "prompt_function": "agieval",
      "hf_repo": "dmayhem93/agieval-gaokao-biology",
      "hf_subset": "default",
      "metric": [
        "loglikelihood_acc",
        "loglikelihood_acc_norm_nospace"
      ],
      "hf_avail_splits": [
        "test"
      ],
      "evaluation_splits": [
        "test"
      ],
      "few_shots_split": null,
      "few_shots_select": "random_sampling",
      "generation_size": 1,
      "stop_sequence": null,
      "output_regex": null,
      "frozen": false,
      "suite": [
        "lighteval"
      ],
      "original_num_docs": 210,
      "effective_num_docs": 210,
      "trust_dataset": true,
      "must_remove_duplicate_docs": null
    },
    "lighteval|agieval:gaokao-chemistry": {
      "name": "agieval:gaokao-chemistry",
      "prompt_function": "agieval",
      "hf_repo": "dmayhem93/agieval-gaokao-chemistry",
      "hf_subset": "default",
      "metric": [
        "loglikelihood_acc",
        "loglikelihood_acc_norm_nospace"
      ],
      "hf_avail_splits": [
        "test"
      ],
      "evaluation_splits": [
        "test"
      ],
      "few_shots_split": null,
      "few_shots_select": "random_sampling",
      "generation_size": 1,
      "stop_sequence": null,
      "output_regex": null,
      "frozen": false,
      "suite": [
        "lighteval"
      ],
      "original_num_docs": 207,
      "effective_num_docs": 207,
      "trust_dataset": true,
      "must_remove_duplicate_docs": null
    },
    "lighteval|agieval:gaokao-chinese": {
      "name": "agieval:gaokao-chinese",
      "prompt_function": "agieval",
      "hf_repo": "dmayhem93/agieval-gaokao-chinese",
      "hf_subset": "default",
      "metric": [
        "loglikelihood_acc",
        "loglikelihood_acc_norm_nospace"
      ],
      "hf_avail_splits": [
        "test"
      ],
      "evaluation_splits": [
        "test"
      ],
      "few_shots_split": null,
      "few_shots_select": "random_sampling",
      "generation_size": 1,
      "stop_sequence": null,
      "output_regex": null,
      "frozen": false,
      "suite": [
        "lighteval"
      ],
      "original_num_docs": 246,
      "effective_num_docs": 246,
      "trust_dataset": true,
      "must_remove_duplicate_docs": null
    },
    "lighteval|agieval:gaokao-english": {
      "name": "agieval:gaokao-english",
      "prompt_function": "agieval",
      "hf_repo": "dmayhem93/agieval-gaokao-english",
      "hf_subset": "default",
      "metric": [
        "loglikelihood_acc",
        "loglikelihood_acc_norm_nospace"
      ],
      "hf_avail_splits": [
        "test"
      ],
      "evaluation_splits": [
        "test"
      ],
      "few_shots_split": null,
      "few_shots_select": "random_sampling",
      "generation_size": 1,
      "stop_sequence": null,
      "output_regex": null,
      "frozen": false,
      "suite": [
        "lighteval"
      ],
      "original_num_docs": 306,
      "effective_num_docs": 306,
      "trust_dataset": true,
      "must_remove_duplicate_docs": null
    },
    "lighteval|agieval:gaokao-geography": {
      "name": "agieval:gaokao-geography",
      "prompt_function": "agieval",
      "hf_repo": "dmayhem93/agieval-gaokao-geography",
      "hf_subset": "default",
      "metric": [
        "loglikelihood_acc",
        "loglikelihood_acc_norm_nospace"
      ],
      "hf_avail_splits": [
        "test"
      ],
      "evaluation_splits": [
        "test"
      ],
      "few_shots_split": null,
      "few_shots_select": "random_sampling",
      "generation_size": 1,
      "stop_sequence": null,
      "output_regex": null,
      "frozen": false,
      "suite": [
        "lighteval"
      ],
      "original_num_docs": 199,
      "effective_num_docs": 199,
      "trust_dataset": true,
      "must_remove_duplicate_docs": null
    },
    "lighteval|agieval:gaokao-history": {
      "name": "agieval:gaokao-history",
      "prompt_function": "agieval",
      "hf_repo": "dmayhem93/agieval-gaokao-history",
      "hf_subset": "default",
      "metric": [
        "loglikelihood_acc",
        "loglikelihood_acc_norm_nospace"
      ],
      "hf_avail_splits": [
        "test"
      ],
      "evaluation_splits": [
        "test"
      ],
      "few_shots_split": null,
      "few_shots_select": "random_sampling",
      "generation_size": 1,
      "stop_sequence": null,
      "output_regex": null,
      "frozen": false,
      "suite": [
        "lighteval"
      ],
      "original_num_docs": 235,
      "effective_num_docs": 235,
      "trust_dataset": true,
      "must_remove_duplicate_docs": null
    },
    "lighteval|agieval:gaokao-mathqa": {
      "name": "agieval:gaokao-mathqa",
      "prompt_function": "agieval",
      "hf_repo": "dmayhem93/agieval-gaokao-mathqa",
      "hf_subset": "default",
      "metric": [
        "loglikelihood_acc",
        "loglikelihood_acc_norm_nospace"
      ],
      "hf_avail_splits": [
        "test"
      ],
      "evaluation_splits": [
        "test"
      ],
      "few_shots_split": null,
      "few_shots_select": "random_sampling",
      "generation_size": 1,
      "stop_sequence": null,
      "output_regex": null,
      "frozen": false,
      "suite": [
        "lighteval"
      ],
      "original_num_docs": 351,
      "effective_num_docs": 351,
      "trust_dataset": true,
      "must_remove_duplicate_docs": null
    },
    "lighteval|agieval:gaokao-physics": {
      "name": "agieval:gaokao-physics",
      "prompt_function": "agieval",
      "hf_repo": "dmayhem93/agieval-gaokao-physics",
      "hf_subset": "default",
      "metric": [
        "loglikelihood_acc",
        "loglikelihood_acc_norm_nospace"
      ],
      "hf_avail_splits": [
        "test"
      ],
      "evaluation_splits": [
        "test"
      ],
      "few_shots_split": null,
      "few_shots_select": "random_sampling",
      "generation_size": 1,
      "stop_sequence": null,
      "output_regex": null,
      "frozen": false,
      "suite": [
        "lighteval"
      ],
      "original_num_docs": 200,
      "effective_num_docs": 200,
      "trust_dataset": true,
      "must_remove_duplicate_docs": null
    },
    "lighteval|agieval:logiqa-en": {
      "name": "agieval:logiqa-en",
      "prompt_function": "agieval",
      "hf_repo": "dmayhem93/agieval-logiqa-en",
      "hf_subset": "default",
      "metric": [
        "loglikelihood_acc",
        "loglikelihood_acc_norm_nospace"
      ],
      "hf_avail_splits": [
        "test"
      ],
      "evaluation_splits": [
        "test"
      ],
      "few_shots_split": null,
      "few_shots_select": "random_sampling",
      "generation_size": 1,
      "stop_sequence": null,
      "output_regex": null,
      "frozen": false,
      "suite": [
        "lighteval"
      ],
      "original_num_docs": 651,
      "effective_num_docs": 651,
      "trust_dataset": true,
      "must_remove_duplicate_docs": null
    },
    "lighteval|agieval:logiqa-zh": {
      "name": "agieval:logiqa-zh",
      "prompt_function": "agieval",
      "hf_repo": "dmayhem93/agieval-logiqa-zh",
      "hf_subset": "default",
      "metric": [
        "loglikelihood_acc",
        "loglikelihood_acc_norm_nospace"
      ],
      "hf_avail_splits": [
        "test"
      ],
      "evaluation_splits": [
        "test"
      ],
      "few_shots_split": null,
      "few_shots_select": "random_sampling",
      "generation_size": 1,
      "stop_sequence": null,
      "output_regex": null,
      "frozen": false,
      "suite": [
        "lighteval"
      ],
      "original_num_docs": 651,
      "effective_num_docs": 651,
      "trust_dataset": true,
      "must_remove_duplicate_docs": null
    },
    "lighteval|agieval:lsat-ar": {
      "name": "agieval:lsat-ar",
      "prompt_function": "agieval",
      "hf_repo": "dmayhem93/agieval-lsat-ar",
      "hf_subset": "default",
      "metric": [
        "loglikelihood_acc",
        "loglikelihood_acc_norm_nospace"
      ],
      "hf_avail_splits": [
        "test"
      ],
      "evaluation_splits": [
        "test"
      ],
      "few_shots_split": null,
      "few_shots_select": "random_sampling",
      "generation_size": 1,
      "stop_sequence": null,
      "output_regex": null,
      "frozen": false,
      "suite": [
        "lighteval"
      ],
      "original_num_docs": 230,
      "effective_num_docs": 230,
      "trust_dataset": true,
      "must_remove_duplicate_docs": null
    },
    "lighteval|agieval:lsat-lr": {
      "name": "agieval:lsat-lr",
      "prompt_function": "agieval",
      "hf_repo": "dmayhem93/agieval-lsat-lr",
      "hf_subset": "default",
      "metric": [
        "loglikelihood_acc",
        "loglikelihood_acc_norm_nospace"
      ],
      "hf_avail_splits": [
        "test"
      ],
      "evaluation_splits": [
        "test"
      ],
      "few_shots_split": null,
      "few_shots_select": "random_sampling",
      "generation_size": 1,
      "stop_sequence": null,
      "output_regex": null,
      "frozen": false,
      "suite": [
        "lighteval"
      ],
      "original_num_docs": 510,
      "effective_num_docs": 510,
      "trust_dataset": true,
      "must_remove_duplicate_docs": null
    },
    "lighteval|agieval:lsat-rc": {
      "name": "agieval:lsat-rc",
      "prompt_function": "agieval",
      "hf_repo": "dmayhem93/agieval-lsat-rc",
      "hf_subset": "default",
      "metric": [
        "loglikelihood_acc",
        "loglikelihood_acc_norm_nospace"
      ],
      "hf_avail_splits": [
        "test"
      ],
      "evaluation_splits": [
        "test"
      ],
      "few_shots_split": null,
      "few_shots_select": "random_sampling",
      "generation_size": 1,
      "stop_sequence": null,
      "output_regex": null,
      "frozen": false,
      "suite": [
        "lighteval"
      ],
      "original_num_docs": 269,
      "effective_num_docs": 269,
      "trust_dataset": true,
      "must_remove_duplicate_docs": null
    },
    "lighteval|agieval:sat-en": {
      "name": "agieval:sat-en",
      "prompt_function": "agieval",
      "hf_repo": "dmayhem93/agieval-sat-en",
      "hf_subset": "default",
      "metric": [
        "loglikelihood_acc",
        "loglikelihood_acc_norm_nospace"
      ],
      "hf_avail_splits": [
        "test"
      ],
      "evaluation_splits": [
        "test"
      ],
      "few_shots_split": null,
      "few_shots_select": "random_sampling",
      "generation_size": 1,
      "stop_sequence": null,
      "output_regex": null,
      "frozen": false,
      "suite": [
        "lighteval"
      ],
      "original_num_docs": 206,
      "effective_num_docs": 206,
      "trust_dataset": true,
      "must_remove_duplicate_docs": null
    },
    "lighteval|agieval:sat-en-without-passage": {
      "name": "agieval:sat-en-without-passage",
      "prompt_function": "agieval",
      "hf_repo": "dmayhem93/agieval-sat-en-without-passage",
      "hf_subset": "default",
      "metric": [
        "loglikelihood_acc",
        "loglikelihood_acc_norm_nospace"
      ],
      "hf_avail_splits": [
        "test"
      ],
      "evaluation_splits": [
        "test"
      ],
      "few_shots_split": null,
      "few_shots_select": "random_sampling",
      "generation_size": 1,
      "stop_sequence": null,
      "output_regex": null,
      "frozen": false,
      "suite": [
        "lighteval"
      ],
      "original_num_docs": 206,
      "effective_num_docs": 206,
      "trust_dataset": true,
      "must_remove_duplicate_docs": null
    },
    "lighteval|agieval:sat-math": {
      "name": "agieval:sat-math",
      "prompt_function": "agieval",
      "hf_repo": "dmayhem93/agieval-sat-math",
      "hf_subset": "default",
      "metric": [
        "loglikelihood_acc",
        "loglikelihood_acc_norm_nospace"
      ],
      "hf_avail_splits": [
        "test"
      ],
      "evaluation_splits": [
        "test"
      ],
      "few_shots_split": null,
      "few_shots_select": "random_sampling",
      "generation_size": 1,
      "stop_sequence": null,
      "output_regex": null,
      "frozen": false,
      "suite": [
        "lighteval"
      ],
      "original_num_docs": 220,
      "effective_num_docs": 220,
      "trust_dataset": true,
      "must_remove_duplicate_docs": null
    }
  },
  "summary_tasks": {
    "lighteval|agieval:aqua-rat|0": {
      "hashes": {
        "hash_examples": "f09607f69e5b7525",
        "hash_full_prompts": "c885824d200848ff",
        "hash_input_tokens": "bbc9473dbbe6f16c",
        "hash_cont_tokens": "7c2bcf33ed1e5ae3"
      },
      "truncated": 0,
      "non_truncated": 254,
      "padded": 1270,
      "non_padded": 0,
      "effective_few_shots": 0.0,
      "num_truncated_few_shots": 0
    },
    "lighteval|agieval:gaokao-biology|0": {
      "hashes": {
        "hash_examples": "f262eaf4a72db963",
        "hash_full_prompts": "a57b68a06f7c5dad",
        "hash_input_tokens": "ca1fe44a4eaeed8d",
        "hash_cont_tokens": "b3febf85776696e1"
      },
      "truncated": 0,
      "non_truncated": 210,
      "padded": 840,
      "non_padded": 0,
      "effective_few_shots": 0.0,
      "num_truncated_few_shots": 0
    },
    "lighteval|agieval:gaokao-chemistry|0": {
      "hashes": {
        "hash_examples": "47f2e649f58d9da5",
        "hash_full_prompts": "77b311499beca56b",
        "hash_input_tokens": "43896e11a60e81d3",
        "hash_cont_tokens": "711fb41221b6515e"
      },
      "truncated": 0,
      "non_truncated": 207,
      "padded": 830,
      "non_padded": 1,
      "effective_few_shots": 0.0,
      "num_truncated_few_shots": 0
    },
    "lighteval|agieval:gaokao-chinese|0": {
      "hashes": {
        "hash_examples": "1010b21fde4726ab",
        "hash_full_prompts": "765b83dc3fe58e0e",
        "hash_input_tokens": "c2e6ac114faa28b7",
        "hash_cont_tokens": "393b1820b8f4534f"
      },
      "truncated": 0,
      "non_truncated": 246,
      "padded": 981,
      "non_padded": 3,
      "effective_few_shots": 0.0,
      "num_truncated_few_shots": 0
    },
    "lighteval|agieval:gaokao-english|0": {
      "hashes": {
        "hash_examples": "4864e492a350ae93",
        "hash_full_prompts": "2d90914067cbddb2",
        "hash_input_tokens": "19ee1ddb098e0782",
        "hash_cont_tokens": "dd72ceb1d7224598"
      },
      "truncated": 0,
      "non_truncated": 306,
      "padded": 1222,
      "non_padded": 2,
      "effective_few_shots": 0.0,
      "num_truncated_few_shots": 0
    },
    "lighteval|agieval:gaokao-geography|0": {
      "hashes": {
        "hash_examples": "ec3a021e37650e7d",
        "hash_full_prompts": "26abf6b8b06b03fd",
        "hash_input_tokens": "9c63079da9deb525",
        "hash_cont_tokens": "e06462bcee629ea8"
      },
      "truncated": 0,
      "non_truncated": 199,
      "padded": 794,
      "non_padded": 2,
      "effective_few_shots": 0.0,
      "num_truncated_few_shots": 0
    },
    "lighteval|agieval:gaokao-history|0": {
      "hashes": {
        "hash_examples": "b3fad1596f1ae1f9",
        "hash_full_prompts": "d22e901442960499",
        "hash_input_tokens": "cd1d7e17c8e3016d",
        "hash_cont_tokens": "87cce58a4ec6cfd8"
      },
      "truncated": 0,
      "non_truncated": 235,
      "padded": 935,
      "non_padded": 5,
      "effective_few_shots": 0.0,
      "num_truncated_few_shots": 0
    },
    "lighteval|agieval:gaokao-mathqa|0": {
      "hashes": {
        "hash_examples": "1d1088556861b0b0",
        "hash_full_prompts": "8f130ed7923a2e96",
        "hash_input_tokens": "d4e27a91afe699a4",
        "hash_cont_tokens": "771402d59229cbae"
      },
      "truncated": 0,
      "non_truncated": 351,
      "padded": 1401,
      "non_padded": 3,
      "effective_few_shots": 0.0,
      "num_truncated_few_shots": 0
    },
    "lighteval|agieval:gaokao-physics|0": {
      "hashes": {
        "hash_examples": "eb05f035c7bfca2f",
        "hash_full_prompts": "04c26f6aa4512fa0",
        "hash_input_tokens": "c571bd43f31938ce",
        "hash_cont_tokens": "f1574dddbe4231e1"
      },
      "truncated": 0,
      "non_truncated": 200,
      "padded": 797,
      "non_padded": 3,
      "effective_few_shots": 0.0,
      "num_truncated_few_shots": 0
    },
    "lighteval|agieval:logiqa-en|0": {
      "hashes": {
        "hash_examples": "0a688a45f69c21e0",
        "hash_full_prompts": "1e7fa6f1e62bea8b",
        "hash_input_tokens": "c349474caa787f2c",
        "hash_cont_tokens": "106c1564fdd5ff8e"
      },
      "truncated": 0,
      "non_truncated": 651,
      "padded": 2597,
      "non_padded": 7,
      "effective_few_shots": 0.0,
      "num_truncated_few_shots": 0
    },
    "lighteval|agieval:logiqa-zh|0": {
      "hashes": {
        "hash_examples": "620d6888b6012ea5",
        "hash_full_prompts": "f5e4b12f2be2730a",
        "hash_input_tokens": "7443875e78dae670",
        "hash_cont_tokens": "22d3640f03c44bb0"
      },
      "truncated": 0,
      "non_truncated": 651,
      "padded": 2581,
      "non_padded": 23,
      "effective_few_shots": 0.0,
      "num_truncated_few_shots": 0
    },
    "lighteval|agieval:lsat-ar|0": {
      "hashes": {
        "hash_examples": "627c8f5ccd5da209",
        "hash_full_prompts": "93f01bfefd21f036",
        "hash_input_tokens": "cd6e0cee27c3c3e4",
        "hash_cont_tokens": "9e13af9ad8f5f78e"
      },
      "truncated": 0,
      "non_truncated": 230,
      "padded": 1140,
      "non_padded": 10,
      "effective_few_shots": 0.0,
      "num_truncated_few_shots": 0
    },
    "lighteval|agieval:lsat-lr|0": {
      "hashes": {
        "hash_examples": "794641c86de172f5",
        "hash_full_prompts": "8e81a99c7521258c",
        "hash_input_tokens": "09e76ba415ce4540",
        "hash_cont_tokens": "86333847359ceee5"
      },
      "truncated": 0,
      "non_truncated": 510,
      "padded": 2525,
      "non_padded": 25,
      "effective_few_shots": 0.0,
      "num_truncated_few_shots": 0
    },
    "lighteval|agieval:lsat-rc|0": {
      "hashes": {
        "hash_examples": "35981ed917ea01cf",
        "hash_full_prompts": "b6e8d1adf72e9fb4",
        "hash_input_tokens": "78bda8444d556179",
        "hash_cont_tokens": "cb47c6b984067525"
      },
      "truncated": 0,
      "non_truncated": 269,
      "padded": 1345,
      "non_padded": 0,
      "effective_few_shots": 0.0,
      "num_truncated_few_shots": 0
    },
    "lighteval|agieval:sat-en|0": {
      "hashes": {
        "hash_examples": "041c39c646536a1e",
        "hash_full_prompts": "b9081e0a4bf5d97f",
        "hash_input_tokens": "16dc5e2c7544b9f1",
        "hash_cont_tokens": "cb01422bf828aefe"
      },
      "truncated": 0,
      "non_truncated": 206,
      "padded": 821,
      "non_padded": 0,
      "effective_few_shots": 0.0,
      "num_truncated_few_shots": 0
    },
    "lighteval|agieval:sat-en-without-passage|0": {
      "hashes": {
        "hash_examples": "e4d9284367dff68f",
        "hash_full_prompts": "f1362aa3ba355eaa",
        "hash_input_tokens": "c4236dd09d0088d9",
        "hash_cont_tokens": "cb01422bf828aefe"
      },
      "truncated": 0,
      "non_truncated": 206,
      "padded": 817,
      "non_padded": 4,
      "effective_few_shots": 0.0,
      "num_truncated_few_shots": 0
    },
    "lighteval|agieval:sat-math|0": {
      "hashes": {
        "hash_examples": "01db7291603fc1a0",
        "hash_full_prompts": "8dab33ab73deb698",
        "hash_input_tokens": "e4d4e2e413144ff5",
        "hash_cont_tokens": "0c4980b69a75cb83"
      },
      "truncated": 0,
      "non_truncated": 220,
      "padded": 876,
      "non_padded": 4,
      "effective_few_shots": 0.0,
      "num_truncated_few_shots": 0
    }
  },
  "summary_general": {
    "hashes": {
      "hash_examples": "da3af66181f18ddf",
      "hash_full_prompts": "37ef9da1700eed47",
      "hash_input_tokens": "3a23d7f630bcf75b",
      "hash_cont_tokens": "7f70e41f6e93ac40"
    },
    "truncated": 0,
    "non_truncated": 5151,
    "padded": 21772,
    "non_padded": 92,
    "num_truncated_few_shots": 0
  }
}