{ "schema": "d1-omni-public-oracle/1", "what": "the publisher's model (AutoModel.from_pretrained(trust_remote_code=True), CPU, fp32) on every public record in text mode and, when it has media, in its media mode: per question the ids, the marker positions, the raw marker logits, the temperature, the probabilities, the response", "model": { "hf_id": "LiquidAI/d1-omni-600M", "revision": "414f8d6438174f5b2133a9c21a478fc42625e308", "source_sha256": { "encoder.py": "5f2e20319ee42a7367febf06c7653d7b3f9b28ef99a64cb9ec7739563dcfcffe", "modeling_d1.py": "2b71a5c909ec5d3aa6307378a41a0fd24af1e17d5c1c1f8d510582147051a686", "prompt.py": "a6b29a55ec8345f1fc1fdbfcc4b64d80d473dc4316f095a62c51cb6cce1194cf", "vision.py": "43ad71029f82e20fd26b97574627773d47639d7a3742c8b04fe5bf55f41c4c28", "audio.py": "c5ff09d52d6a079e79d0505c89dbd0582ba2bc9abf2b26595d8d5b1d0c347bdf", "config.json": "dfccae241822f81752426629c88c8b8d7efdd03dd19418b6645c604fad43d619" }, "weights_sha256": "0713bb05270c2685ad106522f4092bceeeb3a93cf79b401f399a712296c911e1" }, "versions": { "python": "3.12.11", "platform": "macOS-27.0-arm64-arm-64bit", "torch": "2.9.0", "transformers": "5.19.0", "numpy": "2.5.3", "tokenizers": "0.23.2", "safetensors": "0.8.0", "huggingface_hub": "1.33.0", "soundfile": "0.14.0", "torchvision": "0.24.0", "pillow": "12.3.0" }, "reference": "AutoModel.from_pretrained(snapshot, trust_remote_code=True, dtype=torch.float32) (modeling_d1.D1OmniModel), CPU fp32, probabilities(state, questions, images, audio) per request", "oracle_files": [ { "file": "records_ref_audio.json", "sha256": "c483f8c511cfcdb05a8211e495a7476b11f9318899a1125c387c9e722a6d8fbc", "schema": "d1-omni-reference/1", "written": "2026-10-08T11:20:06+0900", "fixtures_sha256": "18b4ddff73a8044b63f949c69245210e1af8a26c48c5fdb785fd91a7807ead5b" }, { "file": "records_ref_images.json", "sha256": "681d2d10eef35ecb8980a3caf93d922a77a17e88ee58795468e78b915ac91b08", "schema": "d1-omni-reference/1", "written": "2026-10-08T08:38:10+0900", "fixtures_sha256": "07ea38a2c0f3d8da4afe4dd4ae79e083c46404cb8ab0ba150b9c0a3fca83ae62" }, { "file": "records_ref.json", "sha256": "e7dba6f44d0452a6aea3e401746d417503056d6f468ff72b56c5511883436c5f", "schema": "d1-omni-reference/1", "written": "2026-10-08T04:10:36+0900", "fixtures_sha256": "c0781185cf9b86210024f94fa645f047af843bfaaf7d80c703c4272634030123" } ], "rows_checked": "every row's ids equal the v3 host rows (encode_rows.v3.json, the publisher's encode)", "records": [ { "id": "aud_01", "mode": "audio", "source": "own_audio", "native": true, "prefix": 121, "usage_input_tokens": 502, "questions": [ { "qid": "topic", "type": "choice", "K": 4, "calibrate": false, "temperature_key": null, "T": null, "prefix": 121, "n_ids": 63, "positions": 184, "bucket": 256, "ids": [ 1, 17, 27377, 1139, 6634, 988, 768, 5196, 523, 18, 3493, 856, 779, 21009, 10789, 1337, 540, 19, 16, 7832, 572, 1873, 535, 835, 11746, 933, 768, 15073, 8688, 20, 19, 16, 7832, 572, 7778, 535, 61958, 774, 810, 23325, 20, 19, 16, 7832, 572, 12992, 535, 21040, 3197, 20, 19, 16, 7832, 572, 15296, 535, 36843, 3321, 1337, 779, 3714, 20, 21 ], "markers": [ 18, 31, 42, 51 ], "logits_raw": [ -2.8674635887145996, -11.332260131835938, -12.128717422485352, -0.14766040444374084 ], "probs_raw": [ 0.06181371584534645, 1.3027777640672866e-05, 5.874532689631451e-06, 0.93816739320755 ], "probs": [ 0.06181371584534645, 1.3027777640672866e-05, 5.874532689631451e-06, 0.93816739320755 ], "argmax_index": 3, "argmax": "feedback", "top2_margin": 0.8763536773622036, "near_tie": false, "gold": "delivery", "correct": false, "host_softmax_max_abs_dp": 3.725290298461914e-09, "batch": { "call": 0, "index": 0, "B": 3 } }, { "qid": "wants", "type": "noul", "K": 2, "calibrate": false, "temperature_key": null, "T": null, "prefix": 121, "n_ids": 33, "positions": 154, "bucket": 256, "ids": [ 1, 17, 27377, 1139, 6634, 988, 768, 5196, 523, 18, 4940, 779, 21009, 14065, 875, 3503, 811, 874, 4007, 540, 19, 16, 4831, 535, 1295, 20, 19, 16, 3347, 535, 18672, 20, 21 ], "markers": [ 21, 27 ], "logits_raw": [ -1.4458585977554321, -0.055500663816928864 ], "probs_raw": [ 0.19935061037540436, 0.8006494045257568 ], "probs": [ 0.8006494045257568, 0.19935061037540436 ], "argmax_index": 0, "argmax": "true", "top2_margin": 0.6012987941503525, "near_tie": false, "gold": "true", "correct": true, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 2, "B": 3 } }, { "qid": "urgency", "type": "score", "K": 3, "calibrate": false, "temperature_key": null, "T": null, "prefix": 121, "n_ids": 43, "positions": 164, "bucket": 256, "ids": [ 1, 17, 27377, 1139, 6634, 988, 768, 5196, 523, 18, 4023, 31149, 856, 779, 5940, 540, 19, 16, 2119, 730, 525, 535, 2577, 7498, 20, 19, 16, 2119, 730, 526, 535, 32238, 20, 19, 16, 2119, 730, 527, 535, 17231, 2166, 20, 21 ], "markers": [ 17, 26, 34 ], "logits_raw": [ -11.046607971191406, -1.6502888202667236, -6.171070098876953 ], "probs_raw": [ 8.212868851842359e-05, 0.9891554713249207, 0.010762522928416729 ], "probs": [ 8.212868851842359e-05, 0.9891554713249207, 0.010762522928416729 ], "argmax_index": 1, "argmax": "1", "top2_margin": 0.9783929483965039, "near_tie": false, "gold": "1", "correct": true, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 1, "B": 3 } } ], "response": { "answers": { "topic": { "type": "choice", "choice": "feedback", "confidence": 0.93816739320755, "probabilities": { "delivery": 0.06181371584534645, "billing": 1.3027777640672866e-05, "account": 5.874532689631451e-06, "feedback": 0.93816739320755 } }, "wants": { "type": "noul", "noul": 0.8006494045257568 }, "urgency": { "type": "score", "score": 1.0106805171817541, "confidence": 0.9891554713249207, "probabilities": { "0": 8.212868851842359e-05, "1": 0.9891554713249207, "2": 0.010762522928416729 }, "legend": { "0": "Can wait", "1": "Soon", "2": "Right now" } } }, "usage": { "input_tokens": 502, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "aud_01", "mode": "text", "source": "own_audio", "native": false, "prefix": 0, "usage_input_tokens": 142, "questions": [ { "qid": "topic", "type": "choice", "K": 4, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 56, "positions": 56, "bucket": 256, "ids": [ 1, 17, 27377, 1139, 6634, 988, 768, 5196, 523, 18, 3493, 856, 779, 21009, 10789, 1337, 540, 19, 16, 11746, 535, 835, 11746, 933, 768, 15073, 8688, 20, 19, 16, 5363, 800, 535, 61958, 774, 810, 23325, 20, 19, 16, 3863, 535, 21040, 3197, 20, 19, 16, 12779, 535, 36843, 3321, 1337, 779, 3714, 20, 21 ], "markers": [ 18, 29, 39, 46 ], "logits_raw": [ -2.1100118160247803, -1.3073279857635498, -3.290978193283081, 1.7141824960708618 ], "probs_raw": [ 0.020269835367798805, 0.04523257911205292, 0.0062224725261330605, 0.9282751679420471 ], "probs": [ 0.05386649817228317, 0.0955735445022583, 0.023170966655015945, 0.8273889422416687 ], "argmax_index": 3, "argmax": "feedback", "top2_margin": 0.7318153977394104, "near_tie": false, "gold": null, "correct": null, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 0, "B": 3 } }, { "qid": "wants", "type": "noul", "K": 2, "calibrate": true, "temperature_key": "noul:2", "T": 1.6663223505020142, "prefix": 0, "n_ids": 43, "positions": 43, "bucket": 256, "ids": [ 1, 17, 27377, 1139, 6634, 988, 768, 5196, 523, 18, 4940, 779, 21009, 14065, 875, 3503, 811, 874, 4007, 540, 19, 16, 4831, 535, 1295, 521, 779, 8294, 2177, 1014, 3965, 20, 19, 16, 3347, 535, 18672, 521, 779, 8294, 10995, 20, 21 ], "markers": [ 21, 33 ], "logits_raw": [ 2.693018674850464, -4.27215051651001 ], "probs_raw": [ 0.9990566372871399, 0.0009433126542717218 ], "probs": [ 0.015068508684635162, 0.984931468963623 ], "argmax_index": 1, "argmax": "false", "top2_margin": 0.9698629602789879, "near_tie": false, "gold": null, "correct": null, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 1, "B": 3 } }, { "qid": "urgency", "type": "score", "K": 3, "calibrate": true, "temperature_key": "score:3-5", "T": 1.7301132678985596, "prefix": 0, "n_ids": 43, "positions": 43, "bucket": 256, "ids": [ 1, 17, 27377, 1139, 6634, 988, 768, 5196, 523, 18, 4023, 31149, 856, 779, 5940, 540, 19, 16, 2119, 730, 525, 535, 2577, 7498, 20, 19, 16, 2119, 730, 526, 535, 32238, 20, 19, 16, 2119, 730, 527, 535, 17231, 2166, 20, 21 ], "markers": [ 17, 26, 34 ], "logits_raw": [ -8.044438362121582, -6.318297863006592, -8.428503036499023 ], "probs_raw": [ 0.13698606193065643, 0.7697144746780396, 0.09329947084188461 ], "probs": [ 0.22158345580101013, 0.6009449362754822, 0.17747168242931366 ], "argmax_index": 1, "argmax": "1", "top2_margin": 0.37936148047447205, "near_tie": false, "gold": null, "correct": null, "host_softmax_max_abs_dp": 1.4901161193847656e-08, "batch": { "call": 0, "index": 2, "B": 3 } } ], "response": { "answers": { "topic": { "type": "choice", "choice": "feedback", "confidence": 0.8273889422416687, "probabilities": { "delivery": 0.05386649817228317, "billing": 0.0955735445022583, "account": 0.023170966655015945, "feedback": 0.8273889422416687 } }, "wants": { "type": "noul", "noul": 0.015068508684635162 }, "urgency": { "type": "score", "score": 0.9558883011341095, "confidence": 0.6009449362754822, "probabilities": { "0": 0.22158345580101013, "1": 0.6009449362754822, "2": 0.17747168242931366 }, "legend": { "0": "Can wait", "1": "Soon", "2": "Right now" } } }, "usage": { "input_tokens": 142, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "aud_02", "mode": "audio", "source": "own_audio", "native": true, "prefix": 110, "usage_input_tokens": 469, "questions": [ { "qid": "topic", "type": "choice", "K": 4, "calibrate": false, "temperature_key": null, "T": null, "prefix": 110, "n_ids": 63, "positions": 173, "bucket": 256, "ids": [ 1, 17, 27377, 1139, 6634, 988, 768, 5196, 523, 18, 3493, 856, 779, 21009, 10789, 1337, 540, 19, 16, 7832, 572, 1873, 535, 835, 11746, 933, 768, 15073, 8688, 20, 19, 16, 7832, 572, 7778, 535, 61958, 774, 810, 23325, 20, 19, 16, 7832, 572, 12992, 535, 21040, 3197, 20, 19, 16, 7832, 572, 15296, 535, 36843, 3321, 1337, 779, 3714, 20, 21 ], "markers": [ 18, 31, 42, 51 ], "logits_raw": [ -11.529346466064453, -16.728370666503906, -14.800735473632812, 0.060165852308273315 ], "probs_raw": [ 9.262639650842175e-06, 5.114783263593381e-08, 3.5155119348928565e-07, 0.9999903440475464 ], "probs": [ 9.262639650842175e-06, 5.114783263593381e-08, 3.5155119348928565e-07, 0.9999903440475464 ], "argmax_index": 3, "argmax": "feedback", "top2_margin": 0.9999810814078955, "near_tie": false, "gold": "feedback", "correct": true, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 0, "B": 3 } }, { "qid": "wants", "type": "noul", "K": 2, "calibrate": false, "temperature_key": null, "T": null, "prefix": 110, "n_ids": 33, "positions": 143, "bucket": 256, "ids": [ 1, 17, 27377, 1139, 6634, 988, 768, 5196, 523, 18, 4940, 779, 21009, 14065, 875, 3503, 811, 874, 4007, 540, 19, 16, 4831, 535, 1295, 20, 19, 16, 3347, 535, 18672, 20, 21 ], "markers": [ 21, 27 ], "logits_raw": [ 0.2793175280094147, -3.478671073913574 ], "probs_raw": [ 0.9772013425827026, 0.02279871143400669 ], "probs": [ 0.02279871143400669, 0.9772013425827026 ], "argmax_index": 1, "argmax": "false", "top2_margin": 0.954402631148696, "near_tie": false, "gold": "false", "correct": true, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 2, "B": 3 } }, { "qid": "urgency", "type": "score", "K": 3, "calibrate": false, "temperature_key": null, "T": null, "prefix": 110, "n_ids": 43, "positions": 153, "bucket": 256, "ids": [ 1, 17, 27377, 1139, 6634, 988, 768, 5196, 523, 18, 4023, 31149, 856, 779, 5940, 540, 19, 16, 2119, 730, 525, 535, 2577, 7498, 20, 19, 16, 2119, 730, 526, 535, 32238, 20, 19, 16, 2119, 730, 527, 535, 17231, 2166, 20, 21 ], "markers": [ 17, 26, 34 ], "logits_raw": [ -12.738178253173828, -5.769467830657959, 1.4846340417861938 ], "probs_raw": [ 6.64972333197511e-07, 0.0007067669648677111, 0.9992924928665161 ], "probs": [ 6.64972333197511e-07, 0.0007067669648677111, 0.9992924928665161 ], "argmax_index": 2, "argmax": "2", "top2_margin": 0.9985857259016484, "near_tie": false, "gold": "0", "correct": false, "host_softmax_max_abs_dp": 5.820766091346741e-11, "batch": { "call": 0, "index": 1, "B": 3 } } ], "response": { "answers": { "topic": { "type": "choice", "choice": "feedback", "confidence": 0.9999903440475464, "probabilities": { "delivery": 9.262639650842175e-06, "billing": 5.114783263593381e-08, "account": 3.5155119348928565e-07, "feedback": 0.9999903440475464 } }, "wants": { "type": "noul", "noul": 0.02279871143400669 }, "urgency": { "type": "score", "score": 1.9992917526979, "confidence": 0.9992924928665161, "probabilities": { "0": 6.64972333197511e-07, "1": 0.0007067669648677111, "2": 0.9992924928665161 }, "legend": { "0": "Can wait", "1": "Soon", "2": "Right now" } } }, "usage": { "input_tokens": 469, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "aud_02", "mode": "text", "source": "own_audio", "native": false, "prefix": 0, "usage_input_tokens": 142, "questions": [ { "qid": "topic", "type": "choice", "K": 4, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 56, "positions": 56, "bucket": 256, "ids": [ 1, 17, 27377, 1139, 6634, 988, 768, 5196, 523, 18, 3493, 856, 779, 21009, 10789, 1337, 540, 19, 16, 11746, 535, 835, 11746, 933, 768, 15073, 8688, 20, 19, 16, 5363, 800, 535, 61958, 774, 810, 23325, 20, 19, 16, 3863, 535, 21040, 3197, 20, 19, 16, 12779, 535, 36843, 3321, 1337, 779, 3714, 20, 21 ], "markers": [ 18, 29, 39, 46 ], "logits_raw": [ -2.1100118160247803, -1.3073279857635498, -3.290978193283081, 1.7141824960708618 ], "probs_raw": [ 0.020269835367798805, 0.04523257911205292, 0.0062224725261330605, 0.9282751679420471 ], "probs": [ 0.05386649817228317, 0.0955735445022583, 0.023170966655015945, 0.8273889422416687 ], "argmax_index": 3, "argmax": "feedback", "top2_margin": 0.7318153977394104, "near_tie": false, "gold": null, "correct": null, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 0, "B": 3 } }, { "qid": "wants", "type": "noul", "K": 2, "calibrate": true, "temperature_key": "noul:2", "T": 1.6663223505020142, "prefix": 0, "n_ids": 43, "positions": 43, "bucket": 256, "ids": [ 1, 17, 27377, 1139, 6634, 988, 768, 5196, 523, 18, 4940, 779, 21009, 14065, 875, 3503, 811, 874, 4007, 540, 19, 16, 4831, 535, 1295, 521, 779, 8294, 2177, 1014, 3965, 20, 19, 16, 3347, 535, 18672, 521, 779, 8294, 10995, 20, 21 ], "markers": [ 21, 33 ], "logits_raw": [ 2.693018674850464, -4.27215051651001 ], "probs_raw": [ 0.9990566372871399, 0.0009433126542717218 ], "probs": [ 0.015068508684635162, 0.984931468963623 ], "argmax_index": 1, "argmax": "false", "top2_margin": 0.9698629602789879, "near_tie": false, "gold": null, "correct": null, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 1, "B": 3 } }, { "qid": "urgency", "type": "score", "K": 3, "calibrate": true, "temperature_key": "score:3-5", "T": 1.7301132678985596, "prefix": 0, "n_ids": 43, "positions": 43, "bucket": 256, "ids": [ 1, 17, 27377, 1139, 6634, 988, 768, 5196, 523, 18, 4023, 31149, 856, 779, 5940, 540, 19, 16, 2119, 730, 525, 535, 2577, 7498, 20, 19, 16, 2119, 730, 526, 535, 32238, 20, 19, 16, 2119, 730, 527, 535, 17231, 2166, 20, 21 ], "markers": [ 17, 26, 34 ], "logits_raw": [ -8.044438362121582, -6.318297863006592, -8.428503036499023 ], "probs_raw": [ 0.13698606193065643, 0.7697144746780396, 0.09329947084188461 ], "probs": [ 0.22158345580101013, 0.6009449362754822, 0.17747168242931366 ], "argmax_index": 1, "argmax": "1", "top2_margin": 0.37936148047447205, "near_tie": false, "gold": null, "correct": null, "host_softmax_max_abs_dp": 1.4901161193847656e-08, "batch": { "call": 0, "index": 2, "B": 3 } } ], "response": { "answers": { "topic": { "type": "choice", "choice": "feedback", "confidence": 0.8273889422416687, "probabilities": { "delivery": 0.05386649817228317, "billing": 0.0955735445022583, "account": 0.023170966655015945, "feedback": 0.8273889422416687 } }, "wants": { "type": "noul", "noul": 0.015068508684635162 }, "urgency": { "type": "score", "score": 0.9558883011341095, "confidence": 0.6009449362754822, "probabilities": { "0": 0.22158345580101013, "1": 0.6009449362754822, "2": 0.17747168242931366 }, "legend": { "0": "Can wait", "1": "Soon", "2": "Right now" } } }, "usage": { "input_tokens": 142, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "aud_03", "mode": "audio", "source": "own_audio", "native": true, "prefix": 110, "usage_input_tokens": 451, "questions": [ { "qid": "topic", "type": "choice", "K": 4, "calibrate": false, "temperature_key": null, "T": null, "prefix": 110, "n_ids": 57, "positions": 167, "bucket": 256, "ids": [ 1, 17, 24373, 18, 3493, 856, 779, 21009, 10789, 1337, 540, 19, 16, 7832, 572, 1873, 535, 835, 11746, 933, 768, 15073, 8688, 20, 19, 16, 7832, 572, 7778, 535, 61958, 774, 810, 23325, 20, 19, 16, 7832, 572, 12992, 535, 21040, 3197, 20, 19, 16, 7832, 572, 15296, 535, 36843, 3321, 1337, 779, 3714, 20, 21 ], "markers": [ 12, 25, 36, 45 ], "logits_raw": [ -8.26899528503418, -5.348294258117676, -9.974102020263672, -1.3370122909545898 ], "probs_raw": [ 0.0009576162556186318, 0.0177678894251585, 0.0001740496081765741, 0.9811004400253296 ], "probs": [ 0.0009576162556186318, 0.0177678894251585, 0.0001740496081765741, 0.9811004400253296 ], "argmax_index": 3, "argmax": "feedback", "top2_margin": 0.9633325506001711, "near_tie": false, "gold": "account", "correct": false, "host_softmax_max_abs_dp": 1.4551915228366852e-11, "batch": { "call": 0, "index": 0, "B": 3 } }, { "qid": "wants", "type": "noul", "K": 2, "calibrate": false, "temperature_key": null, "T": null, "prefix": 110, "n_ids": 27, "positions": 137, "bucket": 256, "ids": [ 1, 17, 24373, 18, 4940, 779, 21009, 14065, 875, 3503, 811, 874, 4007, 540, 19, 16, 4831, 535, 1295, 20, 19, 16, 3347, 535, 18672, 20, 21 ], "markers": [ 15, 21 ], "logits_raw": [ -1.357582926750183, -1.9369168281555176 ], "probs_raw": [ 0.6409140825271606, 0.3590858578681946 ], "probs": [ 0.3590858578681946, 0.6409140825271606 ], "argmax_index": 1, "argmax": "false", "top2_margin": 0.28182822465896606, "near_tie": false, "gold": "true", "correct": false, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 2, "B": 3 } }, { "qid": "urgency", "type": "score", "K": 3, "calibrate": false, "temperature_key": null, "T": null, "prefix": 110, "n_ids": 37, "positions": 147, "bucket": 256, "ids": [ 1, 17, 24373, 18, 4023, 31149, 856, 779, 5940, 540, 19, 16, 2119, 730, 525, 535, 2577, 7498, 20, 19, 16, 2119, 730, 526, 535, 32238, 20, 19, 16, 2119, 730, 527, 535, 17231, 2166, 20, 21 ], "markers": [ 11, 20, 28 ], "logits_raw": [ -13.345606803894043, -3.927008628845215, 1.5194247961044312 ], "probs_raw": [ 3.4860235587075294e-07, 0.004293142817914486, 0.9957064986228943 ], "probs": [ 3.4860235587075294e-07, 0.004293142817914486, 0.9957064986228943 ], "argmax_index": 2, "argmax": "2", "top2_margin": 0.9914133558049798, "near_tie": false, "gold": "2", "correct": true, "host_softmax_max_abs_dp": 4.656612873077393e-10, "batch": { "call": 0, "index": 1, "B": 3 } } ], "response": { "answers": { "topic": { "type": "choice", "choice": "feedback", "confidence": 0.9811004400253296, "probabilities": { "delivery": 0.0009576162556186318, "billing": 0.0177678894251585, "account": 0.0001740496081765741, "feedback": 0.9811004400253296 } }, "wants": { "type": "noul", "noul": 0.3590858578681946 }, "urgency": { "type": "score", "score": 1.995706140063703, "confidence": 0.9957064986228943, "probabilities": { "0": 3.4860235587075294e-07, "1": 0.004293142817914486, "2": 0.9957064986228943 }, "legend": { "0": "Can wait", "1": "Soon", "2": "Right now" } } }, "usage": { "input_tokens": 451, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "aud_03", "mode": "text", "source": "own_audio", "native": false, "prefix": 0, "usage_input_tokens": 121, "questions": [ { "qid": "topic", "type": "choice", "K": 4, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 49, "positions": 49, "bucket": 256, "ids": [ 1, 17, 18, 3493, 856, 779, 21009, 10789, 1337, 540, 19, 16, 11746, 535, 835, 11746, 933, 768, 15073, 8688, 20, 19, 16, 5363, 800, 535, 61958, 774, 810, 23325, 20, 19, 16, 3863, 535, 21040, 3197, 20, 19, 16, 12779, 535, 36843, 3321, 1337, 779, 3714, 20, 21 ], "markers": [ 11, 22, 32, 39 ], "logits_raw": [ -4.030765533447266, -2.516629695892334, -4.948546409606934, -1.2068599462509155 ], "probs_raw": [ 0.04388390854001045, 0.19947391748428345, 0.017527427524328232, 0.7391147613525391 ], "probs": [ 0.08343102037906647, 0.24607208371162415, 0.043311357498168945, 0.6271855235099792 ], "argmax_index": 3, "argmax": "feedback", "top2_margin": 0.3811134397983551, "near_tie": false, "gold": null, "correct": null, "host_softmax_max_abs_dp": 1.4901161193847656e-08, "batch": { "call": 0, "index": 0, "B": 3 } }, { "qid": "wants", "type": "noul", "K": 2, "calibrate": true, "temperature_key": "noul:2", "T": 1.6663223505020142, "prefix": 0, "n_ids": 36, "positions": 36, "bucket": 256, "ids": [ 1, 17, 18, 4940, 779, 21009, 14065, 875, 3503, 811, 874, 4007, 540, 19, 16, 4831, 535, 1295, 521, 779, 8294, 2177, 1014, 3965, 20, 19, 16, 3347, 535, 18672, 521, 779, 8294, 10995, 20, 21 ], "markers": [ 14, 26 ], "logits_raw": [ -0.19816279411315918, 3.9951019287109375 ], "probs_raw": [ 0.014872386120259762, 0.9851276278495789 ], "probs": [ 0.9252889752388, 0.07471100240945816 ], "argmax_index": 0, "argmax": "true", "top2_margin": 0.8505779728293419, "near_tie": false, "gold": null, "correct": null, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 1, "B": 3 } }, { "qid": "urgency", "type": "score", "K": 3, "calibrate": true, "temperature_key": "score:3-5", "T": 1.7301132678985596, "prefix": 0, "n_ids": 36, "positions": 36, "bucket": 256, "ids": [ 1, 17, 18, 4023, 31149, 856, 779, 5940, 540, 19, 16, 2119, 730, 525, 535, 2577, 7498, 20, 19, 16, 2119, 730, 526, 535, 32238, 20, 19, 16, 2119, 730, 527, 535, 17231, 2166, 20, 21 ], "markers": [ 10, 19, 27 ], "logits_raw": [ -7.304863452911377, -6.445466995239258, -5.352657318115234 ], "probs_raw": [ 0.09609891474246979, 0.22696013748645782, 0.6769409775733948 ], "probs": [ 0.17439991235733032, 0.28659766912460327, 0.5390024185180664 ], "argmax_index": 2, "argmax": "2", "top2_margin": 0.25240474939346313, "near_tie": false, "gold": null, "correct": null, "host_softmax_max_abs_dp": 2.9802322387695312e-08, "batch": { "call": 0, "index": 2, "B": 3 } } ], "response": { "answers": { "topic": { "type": "choice", "choice": "feedback", "confidence": 0.6271855235099792, "probabilities": { "delivery": 0.08343102037906647, "billing": 0.24607208371162415, "account": 0.043311357498168945, "feedback": 0.6271855235099792 } }, "wants": { "type": "noul", "noul": 0.9252889752388 }, "urgency": { "type": "score", "score": 1.364602506160736, "confidence": 0.5390024185180664, "probabilities": { "0": 0.17439991235733032, "1": 0.28659766912460327, "2": 0.5390024185180664 }, "legend": { "0": "Can wait", "1": "Soon", "2": "Right now" } } }, "usage": { "input_tokens": 121, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "aud_04", "mode": "audio", "source": "own_audio", "native": true, "prefix": 35, "usage_input_tokens": 244, "questions": [ { "qid": "topic", "type": "choice", "K": 4, "calibrate": false, "temperature_key": null, "T": null, "prefix": 35, "n_ids": 63, "positions": 98, "bucket": 256, "ids": [ 1, 17, 27377, 1139, 6634, 988, 768, 5196, 523, 18, 3493, 856, 779, 21009, 10789, 1337, 540, 19, 16, 7832, 572, 1873, 535, 835, 11746, 933, 768, 15073, 8688, 20, 19, 16, 7832, 572, 7778, 535, 61958, 774, 810, 23325, 20, 19, 16, 7832, 572, 12992, 535, 21040, 3197, 20, 19, 16, 7832, 572, 15296, 535, 36843, 3321, 1337, 779, 3714, 20, 21 ], "markers": [ 18, 31, 42, 51 ], "logits_raw": [ -11.180398941040039, -14.908483505249023, -14.098936080932617, -0.7033980488777161 ], "probs_raw": [ 2.8176236810395494e-05, 6.773240102120326e-07, 1.5218730595734087e-06, 0.9999696016311646 ], "probs": [ 2.8176236810395494e-05, 6.773240102120326e-07, 1.5218730595734087e-06, 0.9999696016311646 ], "argmax_index": 3, "argmax": "feedback", "top2_margin": 0.9999414253943542, "near_tie": false, "gold": "account", "correct": false, "host_softmax_max_abs_dp": 5.684341886080802e-14, "batch": { "call": 0, "index": 0, "B": 3 } }, { "qid": "wants", "type": "noul", "K": 2, "calibrate": false, "temperature_key": null, "T": null, "prefix": 35, "n_ids": 33, "positions": 68, "bucket": 256, "ids": [ 1, 17, 27377, 1139, 6634, 988, 768, 5196, 523, 18, 4940, 779, 21009, 14065, 875, 3503, 811, 874, 4007, 540, 19, 16, 4831, 535, 1295, 20, 19, 16, 3347, 535, 18672, 20, 21 ], "markers": [ 21, 27 ], "logits_raw": [ -0.5335029363632202, -3.4489498138427734 ], "probs_raw": [ 0.9486048221588135, 0.0513952411711216 ], "probs": [ 0.0513952411711216, 0.9486048221588135 ], "argmax_index": 1, "argmax": "false", "top2_margin": 0.8972095809876919, "near_tie": false, "gold": "true", "correct": false, "host_softmax_max_abs_dp": 3.725290298461914e-09, "batch": { "call": 0, "index": 2, "B": 3 } }, { "qid": "urgency", "type": "score", "K": 3, "calibrate": false, "temperature_key": null, "T": null, "prefix": 35, "n_ids": 43, "positions": 78, "bucket": 256, "ids": [ 1, 17, 27377, 1139, 6634, 988, 768, 5196, 523, 18, 4023, 31149, 856, 779, 5940, 540, 19, 16, 2119, 730, 525, 535, 2577, 7498, 20, 19, 16, 2119, 730, 526, 535, 32238, 20, 19, 16, 2119, 730, 527, 535, 17231, 2166, 20, 21 ], "markers": [ 17, 26, 34 ], "logits_raw": [ -13.666671752929688, -7.450804710388184, 2.360034465789795 ], "probs_raw": [ 1.0956349427715395e-07, 5.4850806918693706e-05, 0.9999450445175171 ], "probs": [ 1.0956349427715395e-07, 5.4850806918693706e-05, 0.9999450445175171 ], "argmax_index": 2, "argmax": "2", "top2_margin": 0.9998901937105984, "near_tie": false, "gold": "2", "correct": true, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 1, "B": 3 } } ], "response": { "answers": { "topic": { "type": "choice", "choice": "feedback", "confidence": 0.9999696016311646, "probabilities": { "delivery": 2.8176236810395494e-05, "billing": 6.773240102120326e-07, "account": 1.5218730595734087e-06, "feedback": 0.9999696016311646 } }, "wants": { "type": "noul", "noul": 0.0513952411711216 }, "urgency": { "type": "score", "score": 1.9999449398419529, "confidence": 0.9999450445175171, "probabilities": { "0": 1.0956349427715395e-07, "1": 5.4850806918693706e-05, "2": 0.9999450445175171 }, "legend": { "0": "Can wait", "1": "Soon", "2": "Right now" } } }, "usage": { "input_tokens": 244, "output_tokens": 0 } }, "oracle_file": "records_ref_audio.json" }, { "id": "aud_04", "mode": "text", "source": "own_audio", "native": false, "prefix": 0, "usage_input_tokens": 142, "questions": [ { "qid": "topic", "type": "choice", "K": 4, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 56, "positions": 56, "bucket": 256, "ids": [ 1, 17, 27377, 1139, 6634, 988, 768, 5196, 523, 18, 3493, 856, 779, 21009, 10789, 1337, 540, 19, 16, 11746, 535, 835, 11746, 933, 768, 15073, 8688, 20, 19, 16, 5363, 800, 535, 61958, 774, 810, 23325, 20, 19, 16, 3863, 535, 21040, 3197, 20, 19, 16, 12779, 535, 36843, 3321, 1337, 779, 3714, 20, 21 ], "markers": [ 18, 29, 39, 46 ], "logits_raw": [ -2.1100118160247803, -1.3073279857635498, -3.290978193283081, 1.7141824960708618 ], "probs_raw": [ 0.020269835367798805, 0.04523257911205292, 0.0062224725261330605, 0.9282751679420471 ], "probs": [ 0.05386649817228317, 0.0955735445022583, 0.023170966655015945, 0.8273889422416687 ], "argmax_index": 3, "argmax": "feedback", "top2_margin": 0.7318153977394104, "near_tie": false, "gold": null, "correct": null, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 0, "B": 3 } }, { "qid": "wants", "type": "noul", "K": 2, "calibrate": true, "temperature_key": "noul:2", "T": 1.6663223505020142, "prefix": 0, "n_ids": 43, "positions": 43, "bucket": 256, "ids": [ 1, 17, 27377, 1139, 6634, 988, 768, 5196, 523, 18, 4940, 779, 21009, 14065, 875, 3503, 811, 874, 4007, 540, 19, 16, 4831, 535, 1295, 521, 779, 8294, 2177, 1014, 3965, 20, 19, 16, 3347, 535, 18672, 521, 779, 8294, 10995, 20, 21 ], "markers": [ 21, 33 ], "logits_raw": [ 2.693018674850464, -4.27215051651001 ], "probs_raw": [ 0.9990566372871399, 0.0009433126542717218 ], "probs": [ 0.015068508684635162, 0.984931468963623 ], "argmax_index": 1, "argmax": "false", "top2_margin": 0.9698629602789879, "near_tie": false, "gold": null, "correct": null, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 1, "B": 3 } }, { "qid": "urgency", "type": "score", "K": 3, "calibrate": true, "temperature_key": "score:3-5", "T": 1.7301132678985596, "prefix": 0, "n_ids": 43, "positions": 43, "bucket": 256, "ids": [ 1, 17, 27377, 1139, 6634, 988, 768, 5196, 523, 18, 4023, 31149, 856, 779, 5940, 540, 19, 16, 2119, 730, 525, 535, 2577, 7498, 20, 19, 16, 2119, 730, 526, 535, 32238, 20, 19, 16, 2119, 730, 527, 535, 17231, 2166, 20, 21 ], "markers": [ 17, 26, 34 ], "logits_raw": [ -8.044438362121582, -6.318297863006592, -8.428503036499023 ], "probs_raw": [ 0.13698606193065643, 0.7697144746780396, 0.09329947084188461 ], "probs": [ 0.22158345580101013, 0.6009449362754822, 0.17747168242931366 ], "argmax_index": 1, "argmax": "1", "top2_margin": 0.37936148047447205, "near_tie": false, "gold": null, "correct": null, "host_softmax_max_abs_dp": 1.4901161193847656e-08, "batch": { "call": 0, "index": 2, "B": 3 } } ], "response": { "answers": { "topic": { "type": "choice", "choice": "feedback", "confidence": 0.8273889422416687, "probabilities": { "delivery": 0.05386649817228317, "billing": 0.0955735445022583, "account": 0.023170966655015945, "feedback": 0.8273889422416687 } }, "wants": { "type": "noul", "noul": 0.015068508684635162 }, "urgency": { "type": "score", "score": 0.9558883011341095, "confidence": 0.6009449362754822, "probabilities": { "0": 0.22158345580101013, "1": 0.6009449362754822, "2": 0.17747168242931366 }, "legend": { "0": "Can wait", "1": "Soon", "2": "Right now" } } }, "usage": { "input_tokens": 142, "output_tokens": 0 } }, "oracle_file": "records_ref_audio.json" }, { "id": "aud_05", "mode": "audio", "source": "own_audio", "native": true, "prefix": 36, "usage_input_tokens": 229, "questions": [ { "qid": "topic", "type": "choice", "K": 4, "calibrate": false, "temperature_key": null, "T": null, "prefix": 36, "n_ids": 57, "positions": 93, "bucket": 256, "ids": [ 1, 17, 24373, 18, 3493, 856, 779, 21009, 10789, 1337, 540, 19, 16, 7832, 572, 1873, 535, 835, 11746, 933, 768, 15073, 8688, 20, 19, 16, 7832, 572, 7778, 535, 61958, 774, 810, 23325, 20, 19, 16, 7832, 572, 12992, 535, 21040, 3197, 20, 19, 16, 7832, 572, 15296, 535, 36843, 3321, 1337, 779, 3714, 20, 21 ], "markers": [ 12, 25, 36, 45 ], "logits_raw": [ -11.728607177734375, -14.39934253692627, -4.546878814697266, -2.9778451919555664 ], "probs_raw": [ 0.00013103160017635673, 9.067559403774794e-06, 0.17233005166053772, 0.8275298476219177 ], "probs": [ 0.00013103160017635673, 9.067559403774794e-06, 0.17233005166053772, 0.8275298476219177 ], "argmax_index": 3, "argmax": "feedback", "top2_margin": 0.65519979596138, "near_tie": false, "gold": "billing", "correct": false, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 0, "B": 3 } }, { "qid": "wants", "type": "noul", "K": 2, "calibrate": false, "temperature_key": null, "T": null, "prefix": 36, "n_ids": 27, "positions": 63, "bucket": 256, "ids": [ 1, 17, 24373, 18, 4940, 779, 21009, 14065, 875, 3503, 811, 874, 4007, 540, 19, 16, 4831, 535, 1295, 20, 19, 16, 3347, 535, 18672, 20, 21 ], "markers": [ 15, 21 ], "logits_raw": [ -0.6051132678985596, -0.5525320768356323 ], "probs_raw": [ 0.48685774207115173, 0.5131422877311707 ], "probs": [ 0.5131422877311707, 0.48685774207115173 ], "argmax_index": 0, "argmax": "true", "top2_margin": 0.02628454566001892, "near_tie": false, "gold": "true", "correct": true, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 2, "B": 3 } }, { "qid": "urgency", "type": "score", "K": 3, "calibrate": false, "temperature_key": null, "T": null, "prefix": 36, "n_ids": 37, "positions": 73, "bucket": 256, "ids": [ 1, 17, 24373, 18, 4023, 31149, 856, 779, 5940, 540, 19, 16, 2119, 730, 525, 535, 2577, 7498, 20, 19, 16, 2119, 730, 526, 535, 32238, 20, 19, 16, 2119, 730, 527, 535, 17231, 2166, 20, 21 ], "markers": [ 11, 20, 28 ], "logits_raw": [ -13.145566940307617, -6.944738388061523, -7.77335262298584 ], "probs_raw": [ 0.0014094498474150896, 0.6950807571411133, 0.30350977182388306 ], "probs": [ 0.0014094498474150896, 0.6950807571411133, 0.30350977182388306 ], "argmax_index": 1, "argmax": "1", "top2_margin": 0.3915709853172302, "near_tie": false, "gold": "0", "correct": false, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 1, "B": 3 } } ], "response": { "answers": { "topic": { "type": "choice", "choice": "feedback", "confidence": 0.8275298476219177, "probabilities": { "delivery": 0.00013103160017635673, "billing": 9.067559403774794e-06, "account": 0.17233005166053772, "feedback": 0.8275298476219177 } }, "wants": { "type": "noul", "noul": 0.5131422877311707 }, "urgency": { "type": "score", "score": 1.3021003007888794, "confidence": 0.6950807571411133, "probabilities": { "0": 0.0014094498474150896, "1": 0.6950807571411133, "2": 0.30350977182388306 }, "legend": { "0": "Can wait", "1": "Soon", "2": "Right now" } } }, "usage": { "input_tokens": 229, "output_tokens": 0 } }, "oracle_file": "records_ref_audio.json" }, { "id": "aud_05", "mode": "text", "source": "own_audio", "native": false, "prefix": 0, "usage_input_tokens": 121, "questions": [ { "qid": "topic", "type": "choice", "K": 4, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 49, "positions": 49, "bucket": 256, "ids": [ 1, 17, 18, 3493, 856, 779, 21009, 10789, 1337, 540, 19, 16, 11746, 535, 835, 11746, 933, 768, 15073, 8688, 20, 19, 16, 5363, 800, 535, 61958, 774, 810, 23325, 20, 19, 16, 3863, 535, 21040, 3197, 20, 19, 16, 12779, 535, 36843, 3321, 1337, 779, 3714, 20, 21 ], "markers": [ 11, 22, 32, 39 ], "logits_raw": [ -4.030765533447266, -2.516629695892334, -4.948546409606934, -1.2068599462509155 ], "probs_raw": [ 0.04388390854001045, 0.19947391748428345, 0.017527427524328232, 0.7391147613525391 ], "probs": [ 0.08343102037906647, 0.24607208371162415, 0.043311357498168945, 0.6271855235099792 ], "argmax_index": 3, "argmax": "feedback", "top2_margin": 0.3811134397983551, "near_tie": false, "gold": null, "correct": null, "host_softmax_max_abs_dp": 1.4901161193847656e-08, "batch": { "call": 0, "index": 0, "B": 3 } }, { "qid": "wants", "type": "noul", "K": 2, "calibrate": true, "temperature_key": "noul:2", "T": 1.6663223505020142, "prefix": 0, "n_ids": 36, "positions": 36, "bucket": 256, "ids": [ 1, 17, 18, 4940, 779, 21009, 14065, 875, 3503, 811, 874, 4007, 540, 19, 16, 4831, 535, 1295, 521, 779, 8294, 2177, 1014, 3965, 20, 19, 16, 3347, 535, 18672, 521, 779, 8294, 10995, 20, 21 ], "markers": [ 14, 26 ], "logits_raw": [ -0.19816279411315918, 3.9951019287109375 ], "probs_raw": [ 0.014872386120259762, 0.9851276278495789 ], "probs": [ 0.9252889752388, 0.07471100240945816 ], "argmax_index": 0, "argmax": "true", "top2_margin": 0.8505779728293419, "near_tie": false, "gold": null, "correct": null, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 1, "B": 3 } }, { "qid": "urgency", "type": "score", "K": 3, "calibrate": true, "temperature_key": "score:3-5", "T": 1.7301132678985596, "prefix": 0, "n_ids": 36, "positions": 36, "bucket": 256, "ids": [ 1, 17, 18, 4023, 31149, 856, 779, 5940, 540, 19, 16, 2119, 730, 525, 535, 2577, 7498, 20, 19, 16, 2119, 730, 526, 535, 32238, 20, 19, 16, 2119, 730, 527, 535, 17231, 2166, 20, 21 ], "markers": [ 10, 19, 27 ], "logits_raw": [ -7.304863452911377, -6.445466995239258, -5.352657318115234 ], "probs_raw": [ 0.09609891474246979, 0.22696013748645782, 0.6769409775733948 ], "probs": [ 0.17439991235733032, 0.28659766912460327, 0.5390024185180664 ], "argmax_index": 2, "argmax": "2", "top2_margin": 0.25240474939346313, "near_tie": false, "gold": null, "correct": null, "host_softmax_max_abs_dp": 2.9802322387695312e-08, "batch": { "call": 0, "index": 2, "B": 3 } } ], "response": { "answers": { "topic": { "type": "choice", "choice": "feedback", "confidence": 0.6271855235099792, "probabilities": { "delivery": 0.08343102037906647, "billing": 0.24607208371162415, "account": 0.043311357498168945, "feedback": 0.6271855235099792 } }, "wants": { "type": "noul", "noul": 0.9252889752388 }, "urgency": { "type": "score", "score": 1.364602506160736, "confidence": 0.5390024185180664, "probabilities": { "0": 0.17439991235733032, "1": 0.28659766912460327, "2": 0.5390024185180664 }, "legend": { "0": "Can wait", "1": "Soon", "2": "Right now" } } }, "usage": { "input_tokens": 121, "output_tokens": 0 } }, "oracle_file": "records_ref_audio.json" }, { "id": "aud_06", "mode": "audio", "source": "own_audio", "native": true, "prefix": 101, "usage_input_tokens": 442, "questions": [ { "qid": "topic", "type": "choice", "K": 4, "calibrate": false, "temperature_key": null, "T": null, "prefix": 101, "n_ids": 63, "positions": 164, "bucket": 256, "ids": [ 1, 17, 27377, 1139, 6634, 988, 768, 5196, 523, 18, 3493, 856, 779, 21009, 10789, 1337, 540, 19, 16, 7832, 572, 1873, 535, 835, 11746, 933, 768, 15073, 8688, 20, 19, 16, 7832, 572, 7778, 535, 61958, 774, 810, 23325, 20, 19, 16, 7832, 572, 12992, 535, 21040, 3197, 20, 19, 16, 7832, 572, 15296, 535, 36843, 3321, 1337, 779, 3714, 20, 21 ], "markers": [ 18, 31, 42, 51 ], "logits_raw": [ -5.219973564147949, -7.757868766784668, -11.41282844543457, 0.0007779449224472046 ], "probs_raw": [ 0.005371889099478722, 0.00042455419315956533, 1.0980044862662908e-05, 0.9941925406455994 ], "probs": [ 0.005371889099478722, 0.00042455419315956533, 1.0980044862662908e-05, 0.9941925406455994 ], "argmax_index": 3, "argmax": "feedback", "top2_margin": 0.9888206515461206, "near_tie": false, "gold": "delivery", "correct": false, "host_softmax_max_abs_dp": 2.9103830456733704e-11, "batch": { "call": 0, "index": 0, "B": 3 } }, { "qid": "wants", "type": "noul", "K": 2, "calibrate": false, "temperature_key": null, "T": null, "prefix": 101, "n_ids": 33, "positions": 134, "bucket": 256, "ids": [ 1, 17, 27377, 1139, 6634, 988, 768, 5196, 523, 18, 4940, 779, 21009, 14065, 875, 3503, 811, 874, 4007, 540, 19, 16, 4831, 535, 1295, 20, 19, 16, 3347, 535, 18672, 20, 21 ], "markers": [ 21, 27 ], "logits_raw": [ -0.22706714272499084, -3.0375657081604004 ], "probs_raw": [ 0.9432405829429626, 0.05675949156284332 ], "probs": [ 0.05675949156284332, 0.9432405829429626 ], "argmax_index": 1, "argmax": "false", "top2_margin": 0.8864810913801193, "near_tie": false, "gold": "true", "correct": false, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 2, "B": 3 } }, { "qid": "urgency", "type": "score", "K": 3, "calibrate": false, "temperature_key": null, "T": null, "prefix": 101, "n_ids": 43, "positions": 144, "bucket": 256, "ids": [ 1, 17, 27377, 1139, 6634, 988, 768, 5196, 523, 18, 4023, 31149, 856, 779, 5940, 540, 19, 16, 2119, 730, 525, 535, 2577, 7498, 20, 19, 16, 2119, 730, 526, 535, 32238, 20, 19, 16, 2119, 730, 527, 535, 17231, 2166, 20, 21 ], "markers": [ 17, 26, 34 ], "logits_raw": [ -13.540458679199219, -7.2855424880981445, -5.484748840332031 ], "probs_raw": [ 0.0002722343779169023, 0.1417158991098404, 0.8580119013786316 ], "probs": [ 0.0002722343779169023, 0.1417158991098404, 0.8580119013786316 ], "argmax_index": 2, "argmax": "2", "top2_margin": 0.7162960022687912, "near_tie": false, "gold": "1", "correct": false, "host_softmax_max_abs_dp": 1.4901161193847656e-08, "batch": { "call": 0, "index": 1, "B": 3 } } ], "response": { "answers": { "topic": { "type": "choice", "choice": "feedback", "confidence": 0.9941925406455994, "probabilities": { "delivery": 0.005371889099478722, "billing": 0.00042455419315956533, "account": 1.0980044862662908e-05, "feedback": 0.9941925406455994 } }, "wants": { "type": "noul", "noul": 0.05675949156284332 }, "urgency": { "type": "score", "score": 1.8577397018671036, "confidence": 0.8580119013786316, "probabilities": { "0": 0.0002722343779169023, "1": 0.1417158991098404, "2": 0.8580119013786316 }, "legend": { "0": "Can wait", "1": "Soon", "2": "Right now" } } }, "usage": { "input_tokens": 442, "output_tokens": 0 } }, "oracle_file": "records_ref_audio.json" }, { "id": "aud_06", "mode": "text", "source": "own_audio", "native": false, "prefix": 0, "usage_input_tokens": 142, "questions": [ { "qid": "topic", "type": "choice", "K": 4, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 56, "positions": 56, "bucket": 256, "ids": [ 1, 17, 27377, 1139, 6634, 988, 768, 5196, 523, 18, 3493, 856, 779, 21009, 10789, 1337, 540, 19, 16, 11746, 535, 835, 11746, 933, 768, 15073, 8688, 20, 19, 16, 5363, 800, 535, 61958, 774, 810, 23325, 20, 19, 16, 3863, 535, 21040, 3197, 20, 19, 16, 12779, 535, 36843, 3321, 1337, 779, 3714, 20, 21 ], "markers": [ 18, 29, 39, 46 ], "logits_raw": [ -2.1100118160247803, -1.3073279857635498, -3.290978193283081, 1.7141824960708618 ], "probs_raw": [ 0.020269835367798805, 0.04523257911205292, 0.0062224725261330605, 0.9282751679420471 ], "probs": [ 0.05386649817228317, 0.0955735445022583, 0.023170966655015945, 0.8273889422416687 ], "argmax_index": 3, "argmax": "feedback", "top2_margin": 0.7318153977394104, "near_tie": false, "gold": null, "correct": null, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 0, "B": 3 } }, { "qid": "wants", "type": "noul", "K": 2, "calibrate": true, "temperature_key": "noul:2", "T": 1.6663223505020142, "prefix": 0, "n_ids": 43, "positions": 43, "bucket": 256, "ids": [ 1, 17, 27377, 1139, 6634, 988, 768, 5196, 523, 18, 4940, 779, 21009, 14065, 875, 3503, 811, 874, 4007, 540, 19, 16, 4831, 535, 1295, 521, 779, 8294, 2177, 1014, 3965, 20, 19, 16, 3347, 535, 18672, 521, 779, 8294, 10995, 20, 21 ], "markers": [ 21, 33 ], "logits_raw": [ 2.693018674850464, -4.27215051651001 ], "probs_raw": [ 0.9990566372871399, 0.0009433126542717218 ], "probs": [ 0.015068508684635162, 0.984931468963623 ], "argmax_index": 1, "argmax": "false", "top2_margin": 0.9698629602789879, "near_tie": false, "gold": null, "correct": null, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 1, "B": 3 } }, { "qid": "urgency", "type": "score", "K": 3, "calibrate": true, "temperature_key": "score:3-5", "T": 1.7301132678985596, "prefix": 0, "n_ids": 43, "positions": 43, "bucket": 256, "ids": [ 1, 17, 27377, 1139, 6634, 988, 768, 5196, 523, 18, 4023, 31149, 856, 779, 5940, 540, 19, 16, 2119, 730, 525, 535, 2577, 7498, 20, 19, 16, 2119, 730, 526, 535, 32238, 20, 19, 16, 2119, 730, 527, 535, 17231, 2166, 20, 21 ], "markers": [ 17, 26, 34 ], "logits_raw": [ -8.044438362121582, -6.318297863006592, -8.428503036499023 ], "probs_raw": [ 0.13698606193065643, 0.7697144746780396, 0.09329947084188461 ], "probs": [ 0.22158345580101013, 0.6009449362754822, 0.17747168242931366 ], "argmax_index": 1, "argmax": "1", "top2_margin": 0.37936148047447205, "near_tie": false, "gold": null, "correct": null, "host_softmax_max_abs_dp": 1.4901161193847656e-08, "batch": { "call": 0, "index": 2, "B": 3 } } ], "response": { "answers": { "topic": { "type": "choice", "choice": "feedback", "confidence": 0.8273889422416687, "probabilities": { "delivery": 0.05386649817228317, "billing": 0.0955735445022583, "account": 0.023170966655015945, "feedback": 0.8273889422416687 } }, "wants": { "type": "noul", "noul": 0.015068508684635162 }, "urgency": { "type": "score", "score": 0.9558883011341095, "confidence": 0.6009449362754822, "probabilities": { "0": 0.22158345580101013, "1": 0.6009449362754822, "2": 0.17747168242931366 }, "legend": { "0": "Can wait", "1": "Soon", "2": "Right now" } } }, "usage": { "input_tokens": 142, "output_tokens": 0 } }, "oracle_file": "records_ref_audio.json" }, { "id": "aud_07", "mode": "audio", "source": "own_audio", "native": true, "prefix": 106, "usage_input_tokens": 439, "questions": [ { "qid": "topic", "type": "choice", "K": 4, "calibrate": false, "temperature_key": null, "T": null, "prefix": 106, "n_ids": 57, "positions": 163, "bucket": 256, "ids": [ 1, 17, 24373, 18, 3493, 856, 779, 21009, 10789, 1337, 540, 19, 16, 7832, 572, 1873, 535, 835, 11746, 933, 768, 15073, 8688, 20, 19, 16, 7832, 572, 7778, 535, 61958, 774, 810, 23325, 20, 19, 16, 7832, 572, 12992, 535, 21040, 3197, 20, 19, 16, 7832, 572, 15296, 535, 36843, 3321, 1337, 779, 3714, 20, 21 ], "markers": [ 12, 25, 36, 45 ], "logits_raw": [ -13.036137580871582, -11.367012023925781, -12.230705261230469, -2.8472213745117188 ], "probs_raw": [ 3.757253216463141e-05, 0.00019941713253501803, 8.407468703808263e-05, 0.9996789693832397 ], "probs": [ 3.757253216463141e-05, 0.00019941713253501803, 8.407468703808263e-05, 0.9996789693832397 ], "argmax_index": 3, "argmax": "feedback", "top2_margin": 0.9994795522507047, "near_tie": false, "gold": "feedback", "correct": true, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 0, "B": 3 } }, { "qid": "wants", "type": "noul", "K": 2, "calibrate": false, "temperature_key": null, "T": null, "prefix": 106, "n_ids": 27, "positions": 133, "bucket": 256, "ids": [ 1, 17, 24373, 18, 4940, 779, 21009, 14065, 875, 3503, 811, 874, 4007, 540, 19, 16, 4831, 535, 1295, 20, 19, 16, 3347, 535, 18672, 20, 21 ], "markers": [ 15, 21 ], "logits_raw": [ 0.03376114368438721, -3.4692351818084717 ], "probs_raw": [ 0.970772922039032, 0.02922709286212921 ], "probs": [ 0.02922709286212921, 0.970772922039032 ], "argmax_index": 1, "argmax": "false", "top2_margin": 0.9415458291769028, "near_tie": false, "gold": "false", "correct": true, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 2, "B": 3 } }, { "qid": "urgency", "type": "score", "K": 3, "calibrate": false, "temperature_key": null, "T": null, "prefix": 106, "n_ids": 37, "positions": 143, "bucket": 256, "ids": [ 1, 17, 24373, 18, 4023, 31149, 856, 779, 5940, 540, 19, 16, 2119, 730, 525, 535, 2577, 7498, 20, 19, 16, 2119, 730, 526, 535, 32238, 20, 19, 16, 2119, 730, 527, 535, 17231, 2166, 20, 21 ], "markers": [ 11, 20, 28 ], "logits_raw": [ -9.906222343444824, -5.438223838806152, 1.249060869216919 ], "probs_raw": [ 1.4281534276960883e-05, 0.0012450935319066048, 0.9987406134605408 ], "probs": [ 1.4281534276960883e-05, 0.0012450935319066048, 0.9987406134605408 ], "argmax_index": 2, "argmax": "2", "top2_margin": 0.9974955199286342, "near_tie": false, "gold": "0", "correct": false, "host_softmax_max_abs_dp": 1.1641532182693481e-10, "batch": { "call": 0, "index": 1, "B": 3 } } ], "response": { "answers": { "topic": { "type": "choice", "choice": "feedback", "confidence": 0.9996789693832397, "probabilities": { "delivery": 3.757253216463141e-05, "billing": 0.00019941713253501803, "account": 8.407468703808263e-05, "feedback": 0.9996789693832397 } }, "wants": { "type": "noul", "noul": 0.02922709286212921 }, "urgency": { "type": "score", "score": 1.9987263204529881, "confidence": 0.9987406134605408, "probabilities": { "0": 1.4281534276960883e-05, "1": 0.0012450935319066048, "2": 0.9987406134605408 }, "legend": { "0": "Can wait", "1": "Soon", "2": "Right now" } } }, "usage": { "input_tokens": 439, "output_tokens": 0 } }, "oracle_file": "records_ref_audio.json" }, { "id": "aud_07", "mode": "text", "source": "own_audio", "native": false, "prefix": 0, "usage_input_tokens": 121, "questions": [ { "qid": "topic", "type": "choice", "K": 4, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 49, "positions": 49, "bucket": 256, "ids": [ 1, 17, 18, 3493, 856, 779, 21009, 10789, 1337, 540, 19, 16, 11746, 535, 835, 11746, 933, 768, 15073, 8688, 20, 19, 16, 5363, 800, 535, 61958, 774, 810, 23325, 20, 19, 16, 3863, 535, 21040, 3197, 20, 19, 16, 12779, 535, 36843, 3321, 1337, 779, 3714, 20, 21 ], "markers": [ 11, 22, 32, 39 ], "logits_raw": [ -4.030765533447266, -2.516629695892334, -4.948546409606934, -1.2068599462509155 ], "probs_raw": [ 0.04388390854001045, 0.19947391748428345, 0.017527427524328232, 0.7391147613525391 ], "probs": [ 0.08343102037906647, 0.24607208371162415, 0.043311357498168945, 0.6271855235099792 ], "argmax_index": 3, "argmax": "feedback", "top2_margin": 0.3811134397983551, "near_tie": false, "gold": null, "correct": null, "host_softmax_max_abs_dp": 1.4901161193847656e-08, "batch": { "call": 0, "index": 0, "B": 3 } }, { "qid": "wants", "type": "noul", "K": 2, "calibrate": true, "temperature_key": "noul:2", "T": 1.6663223505020142, "prefix": 0, "n_ids": 36, "positions": 36, "bucket": 256, "ids": [ 1, 17, 18, 4940, 779, 21009, 14065, 875, 3503, 811, 874, 4007, 540, 19, 16, 4831, 535, 1295, 521, 779, 8294, 2177, 1014, 3965, 20, 19, 16, 3347, 535, 18672, 521, 779, 8294, 10995, 20, 21 ], "markers": [ 14, 26 ], "logits_raw": [ -0.19816279411315918, 3.9951019287109375 ], "probs_raw": [ 0.014872386120259762, 0.9851276278495789 ], "probs": [ 0.9252889752388, 0.07471100240945816 ], "argmax_index": 0, "argmax": "true", "top2_margin": 0.8505779728293419, "near_tie": false, "gold": null, "correct": null, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 1, "B": 3 } }, { "qid": "urgency", "type": "score", "K": 3, "calibrate": true, "temperature_key": "score:3-5", "T": 1.7301132678985596, "prefix": 0, "n_ids": 36, "positions": 36, "bucket": 256, "ids": [ 1, 17, 18, 4023, 31149, 856, 779, 5940, 540, 19, 16, 2119, 730, 525, 535, 2577, 7498, 20, 19, 16, 2119, 730, 526, 535, 32238, 20, 19, 16, 2119, 730, 527, 535, 17231, 2166, 20, 21 ], "markers": [ 10, 19, 27 ], "logits_raw": [ -7.304863452911377, -6.445466995239258, -5.352657318115234 ], "probs_raw": [ 0.09609891474246979, 0.22696013748645782, 0.6769409775733948 ], "probs": [ 0.17439991235733032, 0.28659766912460327, 0.5390024185180664 ], "argmax_index": 2, "argmax": "2", "top2_margin": 0.25240474939346313, "near_tie": false, "gold": null, "correct": null, "host_softmax_max_abs_dp": 2.9802322387695312e-08, "batch": { "call": 0, "index": 2, "B": 3 } } ], "response": { "answers": { "topic": { "type": "choice", "choice": "feedback", "confidence": 0.6271855235099792, "probabilities": { "delivery": 0.08343102037906647, "billing": 0.24607208371162415, "account": 0.043311357498168945, "feedback": 0.6271855235099792 } }, "wants": { "type": "noul", "noul": 0.9252889752388 }, "urgency": { "type": "score", "score": 1.364602506160736, "confidence": 0.5390024185180664, "probabilities": { "0": 0.17439991235733032, "1": 0.28659766912460327, "2": 0.5390024185180664 }, "legend": { "0": "Can wait", "1": "Soon", "2": "Right now" } } }, "usage": { "input_tokens": 121, "output_tokens": 0 } }, "oracle_file": "records_ref_audio.json" }, { "id": "aud_08", "mode": "audio", "source": "own_audio", "native": true, "prefix": 199, "usage_input_tokens": 736, "questions": [ { "qid": "topic", "type": "choice", "K": 4, "calibrate": false, "temperature_key": null, "T": null, "prefix": 199, "n_ids": 63, "positions": 262, "bucket": 512, "ids": [ 1, 17, 27377, 1139, 6634, 988, 768, 5196, 523, 18, 3493, 856, 779, 21009, 10789, 1337, 540, 19, 16, 7832, 572, 1873, 535, 835, 11746, 933, 768, 15073, 8688, 20, 19, 16, 7832, 572, 7778, 535, 61958, 774, 810, 23325, 20, 19, 16, 7832, 572, 12992, 535, 21040, 3197, 20, 19, 16, 7832, 572, 15296, 535, 36843, 3321, 1337, 779, 3714, 20, 21 ], "markers": [ 18, 31, 42, 51 ], "logits_raw": [ -8.704425811767578, -3.4039201736450195, -11.297805786132812, -0.0038675889372825623 ], "probs_raw": [ 0.00016108817362692207, 0.03228823095560074, 1.204406453325646e-05, 0.967538595199585 ], "probs": [ 0.00016108817362692207, 0.03228823095560074, 1.204406453325646e-05, 0.967538595199585 ], "argmax_index": 3, "argmax": "feedback", "top2_margin": 0.9352503642439842, "near_tie": false, "gold": "billing", "correct": false, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 0, "B": 3 } }, { "qid": "wants", "type": "noul", "K": 2, "calibrate": false, "temperature_key": null, "T": null, "prefix": 199, "n_ids": 33, "positions": 232, "bucket": 256, "ids": [ 1, 17, 27377, 1139, 6634, 988, 768, 5196, 523, 18, 4940, 779, 21009, 14065, 875, 3503, 811, 874, 4007, 540, 19, 16, 4831, 535, 1295, 20, 19, 16, 3347, 535, 18672, 20, 21 ], "markers": [ 21, 27 ], "logits_raw": [ -1.890669584274292, 1.8680564165115356 ], "probs_raw": [ 0.022782286629080772, 0.9772177338600159 ], "probs": [ 0.9772177338600159, 0.022782286629080772 ], "argmax_index": 0, "argmax": "true", "top2_margin": 0.9544354472309351, "near_tie": false, "gold": "true", "correct": true, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 2, "B": 3 } }, { "qid": "urgency", "type": "score", "K": 3, "calibrate": false, "temperature_key": null, "T": null, "prefix": 199, "n_ids": 43, "positions": 242, "bucket": 256, "ids": [ 1, 17, 27377, 1139, 6634, 988, 768, 5196, 523, 18, 4023, 31149, 856, 779, 5940, 540, 19, 16, 2119, 730, 525, 535, 2577, 7498, 20, 19, 16, 2119, 730, 526, 535, 32238, 20, 19, 16, 2119, 730, 527, 535, 17231, 2166, 20, 21 ], "markers": [ 17, 26, 34 ], "logits_raw": [ -13.792383193969727, -4.434501647949219, -5.2392988204956055 ], "probs_raw": [ 5.961775968899019e-05, 0.6909584999084473, 0.30898186564445496 ], "probs": [ 5.961775968899019e-05, 0.6909584999084473, 0.30898186564445496 ], "argmax_index": 1, "argmax": "1", "top2_margin": 0.3819766342639923, "near_tie": false, "gold": "1", "correct": true, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 1, "B": 3 } } ], "response": { "answers": { "topic": { "type": "choice", "choice": "feedback", "confidence": 0.967538595199585, "probabilities": { "delivery": 0.00016108817362692207, "billing": 0.03228823095560074, "account": 1.204406453325646e-05, "feedback": 0.967538595199585 } }, "wants": { "type": "noul", "noul": 0.9772177338600159 }, "urgency": { "type": "score", "score": 1.3089222311973572, "confidence": 0.6909584999084473, "probabilities": { "0": 5.961775968899019e-05, "1": 0.6909584999084473, "2": 0.30898186564445496 }, "legend": { "0": "Can wait", "1": "Soon", "2": "Right now" } } }, "usage": { "input_tokens": 736, "output_tokens": 0 } }, "oracle_file": "records_ref_audio.json" }, { "id": "aud_08", "mode": "text", "source": "own_audio", "native": false, "prefix": 0, "usage_input_tokens": 142, "questions": [ { "qid": "topic", "type": "choice", "K": 4, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 56, "positions": 56, "bucket": 256, "ids": [ 1, 17, 27377, 1139, 6634, 988, 768, 5196, 523, 18, 3493, 856, 779, 21009, 10789, 1337, 540, 19, 16, 11746, 535, 835, 11746, 933, 768, 15073, 8688, 20, 19, 16, 5363, 800, 535, 61958, 774, 810, 23325, 20, 19, 16, 3863, 535, 21040, 3197, 20, 19, 16, 12779, 535, 36843, 3321, 1337, 779, 3714, 20, 21 ], "markers": [ 18, 29, 39, 46 ], "logits_raw": [ -2.1100118160247803, -1.3073279857635498, -3.290978193283081, 1.7141824960708618 ], "probs_raw": [ 0.020269835367798805, 0.04523257911205292, 0.0062224725261330605, 0.9282751679420471 ], "probs": [ 0.05386649817228317, 0.0955735445022583, 0.023170966655015945, 0.8273889422416687 ], "argmax_index": 3, "argmax": "feedback", "top2_margin": 0.7318153977394104, "near_tie": false, "gold": null, "correct": null, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 0, "B": 3 } }, { "qid": "wants", "type": "noul", "K": 2, "calibrate": true, "temperature_key": "noul:2", "T": 1.6663223505020142, "prefix": 0, "n_ids": 43, "positions": 43, "bucket": 256, "ids": [ 1, 17, 27377, 1139, 6634, 988, 768, 5196, 523, 18, 4940, 779, 21009, 14065, 875, 3503, 811, 874, 4007, 540, 19, 16, 4831, 535, 1295, 521, 779, 8294, 2177, 1014, 3965, 20, 19, 16, 3347, 535, 18672, 521, 779, 8294, 10995, 20, 21 ], "markers": [ 21, 33 ], "logits_raw": [ 2.693018674850464, -4.27215051651001 ], "probs_raw": [ 0.9990566372871399, 0.0009433126542717218 ], "probs": [ 0.015068508684635162, 0.984931468963623 ], "argmax_index": 1, "argmax": "false", "top2_margin": 0.9698629602789879, "near_tie": false, "gold": null, "correct": null, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 1, "B": 3 } }, { "qid": "urgency", "type": "score", "K": 3, "calibrate": true, "temperature_key": "score:3-5", "T": 1.7301132678985596, "prefix": 0, "n_ids": 43, "positions": 43, "bucket": 256, "ids": [ 1, 17, 27377, 1139, 6634, 988, 768, 5196, 523, 18, 4023, 31149, 856, 779, 5940, 540, 19, 16, 2119, 730, 525, 535, 2577, 7498, 20, 19, 16, 2119, 730, 526, 535, 32238, 20, 19, 16, 2119, 730, 527, 535, 17231, 2166, 20, 21 ], "markers": [ 17, 26, 34 ], "logits_raw": [ -8.044438362121582, -6.318297863006592, -8.428503036499023 ], "probs_raw": [ 0.13698606193065643, 0.7697144746780396, 0.09329947084188461 ], "probs": [ 0.22158345580101013, 0.6009449362754822, 0.17747168242931366 ], "argmax_index": 1, "argmax": "1", "top2_margin": 0.37936148047447205, "near_tie": false, "gold": null, "correct": null, "host_softmax_max_abs_dp": 1.4901161193847656e-08, "batch": { "call": 0, "index": 2, "B": 3 } } ], "response": { "answers": { "topic": { "type": "choice", "choice": "feedback", "confidence": 0.8273889422416687, "probabilities": { "delivery": 0.05386649817228317, "billing": 0.0955735445022583, "account": 0.023170966655015945, "feedback": 0.8273889422416687 } }, "wants": { "type": "noul", "noul": 0.015068508684635162 }, "urgency": { "type": "score", "score": 0.9558883011341095, "confidence": 0.6009449362754822, "probabilities": { "0": 0.22158345580101013, "1": 0.6009449362754822, "2": 0.17747168242931366 }, "legend": { "0": "Can wait", "1": "Soon", "2": "Right now" } } }, "usage": { "input_tokens": 142, "output_tokens": 0 } }, "oracle_file": "records_ref_audio.json" }, { "id": "aud_09", "mode": "audio", "source": "own_audio", "native": true, "prefix": 172, "usage_input_tokens": 637, "questions": [ { "qid": "topic", "type": "choice", "K": 4, "calibrate": false, "temperature_key": null, "T": null, "prefix": 172, "n_ids": 57, "positions": 229, "bucket": 256, "ids": [ 1, 17, 24373, 18, 3493, 856, 779, 21009, 10789, 1337, 540, 19, 16, 7832, 572, 1873, 535, 835, 11746, 933, 768, 15073, 8688, 20, 19, 16, 7832, 572, 7778, 535, 61958, 774, 810, 23325, 20, 19, 16, 7832, 572, 12992, 535, 21040, 3197, 20, 19, 16, 7832, 572, 15296, 535, 36843, 3321, 1337, 779, 3714, 20, 21 ], "markers": [ 12, 25, 36, 45 ], "logits_raw": [ -10.309013366699219, -12.59838581085205, -9.959497451782227, -2.7299599647521973 ], "probs_raw": [ 0.0005103875882923603, 5.1717615860980004e-05, 0.0007239239639602602, 0.9987140893936157 ], "probs": [ 0.0005103875882923603, 5.1717615860980004e-05, 0.0007239239639602602, 0.9987140893936157 ], "argmax_index": 3, "argmax": "feedback", "top2_margin": 0.9979901654296555, "near_tie": false, "gold": "account", "correct": false, "host_softmax_max_abs_dp": 1.1920928955078125e-07, "batch": { "call": 0, "index": 0, "B": 3 } }, { "qid": "wants", "type": "noul", "K": 2, "calibrate": false, "temperature_key": null, "T": null, "prefix": 172, "n_ids": 27, "positions": 199, "bucket": 256, "ids": [ 1, 17, 24373, 18, 4940, 779, 21009, 14065, 875, 3503, 811, 874, 4007, 540, 19, 16, 4831, 535, 1295, 20, 19, 16, 3347, 535, 18672, 20, 21 ], "markers": [ 15, 21 ], "logits_raw": [ -1.9508432149887085, -2.6538851261138916 ], "probs_raw": [ 0.6688618659973145, 0.33113816380500793 ], "probs": [ 0.33113816380500793, 0.6688618659973145 ], "argmax_index": 1, "argmax": "false", "top2_margin": 0.3377237021923065, "near_tie": false, "gold": "true", "correct": false, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 2, "B": 3 } }, { "qid": "urgency", "type": "score", "K": 3, "calibrate": false, "temperature_key": null, "T": null, "prefix": 172, "n_ids": 37, "positions": 209, "bucket": 256, "ids": [ 1, 17, 24373, 18, 4023, 31149, 856, 779, 5940, 540, 19, 16, 2119, 730, 525, 535, 2577, 7498, 20, 19, 16, 2119, 730, 526, 535, 32238, 20, 19, 16, 2119, 730, 527, 535, 17231, 2166, 20, 21 ], "markers": [ 11, 20, 28 ], "logits_raw": [ -12.979911804199219, -4.641338348388672, 3.067082166671753 ], "probs_raw": [ 1.0732090771625735e-07, 0.00044882844667881727, 0.999550998210907 ], "probs": [ 1.0732090771625735e-07, 0.00044882844667881727, 0.999550998210907 ], "argmax_index": 2, "argmax": "2", "top2_margin": 0.9991021697642282, "near_tie": false, "gold": "2", "correct": true, "host_softmax_max_abs_dp": 7.105427357601002e-15, "batch": { "call": 0, "index": 1, "B": 3 } } ], "response": { "answers": { "topic": { "type": "choice", "choice": "feedback", "confidence": 0.9987140893936157, "probabilities": { "delivery": 0.0005103875882923603, "billing": 5.1717615860980004e-05, "account": 0.0007239239639602602, "feedback": 0.9987140893936157 } }, "wants": { "type": "noul", "noul": 0.33113816380500793 }, "urgency": { "type": "score", "score": 1.9995508248684928, "confidence": 0.999550998210907, "probabilities": { "0": 1.0732090771625735e-07, "1": 0.00044882844667881727, "2": 0.999550998210907 }, "legend": { "0": "Can wait", "1": "Soon", "2": "Right now" } } }, "usage": { "input_tokens": 637, "output_tokens": 0 } }, "oracle_file": "records_ref_audio.json" }, { "id": "aud_09", "mode": "text", "source": "own_audio", "native": false, "prefix": 0, "usage_input_tokens": 121, "questions": [ { "qid": "topic", "type": "choice", "K": 4, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 49, "positions": 49, "bucket": 256, "ids": [ 1, 17, 18, 3493, 856, 779, 21009, 10789, 1337, 540, 19, 16, 11746, 535, 835, 11746, 933, 768, 15073, 8688, 20, 19, 16, 5363, 800, 535, 61958, 774, 810, 23325, 20, 19, 16, 3863, 535, 21040, 3197, 20, 19, 16, 12779, 535, 36843, 3321, 1337, 779, 3714, 20, 21 ], "markers": [ 11, 22, 32, 39 ], "logits_raw": [ -4.030765533447266, -2.516629695892334, -4.948546409606934, -1.2068599462509155 ], "probs_raw": [ 0.04388390854001045, 0.19947391748428345, 0.017527427524328232, 0.7391147613525391 ], "probs": [ 0.08343102037906647, 0.24607208371162415, 0.043311357498168945, 0.6271855235099792 ], "argmax_index": 3, "argmax": "feedback", "top2_margin": 0.3811134397983551, "near_tie": false, "gold": null, "correct": null, "host_softmax_max_abs_dp": 1.4901161193847656e-08, "batch": { "call": 0, "index": 0, "B": 3 } }, { "qid": "wants", "type": "noul", "K": 2, "calibrate": true, "temperature_key": "noul:2", "T": 1.6663223505020142, "prefix": 0, "n_ids": 36, "positions": 36, "bucket": 256, "ids": [ 1, 17, 18, 4940, 779, 21009, 14065, 875, 3503, 811, 874, 4007, 540, 19, 16, 4831, 535, 1295, 521, 779, 8294, 2177, 1014, 3965, 20, 19, 16, 3347, 535, 18672, 521, 779, 8294, 10995, 20, 21 ], "markers": [ 14, 26 ], "logits_raw": [ -0.19816279411315918, 3.9951019287109375 ], "probs_raw": [ 0.014872386120259762, 0.9851276278495789 ], "probs": [ 0.9252889752388, 0.07471100240945816 ], "argmax_index": 0, "argmax": "true", "top2_margin": 0.8505779728293419, "near_tie": false, "gold": null, "correct": null, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 1, "B": 3 } }, { "qid": "urgency", "type": "score", "K": 3, "calibrate": true, "temperature_key": "score:3-5", "T": 1.7301132678985596, "prefix": 0, "n_ids": 36, "positions": 36, "bucket": 256, "ids": [ 1, 17, 18, 4023, 31149, 856, 779, 5940, 540, 19, 16, 2119, 730, 525, 535, 2577, 7498, 20, 19, 16, 2119, 730, 526, 535, 32238, 20, 19, 16, 2119, 730, 527, 535, 17231, 2166, 20, 21 ], "markers": [ 10, 19, 27 ], "logits_raw": [ -7.304863452911377, -6.445466995239258, -5.352657318115234 ], "probs_raw": [ 0.09609891474246979, 0.22696013748645782, 0.6769409775733948 ], "probs": [ 0.17439991235733032, 0.28659766912460327, 0.5390024185180664 ], "argmax_index": 2, "argmax": "2", "top2_margin": 0.25240474939346313, "near_tie": false, "gold": null, "correct": null, "host_softmax_max_abs_dp": 2.9802322387695312e-08, "batch": { "call": 0, "index": 2, "B": 3 } } ], "response": { "answers": { "topic": { "type": "choice", "choice": "feedback", "confidence": 0.6271855235099792, "probabilities": { "delivery": 0.08343102037906647, "billing": 0.24607208371162415, "account": 0.043311357498168945, "feedback": 0.6271855235099792 } }, "wants": { "type": "noul", "noul": 0.9252889752388 }, "urgency": { "type": "score", "score": 1.364602506160736, "confidence": 0.5390024185180664, "probabilities": { "0": 0.17439991235733032, "1": 0.28659766912460327, "2": 0.5390024185180664 }, "legend": { "0": "Can wait", "1": "Soon", "2": "Right now" } } }, "usage": { "input_tokens": 121, "output_tokens": 0 } }, "oracle_file": "records_ref_audio.json" }, { "id": "aud_10", "mode": "audio", "source": "own_audio", "native": true, "prefix": 269, "usage_input_tokens": 946, "questions": [ { "qid": "topic", "type": "choice", "K": 4, "calibrate": false, "temperature_key": null, "T": null, "prefix": 269, "n_ids": 63, "positions": 332, "bucket": 512, "ids": [ 1, 17, 27377, 1139, 6634, 988, 768, 5196, 523, 18, 3493, 856, 779, 21009, 10789, 1337, 540, 19, 16, 7832, 572, 1873, 535, 835, 11746, 933, 768, 15073, 8688, 20, 19, 16, 7832, 572, 7778, 535, 61958, 774, 810, 23325, 20, 19, 16, 7832, 572, 12992, 535, 21040, 3197, 20, 19, 16, 7832, 572, 15296, 535, 36843, 3321, 1337, 779, 3714, 20, 21 ], "markers": [ 18, 31, 42, 51 ], "logits_raw": [ -6.047428131103516, -11.045299530029297, -10.394746780395508, -0.31905287504196167 ], "probs_raw": [ 0.003241606056690216, 2.1888312403461896e-05, 4.195113797322847e-05, 0.9966945648193359 ], "probs": [ 0.003241606056690216, 2.1888312403461896e-05, 4.195113797322847e-05, 0.9966945648193359 ], "argmax_index": 3, "argmax": "feedback", "top2_margin": 0.9934529587626457, "near_tie": false, "gold": "feedback", "correct": true, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 0, "B": 3 } }, { "qid": "wants", "type": "noul", "K": 2, "calibrate": false, "temperature_key": null, "T": null, "prefix": 269, "n_ids": 33, "positions": 302, "bucket": 512, "ids": [ 1, 17, 27377, 1139, 6634, 988, 768, 5196, 523, 18, 4940, 779, 21009, 14065, 875, 3503, 811, 874, 4007, 540, 19, 16, 4831, 535, 1295, 20, 19, 16, 3347, 535, 18672, 20, 21 ], "markers": [ 21, 27 ], "logits_raw": [ -0.8847565054893494, -3.556145191192627 ], "probs_raw": [ 0.9353170394897461, 0.06468290090560913 ], "probs": [ 0.06468290090560913, 0.9353170394897461 ], "argmax_index": 1, "argmax": "false", "top2_margin": 0.870634138584137, "near_tie": false, "gold": "false", "correct": true, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 2, "B": 3 } }, { "qid": "urgency", "type": "score", "K": 3, "calibrate": false, "temperature_key": null, "T": null, "prefix": 269, "n_ids": 43, "positions": 312, "bucket": 512, "ids": [ 1, 17, 27377, 1139, 6634, 988, 768, 5196, 523, 18, 4023, 31149, 856, 779, 5940, 540, 19, 16, 2119, 730, 525, 535, 2577, 7498, 20, 19, 16, 2119, 730, 526, 535, 32238, 20, 19, 16, 2119, 730, 527, 535, 17231, 2166, 20, 21 ], "markers": [ 17, 26, 34 ], "logits_raw": [ -12.225658416748047, -3.0319995880126953, -1.7885916233062744 ], "probs_raw": [ 2.27603468374582e-05, 0.22383823990821838, 0.776138961315155 ], "probs": [ 2.27603468374582e-05, 0.22383823990821838, 0.776138961315155 ], "argmax_index": 2, "argmax": "2", "top2_margin": 0.5523007214069366, "near_tie": false, "gold": "0", "correct": false, "host_softmax_max_abs_dp": 1.4901161193847656e-08, "batch": { "call": 0, "index": 1, "B": 3 } } ], "response": { "answers": { "topic": { "type": "choice", "choice": "feedback", "confidence": 0.9966945648193359, "probabilities": { "delivery": 0.003241606056690216, "billing": 2.1888312403461896e-05, "account": 4.195113797322847e-05, "feedback": 0.9966945648193359 } }, "wants": { "type": "noul", "noul": 0.06468290090560913 }, "urgency": { "type": "score", "score": 1.7761161625385284, "confidence": 0.776138961315155, "probabilities": { "0": 2.27603468374582e-05, "1": 0.22383823990821838, "2": 0.776138961315155 }, "legend": { "0": "Can wait", "1": "Soon", "2": "Right now" } } }, "usage": { "input_tokens": 946, "output_tokens": 0 } }, "oracle_file": "records_ref_audio.json" }, { "id": "aud_10", "mode": "text", "source": "own_audio", "native": false, "prefix": 0, "usage_input_tokens": 142, "questions": [ { "qid": "topic", "type": "choice", "K": 4, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 56, "positions": 56, "bucket": 256, "ids": [ 1, 17, 27377, 1139, 6634, 988, 768, 5196, 523, 18, 3493, 856, 779, 21009, 10789, 1337, 540, 19, 16, 11746, 535, 835, 11746, 933, 768, 15073, 8688, 20, 19, 16, 5363, 800, 535, 61958, 774, 810, 23325, 20, 19, 16, 3863, 535, 21040, 3197, 20, 19, 16, 12779, 535, 36843, 3321, 1337, 779, 3714, 20, 21 ], "markers": [ 18, 29, 39, 46 ], "logits_raw": [ -2.1100118160247803, -1.3073279857635498, -3.290978193283081, 1.7141824960708618 ], "probs_raw": [ 0.020269835367798805, 0.04523257911205292, 0.0062224725261330605, 0.9282751679420471 ], "probs": [ 0.05386649817228317, 0.0955735445022583, 0.023170966655015945, 0.8273889422416687 ], "argmax_index": 3, "argmax": "feedback", "top2_margin": 0.7318153977394104, "near_tie": false, "gold": null, "correct": null, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 0, "B": 3 } }, { "qid": "wants", "type": "noul", "K": 2, "calibrate": true, "temperature_key": "noul:2", "T": 1.6663223505020142, "prefix": 0, "n_ids": 43, "positions": 43, "bucket": 256, "ids": [ 1, 17, 27377, 1139, 6634, 988, 768, 5196, 523, 18, 4940, 779, 21009, 14065, 875, 3503, 811, 874, 4007, 540, 19, 16, 4831, 535, 1295, 521, 779, 8294, 2177, 1014, 3965, 20, 19, 16, 3347, 535, 18672, 521, 779, 8294, 10995, 20, 21 ], "markers": [ 21, 33 ], "logits_raw": [ 2.693018674850464, -4.27215051651001 ], "probs_raw": [ 0.9990566372871399, 0.0009433126542717218 ], "probs": [ 0.015068508684635162, 0.984931468963623 ], "argmax_index": 1, "argmax": "false", "top2_margin": 0.9698629602789879, "near_tie": false, "gold": null, "correct": null, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 1, "B": 3 } }, { "qid": "urgency", "type": "score", "K": 3, "calibrate": true, "temperature_key": "score:3-5", "T": 1.7301132678985596, "prefix": 0, "n_ids": 43, "positions": 43, "bucket": 256, "ids": [ 1, 17, 27377, 1139, 6634, 988, 768, 5196, 523, 18, 4023, 31149, 856, 779, 5940, 540, 19, 16, 2119, 730, 525, 535, 2577, 7498, 20, 19, 16, 2119, 730, 526, 535, 32238, 20, 19, 16, 2119, 730, 527, 535, 17231, 2166, 20, 21 ], "markers": [ 17, 26, 34 ], "logits_raw": [ -8.044438362121582, -6.318297863006592, -8.428503036499023 ], "probs_raw": [ 0.13698606193065643, 0.7697144746780396, 0.09329947084188461 ], "probs": [ 0.22158345580101013, 0.6009449362754822, 0.17747168242931366 ], "argmax_index": 1, "argmax": "1", "top2_margin": 0.37936148047447205, "near_tie": false, "gold": null, "correct": null, "host_softmax_max_abs_dp": 1.4901161193847656e-08, "batch": { "call": 0, "index": 2, "B": 3 } } ], "response": { "answers": { "topic": { "type": "choice", "choice": "feedback", "confidence": 0.8273889422416687, "probabilities": { "delivery": 0.05386649817228317, "billing": 0.0955735445022583, "account": 0.023170966655015945, "feedback": 0.8273889422416687 } }, "wants": { "type": "noul", "noul": 0.015068508684635162 }, "urgency": { "type": "score", "score": 0.9558883011341095, "confidence": 0.6009449362754822, "probabilities": { "0": 0.22158345580101013, "1": 0.6009449362754822, "2": 0.17747168242931366 }, "legend": { "0": "Can wait", "1": "Soon", "2": "Right now" } } }, "usage": { "input_tokens": 142, "output_tokens": 0 } }, "oracle_file": "records_ref_audio.json" }, { "id": "aud_11", "mode": "audio", "source": "own_audio", "native": true, "prefix": 263, "usage_input_tokens": 910, "questions": [ { "qid": "topic", "type": "choice", "K": 4, "calibrate": false, "temperature_key": null, "T": null, "prefix": 263, "n_ids": 57, "positions": 320, "bucket": 512, "ids": [ 1, 17, 24373, 18, 3493, 856, 779, 21009, 10789, 1337, 540, 19, 16, 7832, 572, 1873, 535, 835, 11746, 933, 768, 15073, 8688, 20, 19, 16, 7832, 572, 7778, 535, 61958, 774, 810, 23325, 20, 19, 16, 7832, 572, 12992, 535, 21040, 3197, 20, 19, 16, 7832, 572, 15296, 535, 36843, 3321, 1337, 779, 3714, 20, 21 ], "markers": [ 12, 25, 36, 45 ], "logits_raw": [ -5.378786087036133, -6.809049606323242, -13.527311325073242, -2.2166378498077393 ], "probs_raw": [ 0.04022393375635147, 0.009623407386243343, 1.1631209417828359e-05, 0.9501410126686096 ], "probs": [ 0.04022393375635147, 0.009623407386243343, 1.1631209417828359e-05, 0.9501410126686096 ], "argmax_index": 3, "argmax": "feedback", "top2_margin": 0.9099170789122581, "near_tie": false, "gold": "delivery", "correct": false, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 0, "B": 3 } }, { "qid": "wants", "type": "noul", "K": 2, "calibrate": false, "temperature_key": null, "T": null, "prefix": 263, "n_ids": 27, "positions": 290, "bucket": 512, "ids": [ 1, 17, 24373, 18, 4940, 779, 21009, 14065, 875, 3503, 811, 874, 4007, 540, 19, 16, 4831, 535, 1295, 20, 19, 16, 3347, 535, 18672, 20, 21 ], "markers": [ 15, 21 ], "logits_raw": [ -2.169802665710449, -2.4261226654052734 ], "probs_raw": [ 0.5637314319610596, 0.43626853823661804 ], "probs": [ 0.43626853823661804, 0.5637314319610596 ], "argmax_index": 1, "argmax": "false", "top2_margin": 0.12746289372444153, "near_tie": false, "gold": "true", "correct": false, "host_softmax_max_abs_dp": 2.9802322387695312e-08, "batch": { "call": 0, "index": 2, "B": 3 } }, { "qid": "urgency", "type": "score", "K": 3, "calibrate": false, "temperature_key": null, "T": null, "prefix": 263, "n_ids": 37, "positions": 300, "bucket": 512, "ids": [ 1, 17, 24373, 18, 4023, 31149, 856, 779, 5940, 540, 19, 16, 2119, 730, 525, 535, 2577, 7498, 20, 19, 16, 2119, 730, 526, 535, 32238, 20, 19, 16, 2119, 730, 527, 535, 17231, 2166, 20, 21 ], "markers": [ 11, 20, 28 ], "logits_raw": [ -13.344354629516602, -3.0399885177612305, -0.4444863498210907 ], "probs_raw": [ 2.3249160676641623e-06, 0.06942828744649887, 0.9305693507194519 ], "probs": [ 2.3249160676641623e-06, 0.06942828744649887, 0.9305693507194519 ], "argmax_index": 2, "argmax": "2", "top2_margin": 0.861141063272953, "near_tie": false, "gold": "1", "correct": false, "host_softmax_max_abs_dp": 2.2737367544323206e-13, "batch": { "call": 0, "index": 1, "B": 3 } } ], "response": { "answers": { "topic": { "type": "choice", "choice": "feedback", "confidence": 0.9501410126686096, "probabilities": { "delivery": 0.04022393375635147, "billing": 0.009623407386243343, "account": 1.1631209417828359e-05, "feedback": 0.9501410126686096 } }, "wants": { "type": "noul", "noul": 0.43626853823661804 }, "urgency": { "type": "score", "score": 1.9305669888854027, "confidence": 0.9305693507194519, "probabilities": { "0": 2.3249160676641623e-06, "1": 0.06942828744649887, "2": 0.9305693507194519 }, "legend": { "0": "Can wait", "1": "Soon", "2": "Right now" } } }, "usage": { "input_tokens": 910, "output_tokens": 0 } }, "oracle_file": "records_ref_audio.json" }, { "id": "aud_11", "mode": "text", "source": "own_audio", "native": false, "prefix": 0, "usage_input_tokens": 121, "questions": [ { "qid": "topic", "type": "choice", "K": 4, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 49, "positions": 49, "bucket": 256, "ids": [ 1, 17, 18, 3493, 856, 779, 21009, 10789, 1337, 540, 19, 16, 11746, 535, 835, 11746, 933, 768, 15073, 8688, 20, 19, 16, 5363, 800, 535, 61958, 774, 810, 23325, 20, 19, 16, 3863, 535, 21040, 3197, 20, 19, 16, 12779, 535, 36843, 3321, 1337, 779, 3714, 20, 21 ], "markers": [ 11, 22, 32, 39 ], "logits_raw": [ -4.030765533447266, -2.516629695892334, -4.948546409606934, -1.2068599462509155 ], "probs_raw": [ 0.04388390854001045, 0.19947391748428345, 0.017527427524328232, 0.7391147613525391 ], "probs": [ 0.08343102037906647, 0.24607208371162415, 0.043311357498168945, 0.6271855235099792 ], "argmax_index": 3, "argmax": "feedback", "top2_margin": 0.3811134397983551, "near_tie": false, "gold": null, "correct": null, "host_softmax_max_abs_dp": 1.4901161193847656e-08, "batch": { "call": 0, "index": 0, "B": 3 } }, { "qid": "wants", "type": "noul", "K": 2, "calibrate": true, "temperature_key": "noul:2", "T": 1.6663223505020142, "prefix": 0, "n_ids": 36, "positions": 36, "bucket": 256, "ids": [ 1, 17, 18, 4940, 779, 21009, 14065, 875, 3503, 811, 874, 4007, 540, 19, 16, 4831, 535, 1295, 521, 779, 8294, 2177, 1014, 3965, 20, 19, 16, 3347, 535, 18672, 521, 779, 8294, 10995, 20, 21 ], "markers": [ 14, 26 ], "logits_raw": [ -0.19816279411315918, 3.9951019287109375 ], "probs_raw": [ 0.014872386120259762, 0.9851276278495789 ], "probs": [ 0.9252889752388, 0.07471100240945816 ], "argmax_index": 0, "argmax": "true", "top2_margin": 0.8505779728293419, "near_tie": false, "gold": null, "correct": null, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 1, "B": 3 } }, { "qid": "urgency", "type": "score", "K": 3, "calibrate": true, "temperature_key": "score:3-5", "T": 1.7301132678985596, "prefix": 0, "n_ids": 36, "positions": 36, "bucket": 256, "ids": [ 1, 17, 18, 4023, 31149, 856, 779, 5940, 540, 19, 16, 2119, 730, 525, 535, 2577, 7498, 20, 19, 16, 2119, 730, 526, 535, 32238, 20, 19, 16, 2119, 730, 527, 535, 17231, 2166, 20, 21 ], "markers": [ 10, 19, 27 ], "logits_raw": [ -7.304863452911377, -6.445466995239258, -5.352657318115234 ], "probs_raw": [ 0.09609891474246979, 0.22696013748645782, 0.6769409775733948 ], "probs": [ 0.17439991235733032, 0.28659766912460327, 0.5390024185180664 ], "argmax_index": 2, "argmax": "2", "top2_margin": 0.25240474939346313, "near_tie": false, "gold": null, "correct": null, "host_softmax_max_abs_dp": 2.9802322387695312e-08, "batch": { "call": 0, "index": 2, "B": 3 } } ], "response": { "answers": { "topic": { "type": "choice", "choice": "feedback", "confidence": 0.6271855235099792, "probabilities": { "delivery": 0.08343102037906647, "billing": 0.24607208371162415, "account": 0.043311357498168945, "feedback": 0.6271855235099792 } }, "wants": { "type": "noul", "noul": 0.9252889752388 }, "urgency": { "type": "score", "score": 1.364602506160736, "confidence": 0.5390024185180664, "probabilities": { "0": 0.17439991235733032, "1": 0.28659766912460327, "2": 0.5390024185180664 }, "legend": { "0": "Can wait", "1": "Soon", "2": "Right now" } } }, "usage": { "input_tokens": 121, "output_tokens": 0 } }, "oracle_file": "records_ref_audio.json" }, { "id": "aud_12", "mode": "audio", "source": "own_audio", "native": true, "prefix": 350, "usage_input_tokens": 1189, "questions": [ { "qid": "topic", "type": "choice", "K": 4, "calibrate": false, "temperature_key": null, "T": null, "prefix": 350, "n_ids": 63, "positions": 413, "bucket": 512, "ids": [ 1, 17, 27377, 1139, 6634, 988, 768, 5196, 523, 18, 3493, 856, 779, 21009, 10789, 1337, 540, 19, 16, 7832, 572, 1873, 535, 835, 11746, 933, 768, 15073, 8688, 20, 19, 16, 7832, 572, 7778, 535, 61958, 774, 810, 23325, 20, 19, 16, 7832, 572, 12992, 535, 21040, 3197, 20, 19, 16, 7832, 572, 15296, 535, 36843, 3321, 1337, 779, 3714, 20, 21 ], "markers": [ 18, 31, 42, 51 ], "logits_raw": [ -8.041118621826172, -6.483062744140625, -13.2620849609375, -0.41955888271331787 ], "probs_raw": [ 0.0004884006921201944, 0.0023196975234895945, 2.638391379150562e-06, 0.9971892237663269 ], "probs": [ 0.0004884006921201944, 0.0023196975234895945, 2.638391379150562e-06, 0.9971892237663269 ], "argmax_index": 3, "argmax": "feedback", "top2_margin": 0.9948695262428373, "near_tie": false, "gold": "billing", "correct": false, "host_softmax_max_abs_dp": 2.2737367544323206e-13, "batch": { "call": 0, "index": 0, "B": 3 } }, { "qid": "wants", "type": "noul", "K": 2, "calibrate": false, "temperature_key": null, "T": null, "prefix": 350, "n_ids": 33, "positions": 383, "bucket": 512, "ids": [ 1, 17, 27377, 1139, 6634, 988, 768, 5196, 523, 18, 4940, 779, 21009, 14065, 875, 3503, 811, 874, 4007, 540, 19, 16, 4831, 535, 1295, 20, 19, 16, 3347, 535, 18672, 20, 21 ], "markers": [ 21, 27 ], "logits_raw": [ -0.8726661205291748, -2.7932546138763428 ], "probs_raw": [ 0.8722040057182312, 0.12779594957828522 ], "probs": [ 0.12779594957828522, 0.8722040057182312 ], "argmax_index": 1, "argmax": "false", "top2_margin": 0.744408056139946, "near_tie": false, "gold": "true", "correct": false, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 2, "B": 3 } }, { "qid": "urgency", "type": "score", "K": 3, "calibrate": false, "temperature_key": null, "T": null, "prefix": 350, "n_ids": 43, "positions": 393, "bucket": 512, "ids": [ 1, 17, 27377, 1139, 6634, 988, 768, 5196, 523, 18, 4023, 31149, 856, 779, 5940, 540, 19, 16, 2119, 730, 525, 535, 2577, 7498, 20, 19, 16, 2119, 730, 526, 535, 32238, 20, 19, 16, 2119, 730, 527, 535, 17231, 2166, 20, 21 ], "markers": [ 17, 26, 34 ], "logits_raw": [ -10.186288833618164, -1.5687068700790405, 0.817966103553772 ], "probs_raw": [ 1.5230326425808016e-05, 0.08419332653284073, 0.915791392326355 ], "probs": [ 1.5230326425808016e-05, 0.08419332653284073, 0.915791392326355 ], "argmax_index": 2, "argmax": "2", "top2_margin": 0.8315980657935143, "near_tie": false, "gold": "2", "correct": true, "host_softmax_max_abs_dp": 9.094947017729282e-13, "batch": { "call": 0, "index": 1, "B": 3 } } ], "response": { "answers": { "topic": { "type": "choice", "choice": "feedback", "confidence": 0.9971892237663269, "probabilities": { "delivery": 0.0004884006921201944, "billing": 0.0023196975234895945, "account": 2.638391379150562e-06, "feedback": 0.9971892237663269 } }, "wants": { "type": "noul", "noul": 0.12779594957828522 }, "urgency": { "type": "score", "score": 1.9157761111855507, "confidence": 0.915791392326355, "probabilities": { "0": 1.5230326425808016e-05, "1": 0.08419332653284073, "2": 0.915791392326355 }, "legend": { "0": "Can wait", "1": "Soon", "2": "Right now" } } }, "usage": { "input_tokens": 1189, "output_tokens": 0 } }, "oracle_file": "records_ref_audio.json" }, { "id": "aud_12", "mode": "text", "source": "own_audio", "native": false, "prefix": 0, "usage_input_tokens": 142, "questions": [ { "qid": "topic", "type": "choice", "K": 4, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 56, "positions": 56, "bucket": 256, "ids": [ 1, 17, 27377, 1139, 6634, 988, 768, 5196, 523, 18, 3493, 856, 779, 21009, 10789, 1337, 540, 19, 16, 11746, 535, 835, 11746, 933, 768, 15073, 8688, 20, 19, 16, 5363, 800, 535, 61958, 774, 810, 23325, 20, 19, 16, 3863, 535, 21040, 3197, 20, 19, 16, 12779, 535, 36843, 3321, 1337, 779, 3714, 20, 21 ], "markers": [ 18, 29, 39, 46 ], "logits_raw": [ -2.1100118160247803, -1.3073279857635498, -3.290978193283081, 1.7141824960708618 ], "probs_raw": [ 0.020269835367798805, 0.04523257911205292, 0.0062224725261330605, 0.9282751679420471 ], "probs": [ 0.05386649817228317, 0.0955735445022583, 0.023170966655015945, 0.8273889422416687 ], "argmax_index": 3, "argmax": "feedback", "top2_margin": 0.7318153977394104, "near_tie": false, "gold": null, "correct": null, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 0, "B": 3 } }, { "qid": "wants", "type": "noul", "K": 2, "calibrate": true, "temperature_key": "noul:2", "T": 1.6663223505020142, "prefix": 0, "n_ids": 43, "positions": 43, "bucket": 256, "ids": [ 1, 17, 27377, 1139, 6634, 988, 768, 5196, 523, 18, 4940, 779, 21009, 14065, 875, 3503, 811, 874, 4007, 540, 19, 16, 4831, 535, 1295, 521, 779, 8294, 2177, 1014, 3965, 20, 19, 16, 3347, 535, 18672, 521, 779, 8294, 10995, 20, 21 ], "markers": [ 21, 33 ], "logits_raw": [ 2.693018674850464, -4.27215051651001 ], "probs_raw": [ 0.9990566372871399, 0.0009433126542717218 ], "probs": [ 0.015068508684635162, 0.984931468963623 ], "argmax_index": 1, "argmax": "false", "top2_margin": 0.9698629602789879, "near_tie": false, "gold": null, "correct": null, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 1, "B": 3 } }, { "qid": "urgency", "type": "score", "K": 3, "calibrate": true, "temperature_key": "score:3-5", "T": 1.7301132678985596, "prefix": 0, "n_ids": 43, "positions": 43, "bucket": 256, "ids": [ 1, 17, 27377, 1139, 6634, 988, 768, 5196, 523, 18, 4023, 31149, 856, 779, 5940, 540, 19, 16, 2119, 730, 525, 535, 2577, 7498, 20, 19, 16, 2119, 730, 526, 535, 32238, 20, 19, 16, 2119, 730, 527, 535, 17231, 2166, 20, 21 ], "markers": [ 17, 26, 34 ], "logits_raw": [ -8.044438362121582, -6.318297863006592, -8.428503036499023 ], "probs_raw": [ 0.13698606193065643, 0.7697144746780396, 0.09329947084188461 ], "probs": [ 0.22158345580101013, 0.6009449362754822, 0.17747168242931366 ], "argmax_index": 1, "argmax": "1", "top2_margin": 0.37936148047447205, "near_tie": false, "gold": null, "correct": null, "host_softmax_max_abs_dp": 1.4901161193847656e-08, "batch": { "call": 0, "index": 2, "B": 3 } } ], "response": { "answers": { "topic": { "type": "choice", "choice": "feedback", "confidence": 0.8273889422416687, "probabilities": { "delivery": 0.05386649817228317, "billing": 0.0955735445022583, "account": 0.023170966655015945, "feedback": 0.8273889422416687 } }, "wants": { "type": "noul", "noul": 0.015068508684635162 }, "urgency": { "type": "score", "score": 0.9558883011341095, "confidence": 0.6009449362754822, "probabilities": { "0": 0.22158345580101013, "1": 0.6009449362754822, "2": 0.17747168242931366 }, "legend": { "0": "Can wait", "1": "Soon", "2": "Right now" } } }, "usage": { "input_tokens": 142, "output_tokens": 0 } }, "oracle_file": "records_ref_audio.json" }, { "id": "aud_13", "mode": "audio", "source": "own_audio", "native": true, "prefix": 356, "usage_input_tokens": 1189, "questions": [ { "qid": "topic", "type": "choice", "K": 4, "calibrate": false, "temperature_key": null, "T": null, "prefix": 356, "n_ids": 57, "positions": 413, "bucket": 512, "ids": [ 1, 17, 24373, 18, 3493, 856, 779, 21009, 10789, 1337, 540, 19, 16, 7832, 572, 1873, 535, 835, 11746, 933, 768, 15073, 8688, 20, 19, 16, 7832, 572, 7778, 535, 61958, 774, 810, 23325, 20, 19, 16, 7832, 572, 12992, 535, 21040, 3197, 20, 19, 16, 7832, 572, 15296, 535, 36843, 3321, 1337, 779, 3714, 20, 21 ], "markers": [ 12, 25, 36, 45 ], "logits_raw": [ -10.233248710632324, -8.533600807189941, -11.744644165039062, -1.8605585098266602 ], "probs_raw": [ 0.00023073617194313556, 0.0012625929666683078, 5.0900838687084615e-05, 0.9984557628631592 ], "probs": [ 0.00023073617194313556, 0.0012625929666683078, 5.0900838687084615e-05, 0.9984557628631592 ], "argmax_index": 3, "argmax": "feedback", "top2_margin": 0.9971931698964909, "near_tie": false, "gold": "account", "correct": false, "host_softmax_max_abs_dp": 3.637978807091713e-12, "batch": { "call": 0, "index": 0, "B": 3 } }, { "qid": "wants", "type": "noul", "K": 2, "calibrate": false, "temperature_key": null, "T": null, "prefix": 356, "n_ids": 27, "positions": 383, "bucket": 512, "ids": [ 1, 17, 24373, 18, 4940, 779, 21009, 14065, 875, 3503, 811, 874, 4007, 540, 19, 16, 4831, 535, 1295, 20, 19, 16, 3347, 535, 18672, 20, 21 ], "markers": [ 15, 21 ], "logits_raw": [ -1.111641764640808, -1.4849694967269897 ], "probs_raw": [ 0.5922628045082092, 0.4077371656894684 ], "probs": [ 0.4077371656894684, 0.5922628045082092 ], "argmax_index": 1, "argmax": "false", "top2_margin": 0.18452563881874084, "near_tie": false, "gold": "true", "correct": false, "host_softmax_max_abs_dp": 2.9802322387695312e-08, "batch": { "call": 0, "index": 2, "B": 3 } }, { "qid": "urgency", "type": "score", "K": 3, "calibrate": false, "temperature_key": null, "T": null, "prefix": 356, "n_ids": 37, "positions": 393, "bucket": 512, "ids": [ 1, 17, 24373, 18, 4023, 31149, 856, 779, 5940, 540, 19, 16, 2119, 730, 525, 535, 2577, 7498, 20, 19, 16, 2119, 730, 526, 535, 32238, 20, 19, 16, 2119, 730, 527, 535, 17231, 2166, 20, 21 ], "markers": [ 11, 20, 28 ], "logits_raw": [ -8.093282699584961, -3.591269016265869, 1.4334510564804077 ], "probs_raw": [ 7.239612750709057e-05, 0.006530025042593479, 0.99339759349823 ], "probs": [ 7.239612750709057e-05, 0.006530025042593479, 0.99339759349823 ], "argmax_index": 2, "argmax": "2", "top2_margin": 0.9868675684556365, "near_tie": false, "gold": "0", "correct": false, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 1, "B": 3 } } ], "response": { "answers": { "topic": { "type": "choice", "choice": "feedback", "confidence": 0.9984557628631592, "probabilities": { "delivery": 0.00023073617194313556, "billing": 0.0012625929666683078, "account": 5.0900838687084615e-05, "feedback": 0.9984557628631592 } }, "wants": { "type": "noul", "noul": 0.4077371656894684 }, "urgency": { "type": "score", "score": 1.9933252120390534, "confidence": 0.99339759349823, "probabilities": { "0": 7.239612750709057e-05, "1": 0.006530025042593479, "2": 0.99339759349823 }, "legend": { "0": "Can wait", "1": "Soon", "2": "Right now" } } }, "usage": { "input_tokens": 1189, "output_tokens": 0 } }, "oracle_file": "records_ref_audio.json" }, { "id": "aud_13", "mode": "text", "source": "own_audio", "native": false, "prefix": 0, "usage_input_tokens": 121, "questions": [ { "qid": "topic", "type": "choice", "K": 4, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 49, "positions": 49, "bucket": 256, "ids": [ 1, 17, 18, 3493, 856, 779, 21009, 10789, 1337, 540, 19, 16, 11746, 535, 835, 11746, 933, 768, 15073, 8688, 20, 19, 16, 5363, 800, 535, 61958, 774, 810, 23325, 20, 19, 16, 3863, 535, 21040, 3197, 20, 19, 16, 12779, 535, 36843, 3321, 1337, 779, 3714, 20, 21 ], "markers": [ 11, 22, 32, 39 ], "logits_raw": [ -4.030765533447266, -2.516629695892334, -4.948546409606934, -1.2068599462509155 ], "probs_raw": [ 0.04388390854001045, 0.19947391748428345, 0.017527427524328232, 0.7391147613525391 ], "probs": [ 0.08343102037906647, 0.24607208371162415, 0.043311357498168945, 0.6271855235099792 ], "argmax_index": 3, "argmax": "feedback", "top2_margin": 0.3811134397983551, "near_tie": false, "gold": null, "correct": null, "host_softmax_max_abs_dp": 1.4901161193847656e-08, "batch": { "call": 0, "index": 0, "B": 3 } }, { "qid": "wants", "type": "noul", "K": 2, "calibrate": true, "temperature_key": "noul:2", "T": 1.6663223505020142, "prefix": 0, "n_ids": 36, "positions": 36, "bucket": 256, "ids": [ 1, 17, 18, 4940, 779, 21009, 14065, 875, 3503, 811, 874, 4007, 540, 19, 16, 4831, 535, 1295, 521, 779, 8294, 2177, 1014, 3965, 20, 19, 16, 3347, 535, 18672, 521, 779, 8294, 10995, 20, 21 ], "markers": [ 14, 26 ], "logits_raw": [ -0.19816279411315918, 3.9951019287109375 ], "probs_raw": [ 0.014872386120259762, 0.9851276278495789 ], "probs": [ 0.9252889752388, 0.07471100240945816 ], "argmax_index": 0, "argmax": "true", "top2_margin": 0.8505779728293419, "near_tie": false, "gold": null, "correct": null, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 1, "B": 3 } }, { "qid": "urgency", "type": "score", "K": 3, "calibrate": true, "temperature_key": "score:3-5", "T": 1.7301132678985596, "prefix": 0, "n_ids": 36, "positions": 36, "bucket": 256, "ids": [ 1, 17, 18, 4023, 31149, 856, 779, 5940, 540, 19, 16, 2119, 730, 525, 535, 2577, 7498, 20, 19, 16, 2119, 730, 526, 535, 32238, 20, 19, 16, 2119, 730, 527, 535, 17231, 2166, 20, 21 ], "markers": [ 10, 19, 27 ], "logits_raw": [ -7.304863452911377, -6.445466995239258, -5.352657318115234 ], "probs_raw": [ 0.09609891474246979, 0.22696013748645782, 0.6769409775733948 ], "probs": [ 0.17439991235733032, 0.28659766912460327, 0.5390024185180664 ], "argmax_index": 2, "argmax": "2", "top2_margin": 0.25240474939346313, "near_tie": false, "gold": null, "correct": null, "host_softmax_max_abs_dp": 2.9802322387695312e-08, "batch": { "call": 0, "index": 2, "B": 3 } } ], "response": { "answers": { "topic": { "type": "choice", "choice": "feedback", "confidence": 0.6271855235099792, "probabilities": { "delivery": 0.08343102037906647, "billing": 0.24607208371162415, "account": 0.043311357498168945, "feedback": 0.6271855235099792 } }, "wants": { "type": "noul", "noul": 0.9252889752388 }, "urgency": { "type": "score", "score": 1.364602506160736, "confidence": 0.5390024185180664, "probabilities": { "0": 0.17439991235733032, "1": 0.28659766912460327, "2": 0.5390024185180664 }, "legend": { "0": "Can wait", "1": "Soon", "2": "Right now" } } }, "usage": { "input_tokens": 121, "output_tokens": 0 } }, "oracle_file": "records_ref_audio.json" }, { "id": "aud_14", "mode": "audio", "source": "own_audio", "native": true, "prefix": 7, "usage_input_tokens": 142, "questions": [ { "qid": "topic", "type": "choice", "K": 4, "calibrate": false, "temperature_key": null, "T": null, "prefix": 7, "n_ids": 57, "positions": 64, "bucket": 256, "ids": [ 1, 17, 24373, 18, 3493, 856, 779, 21009, 10789, 1337, 540, 19, 16, 7832, 572, 1873, 535, 835, 11746, 933, 768, 15073, 8688, 20, 19, 16, 7832, 572, 7778, 535, 61958, 774, 810, 23325, 20, 19, 16, 7832, 572, 12992, 535, 21040, 3197, 20, 19, 16, 7832, 572, 15296, 535, 36843, 3321, 1337, 779, 3714, 20, 21 ], "markers": [ 12, 25, 36, 45 ], "logits_raw": [ -14.164146423339844, -10.380624771118164, -15.470945358276367, -2.3223929405212402 ], "probs_raw": [ 7.1953281803871505e-06, 0.0003163830260746181, 1.9476683519314975e-06, 0.9996744394302368 ], "probs": [ 7.1953281803871505e-06, 0.0003163830260746181, 1.9476683519314975e-06, 0.9996744394302368 ], "argmax_index": 3, "argmax": "feedback", "top2_margin": 0.9993580564041622, "near_tie": false, "gold": "billing", "correct": false, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 0, "B": 3 } }, { "qid": "wants", "type": "noul", "K": 2, "calibrate": false, "temperature_key": null, "T": null, "prefix": 7, "n_ids": 27, "positions": 34, "bucket": 256, "ids": [ 1, 17, 24373, 18, 4940, 779, 21009, 14065, 875, 3503, 811, 874, 4007, 540, 19, 16, 4831, 535, 1295, 20, 19, 16, 3347, 535, 18672, 20, 21 ], "markers": [ 15, 21 ], "logits_raw": [ -1.1729061603546143, -0.18968117237091064 ], "probs_raw": [ 0.27225232124328613, 0.7277476191520691 ], "probs": [ 0.7277476191520691, 0.27225232124328613 ], "argmax_index": 0, "argmax": "true", "top2_margin": 0.45549529790878296, "near_tie": false, "gold": "true", "correct": true, "host_softmax_max_abs_dp": 2.9802322387695312e-08, "batch": { "call": 0, "index": 2, "B": 3 } }, { "qid": "urgency", "type": "score", "K": 3, "calibrate": false, "temperature_key": null, "T": null, "prefix": 7, "n_ids": 37, "positions": 44, "bucket": 256, "ids": [ 1, 17, 24373, 18, 4023, 31149, 856, 779, 5940, 540, 19, 16, 2119, 730, 525, 535, 2577, 7498, 20, 19, 16, 2119, 730, 526, 535, 32238, 20, 19, 16, 2119, 730, 527, 535, 17231, 2166, 20, 21 ], "markers": [ 11, 20, 28 ], "logits_raw": [ -10.382744789123535, -6.77346658706665, -5.851489067077637 ], "probs_raw": [ 0.007644413039088249, 0.2823798954486847, 0.7099757194519043 ], "probs": [ 0.007644413039088249, 0.2823798954486847, 0.7099757194519043 ], "argmax_index": 2, "argmax": "2", "top2_margin": 0.4275958240032196, "near_tie": false, "gold": null, "correct": null, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 1, "B": 3 } } ], "response": { "answers": { "topic": { "type": "choice", "choice": "feedback", "confidence": 0.9996744394302368, "probabilities": { "delivery": 7.1953281803871505e-06, "billing": 0.0003163830260746181, "account": 1.9476683519314975e-06, "feedback": 0.9996744394302368 } }, "wants": { "type": "noul", "noul": 0.7277476191520691 }, "urgency": { "type": "score", "score": 1.7023313343524933, "confidence": 0.7099757194519043, "probabilities": { "0": 0.007644413039088249, "1": 0.2823798954486847, "2": 0.7099757194519043 }, "legend": { "0": "Can wait", "1": "Soon", "2": "Right now" } } }, "usage": { "input_tokens": 142, "output_tokens": 0 } }, "oracle_file": "records_ref_audio.json" }, { "id": "aud_14", "mode": "text", "source": "own_audio", "native": false, "prefix": 0, "usage_input_tokens": 121, "questions": [ { "qid": "topic", "type": "choice", "K": 4, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 49, "positions": 49, "bucket": 256, "ids": [ 1, 17, 18, 3493, 856, 779, 21009, 10789, 1337, 540, 19, 16, 11746, 535, 835, 11746, 933, 768, 15073, 8688, 20, 19, 16, 5363, 800, 535, 61958, 774, 810, 23325, 20, 19, 16, 3863, 535, 21040, 3197, 20, 19, 16, 12779, 535, 36843, 3321, 1337, 779, 3714, 20, 21 ], "markers": [ 11, 22, 32, 39 ], "logits_raw": [ -4.030765533447266, -2.516629695892334, -4.948546409606934, -1.2068599462509155 ], "probs_raw": [ 0.04388390854001045, 0.19947391748428345, 0.017527427524328232, 0.7391147613525391 ], "probs": [ 0.08343102037906647, 0.24607208371162415, 0.043311357498168945, 0.6271855235099792 ], "argmax_index": 3, "argmax": "feedback", "top2_margin": 0.3811134397983551, "near_tie": false, "gold": null, "correct": null, "host_softmax_max_abs_dp": 1.4901161193847656e-08, "batch": { "call": 0, "index": 0, "B": 3 } }, { "qid": "wants", "type": "noul", "K": 2, "calibrate": true, "temperature_key": "noul:2", "T": 1.6663223505020142, "prefix": 0, "n_ids": 36, "positions": 36, "bucket": 256, "ids": [ 1, 17, 18, 4940, 779, 21009, 14065, 875, 3503, 811, 874, 4007, 540, 19, 16, 4831, 535, 1295, 521, 779, 8294, 2177, 1014, 3965, 20, 19, 16, 3347, 535, 18672, 521, 779, 8294, 10995, 20, 21 ], "markers": [ 14, 26 ], "logits_raw": [ -0.19816279411315918, 3.9951019287109375 ], "probs_raw": [ 0.014872386120259762, 0.9851276278495789 ], "probs": [ 0.9252889752388, 0.07471100240945816 ], "argmax_index": 0, "argmax": "true", "top2_margin": 0.8505779728293419, "near_tie": false, "gold": null, "correct": null, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 1, "B": 3 } }, { "qid": "urgency", "type": "score", "K": 3, "calibrate": true, "temperature_key": "score:3-5", "T": 1.7301132678985596, "prefix": 0, "n_ids": 36, "positions": 36, "bucket": 256, "ids": [ 1, 17, 18, 4023, 31149, 856, 779, 5940, 540, 19, 16, 2119, 730, 525, 535, 2577, 7498, 20, 19, 16, 2119, 730, 526, 535, 32238, 20, 19, 16, 2119, 730, 527, 535, 17231, 2166, 20, 21 ], "markers": [ 10, 19, 27 ], "logits_raw": [ -7.304863452911377, -6.445466995239258, -5.352657318115234 ], "probs_raw": [ 0.09609891474246979, 0.22696013748645782, 0.6769409775733948 ], "probs": [ 0.17439991235733032, 0.28659766912460327, 0.5390024185180664 ], "argmax_index": 2, "argmax": "2", "top2_margin": 0.25240474939346313, "near_tie": false, "gold": null, "correct": null, "host_softmax_max_abs_dp": 2.9802322387695312e-08, "batch": { "call": 0, "index": 2, "B": 3 } } ], "response": { "answers": { "topic": { "type": "choice", "choice": "feedback", "confidence": 0.6271855235099792, "probabilities": { "delivery": 0.08343102037906647, "billing": 0.24607208371162415, "account": 0.043311357498168945, "feedback": 0.6271855235099792 } }, "wants": { "type": "noul", "noul": 0.9252889752388 }, "urgency": { "type": "score", "score": 1.364602506160736, "confidence": 0.5390024185180664, "probabilities": { "0": 0.17439991235733032, "1": 0.28659766912460327, "2": 0.5390024185180664 }, "legend": { "0": "Can wait", "1": "Soon", "2": "Right now" } } }, "usage": { "input_tokens": 121, "output_tokens": 0 } }, "oracle_file": "records_ref_audio.json" }, { "id": "aud_15", "mode": "audio", "source": "own_audio", "native": true, "prefix": 375, "usage_input_tokens": 1264, "questions": [ { "qid": "topic", "type": "choice", "K": 4, "calibrate": false, "temperature_key": null, "T": null, "prefix": 375, "n_ids": 63, "positions": 438, "bucket": 512, "ids": [ 1, 17, 27377, 1139, 6634, 988, 768, 5196, 523, 18, 3493, 856, 779, 21009, 10789, 1337, 540, 19, 16, 7832, 572, 1873, 535, 835, 11746, 933, 768, 15073, 8688, 20, 19, 16, 7832, 572, 7778, 535, 61958, 774, 810, 23325, 20, 19, 16, 7832, 572, 12992, 535, 21040, 3197, 20, 19, 16, 7832, 572, 15296, 535, 36843, 3321, 1337, 779, 3714, 20, 21 ], "markers": [ 18, 31, 42, 51 ], "logits_raw": [ -7.3204827308654785, -7.056801795959473, -11.541250228881836, -1.0992538928985596 ], "probs_raw": [ 0.001977700972929597, 0.0025743974838405848, 2.9047207135590725e-05, 0.995418906211853 ], "probs": [ 0.001977700972929597, 0.0025743974838405848, 2.9047207135590725e-05, 0.995418906211853 ], "argmax_index": 3, "argmax": "feedback", "top2_margin": 0.9928445087280124, "near_tie": false, "gold": "delivery", "correct": false, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 0, "B": 3 } }, { "qid": "wants", "type": "noul", "K": 2, "calibrate": false, "temperature_key": null, "T": null, "prefix": 375, "n_ids": 33, "positions": 408, "bucket": 512, "ids": [ 1, 17, 27377, 1139, 6634, 988, 768, 5196, 523, 18, 4940, 779, 21009, 14065, 875, 3503, 811, 874, 4007, 540, 19, 16, 4831, 535, 1295, 20, 19, 16, 3347, 535, 18672, 20, 21 ], "markers": [ 21, 27 ], "logits_raw": [ -1.6256054639816284, 0.10680553317070007 ], "probs_raw": [ 0.1502794474363327, 0.8497205972671509 ], "probs": [ 0.8497205972671509, 0.1502794474363327 ], "argmax_index": 0, "argmax": "true", "top2_margin": 0.6994411498308182, "near_tie": false, "gold": "true", "correct": true, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 2, "B": 3 } }, { "qid": "urgency", "type": "score", "K": 3, "calibrate": false, "temperature_key": null, "T": null, "prefix": 375, "n_ids": 43, "positions": 418, "bucket": 512, "ids": [ 1, 17, 27377, 1139, 6634, 988, 768, 5196, 523, 18, 4023, 31149, 856, 779, 5940, 540, 19, 16, 2119, 730, 525, 535, 2577, 7498, 20, 19, 16, 2119, 730, 526, 535, 32238, 20, 19, 16, 2119, 730, 527, 535, 17231, 2166, 20, 21 ], "markers": [ 17, 26, 34 ], "logits_raw": [ -12.868496894836426, -3.0056636333465576, 4.093652725219727 ], "probs_raw": [ 4.2960873969377644e-08, 0.0008249877719208598, 0.9991748929023743 ], "probs": [ 4.2960873969377644e-08, 0.0008249877719208598, 0.9991748929023743 ], "argmax_index": 2, "argmax": "2", "top2_margin": 0.9983499051304534, "near_tie": false, "gold": "2", "correct": true, "host_softmax_max_abs_dp": 5.820766091346741e-11, "batch": { "call": 0, "index": 1, "B": 3 } } ], "response": { "answers": { "topic": { "type": "choice", "choice": "feedback", "confidence": 0.995418906211853, "probabilities": { "delivery": 0.001977700972929597, "billing": 0.0025743974838405848, "account": 2.9047207135590725e-05, "feedback": 0.995418906211853 } }, "wants": { "type": "noul", "noul": 0.8497205972671509 }, "urgency": { "type": "score", "score": 1.9991747735766694, "confidence": 0.9991748929023743, "probabilities": { "0": 4.2960873969377644e-08, "1": 0.0008249877719208598, "2": 0.9991748929023743 }, "legend": { "0": "Can wait", "1": "Soon", "2": "Right now" } } }, "usage": { "input_tokens": 1264, "output_tokens": 0 } }, "oracle_file": "records_ref_audio.json" }, { "id": "aud_15", "mode": "text", "source": "own_audio", "native": false, "prefix": 0, "usage_input_tokens": 142, "questions": [ { "qid": "topic", "type": "choice", "K": 4, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 56, "positions": 56, "bucket": 256, "ids": [ 1, 17, 27377, 1139, 6634, 988, 768, 5196, 523, 18, 3493, 856, 779, 21009, 10789, 1337, 540, 19, 16, 11746, 535, 835, 11746, 933, 768, 15073, 8688, 20, 19, 16, 5363, 800, 535, 61958, 774, 810, 23325, 20, 19, 16, 3863, 535, 21040, 3197, 20, 19, 16, 12779, 535, 36843, 3321, 1337, 779, 3714, 20, 21 ], "markers": [ 18, 29, 39, 46 ], "logits_raw": [ -2.1100118160247803, -1.3073279857635498, -3.290978193283081, 1.7141824960708618 ], "probs_raw": [ 0.020269835367798805, 0.04523257911205292, 0.0062224725261330605, 0.9282751679420471 ], "probs": [ 0.05386649817228317, 0.0955735445022583, 0.023170966655015945, 0.8273889422416687 ], "argmax_index": 3, "argmax": "feedback", "top2_margin": 0.7318153977394104, "near_tie": false, "gold": null, "correct": null, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 0, "B": 3 } }, { "qid": "wants", "type": "noul", "K": 2, "calibrate": true, "temperature_key": "noul:2", "T": 1.6663223505020142, "prefix": 0, "n_ids": 43, "positions": 43, "bucket": 256, "ids": [ 1, 17, 27377, 1139, 6634, 988, 768, 5196, 523, 18, 4940, 779, 21009, 14065, 875, 3503, 811, 874, 4007, 540, 19, 16, 4831, 535, 1295, 521, 779, 8294, 2177, 1014, 3965, 20, 19, 16, 3347, 535, 18672, 521, 779, 8294, 10995, 20, 21 ], "markers": [ 21, 33 ], "logits_raw": [ 2.693018674850464, -4.27215051651001 ], "probs_raw": [ 0.9990566372871399, 0.0009433126542717218 ], "probs": [ 0.015068508684635162, 0.984931468963623 ], "argmax_index": 1, "argmax": "false", "top2_margin": 0.9698629602789879, "near_tie": false, "gold": null, "correct": null, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 1, "B": 3 } }, { "qid": "urgency", "type": "score", "K": 3, "calibrate": true, "temperature_key": "score:3-5", "T": 1.7301132678985596, "prefix": 0, "n_ids": 43, "positions": 43, "bucket": 256, "ids": [ 1, 17, 27377, 1139, 6634, 988, 768, 5196, 523, 18, 4023, 31149, 856, 779, 5940, 540, 19, 16, 2119, 730, 525, 535, 2577, 7498, 20, 19, 16, 2119, 730, 526, 535, 32238, 20, 19, 16, 2119, 730, 527, 535, 17231, 2166, 20, 21 ], "markers": [ 17, 26, 34 ], "logits_raw": [ -8.044438362121582, -6.318297863006592, -8.428503036499023 ], "probs_raw": [ 0.13698606193065643, 0.7697144746780396, 0.09329947084188461 ], "probs": [ 0.22158345580101013, 0.6009449362754822, 0.17747168242931366 ], "argmax_index": 1, "argmax": "1", "top2_margin": 0.37936148047447205, "near_tie": false, "gold": null, "correct": null, "host_softmax_max_abs_dp": 1.4901161193847656e-08, "batch": { "call": 0, "index": 2, "B": 3 } } ], "response": { "answers": { "topic": { "type": "choice", "choice": "feedback", "confidence": 0.8273889422416687, "probabilities": { "delivery": 0.05386649817228317, "billing": 0.0955735445022583, "account": 0.023170966655015945, "feedback": 0.8273889422416687 } }, "wants": { "type": "noul", "noul": 0.015068508684635162 }, "urgency": { "type": "score", "score": 0.9558883011341095, "confidence": 0.6009449362754822, "probabilities": { "0": 0.22158345580101013, "1": 0.6009449362754822, "2": 0.17747168242931366 }, "legend": { "0": "Can wait", "1": "Soon", "2": "Right now" } } }, "usage": { "input_tokens": 142, "output_tokens": 0 } }, "oracle_file": "records_ref_audio.json" }, { "id": "card_batch_00", "mode": "text", "source": "card", "native": true, "prefix": 0, "usage_input_tokens": 53, "questions": [ { "qid": "team", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 53, "positions": 53, "bucket": 256, "ids": [ 1, 17, 15090, 856, 1727, 61026, 540, 1311, 953, 2888, 17980, 523, 18, 28828, 3410, 1753, 9936, 1033, 540, 19, 16, 5363, 800, 535, 61958, 774, 521, 53467, 592, 521, 60827, 1864, 20, 19, 16, 10017, 535, 5135, 933, 3475, 50022, 20, 19, 16, 24744, 535, 11631, 7937, 59006, 2851, 1517, 20, 21 ], "markers": [ 20, 34, 43 ], "logits_raw": [ -4.570059776306152, -1.822887659072876, -7.762394428253174 ], "probs_raw": [ 0.060097839683294296, 0.937433660030365, 0.0024685710668563843 ], "probs": [ 0.12167388945817947, 0.865885853767395, 0.012440279126167297 ], "argmax_index": 1, "argmax": "technical", "top2_margin": 0.7442119643092155, "near_tie": false, "gold": null, "correct": null, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "team": { "type": "choice", "choice": "technical", "confidence": 0.865885853767395, "probabilities": { "billing": 0.12167388945817947, "technical": 0.865885853767395, "fraud": 0.012440279126167297 } } }, "usage": { "input_tokens": 53, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "card_batch_01", "mode": "text", "source": "card", "native": true, "prefix": 0, "usage_input_tokens": 51, "questions": [ { "qid": "team", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 51, "positions": 51, "bucket": 256, "ids": [ 1, 17, 1098, 1495, 49085, 1403, 859, 2714, 11063, 523, 18, 28828, 3410, 1753, 9936, 1033, 540, 19, 16, 5363, 800, 535, 61958, 774, 521, 53467, 592, 521, 60827, 1864, 20, 19, 16, 10017, 535, 5135, 933, 3475, 50022, 20, 19, 16, 24744, 535, 11631, 7937, 59006, 2851, 1517, 20, 21 ], "markers": [ 18, 32, 41 ], "logits_raw": [ -4.752628326416016, 4.058452606201172, -4.859061241149902 ], "probs_raw": [ 0.00014902983093634248, 0.9997169375419617, 0.00013398309238255024 ], "probs": [ 0.0018406440503895283, 0.9964535236358643, 0.0017058887751773 ], "argmax_index": 1, "argmax": "technical", "top2_margin": 0.9946128795854747, "near_tie": false, "gold": null, "correct": null, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "team": { "type": "choice", "choice": "technical", "confidence": 0.9964535236358643, "probabilities": { "billing": 0.0018406440503895283, "technical": 0.9964535236358643, "fraud": 0.0017058887751773 } } }, "usage": { "input_tokens": 51, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "card_text", "mode": "text", "source": "card", "native": true, "prefix": 0, "usage_input_tokens": 154, "questions": [ { "qid": "refund", "type": "noul", "K": 2, "calibrate": true, "temperature_key": "noul:2", "T": 1.6663223505020142, "prefix": 0, "n_ids": 47, "positions": 47, "bucket": 256, "ids": [ 1, 17, 550, 953, 15418, 11126, 1033, 5805, 521, 8694, 53467, 1235, 803, 1377, 523, 18, 4940, 779, 13608, 14065, 875, 768, 53467, 540, 19, 16, 4831, 535, 1295, 521, 779, 8294, 2177, 1014, 3965, 20, 19, 16, 3347, 535, 18672, 521, 779, 8294, 10995, 20, 21 ], "markers": [ 25, 37 ], "logits_raw": [ -2.0755720138549805, 8.723648071289062 ], "probs_raw": [ 2.0415003746165894e-05, 0.9999796152114868 ], "probs": [ 0.998469889163971, 0.001530130160972476 ], "argmax_index": 0, "argmax": "true", "top2_margin": 0.9969397590029985, "near_tie": false, "gold": null, "correct": null, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 2, "B": 3 } }, { "qid": "team", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 56, "positions": 56, "bucket": 256, "ids": [ 1, 17, 550, 953, 15418, 11126, 1033, 5805, 521, 8694, 53467, 1235, 803, 1377, 523, 18, 28828, 3410, 1753, 9936, 1033, 540, 19, 16, 5363, 800, 535, 61958, 774, 521, 53467, 592, 521, 60827, 1864, 20, 19, 16, 10017, 535, 5135, 933, 3475, 50022, 20, 19, 16, 24744, 535, 11631, 7937, 59006, 2851, 1517, 20, 21 ], "markers": [ 23, 37, 46 ], "logits_raw": [ 6.0775146484375, -1.2129172086715698, -2.832930088043213 ], "probs_raw": [ 0.999183714389801, 0.0006814766093157232, 0.00013486169336829334 ], "probs": [ 0.9928572177886963, 0.005434439517557621, 0.0017083368729799986 ], "argmax_index": 0, "argmax": "billing", "top2_margin": 0.9874227782711387, "near_tie": false, "gold": null, "correct": null, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 0, "B": 3 } }, { "qid": "urgency", "type": "score", "K": 3, "calibrate": true, "temperature_key": "score:3-5", "T": 1.7301132678985596, "prefix": 0, "n_ids": 51, "positions": 51, "bucket": 256, "ids": [ 1, 17, 550, 953, 15418, 11126, 1033, 5805, 521, 8694, 53467, 1235, 803, 1377, 523, 18, 4023, 31149, 856, 1033, 540, 19, 16, 2119, 730, 525, 535, 2577, 7498, 20, 19, 16, 2119, 730, 526, 535, 11268, 20, 19, 16, 2119, 730, 527, 535, 18578, 800, 779, 13608, 2166, 20, 21 ], "markers": [ 22, 31, 39 ], "logits_raw": [ -12.788542747497559, -6.237281322479248, -3.7804148197174072 ], "probs_raw": [ 0.00011273521522525698, 0.07892896980047226, 0.9209583401679993 ], "probs": [ 0.004393829498440027, 0.19379673898220062, 0.8018094301223755 ], "argmax_index": 2, "argmax": "2", "top2_margin": 0.6080126911401749, "near_tie": false, "gold": null, "correct": null, "host_softmax_max_abs_dp": 4.656612873077393e-10, "batch": { "call": 0, "index": 1, "B": 3 } } ], "response": { "answers": { "refund": { "type": "noul", "noul": 0.998469889163971 }, "team": { "type": "choice", "choice": "billing", "confidence": 0.9928572177886963, "probabilities": { "billing": 0.9928572177886963, "technical": 0.005434439517557621, "fraud": 0.0017083368729799986 } }, "urgency": { "type": "score", "score": 1.7974155992269516, "confidence": 0.8018094301223755, "probabilities": { "0": 0.004393829498440027, "1": 0.19379673898220062, "2": 0.8018094301223755 }, "legend": { "0": "Can wait", "1": "Today", "2": "Blocking the customer now" } } }, "usage": { "input_tokens": 154, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "img_01", "mode": "image", "source": "own_image", "native": true, "prefix": 144, "usage_input_tokens": 530, "questions": [ { "qid": "room", "type": "choice", "K": 4, "calibrate": false, "temperature_key": null, "T": null, "prefix": 144, "n_ids": 39, "positions": 183, "bucket": 256, "ids": [ 1, 17, 18, 3493, 3746, 803, 5362, 856, 1033, 540, 19, 16, 15936, 535, 835, 15936, 20, 19, 16, 32874, 535, 835, 32874, 20, 19, 16, 6756, 535, 1592, 6756, 20, 19, 16, 32062, 535, 835, 32062, 20, 21 ], "markers": [ 11, 18, 25, 32 ], "logits_raw": [ -13.372673034667969, 2.3188087940216064, -10.872926712036133, -5.046749114990234 ], "probs_raw": [ 1.5310895662423718e-07, 0.9993657469749451, 1.8647758679435356e-06, 0.0006322711706161499 ], "probs": [ 1.5310895662423718e-07, 0.9993657469749451, 1.8647758679435356e-06, 0.0006322711706161499 ], "argmax_index": 1, "argmax": "bedroom", "top2_margin": 0.9987334758043289, "near_tie": false, "gold": "bedroom", "correct": true, "host_softmax_max_abs_dp": 5.820766091346741e-11, "batch": { "call": 0, "index": 0, "B": 3 } }, { "qid": "suitcase", "type": "noul", "K": 2, "calibrate": false, "temperature_key": null, "T": null, "prefix": 144, "n_ids": 25, "positions": 169, "bucket": 256, "ids": [ 1, 17, 18, 4940, 1413, 768, 14503, 10511, 797, 779, 5362, 540, 19, 16, 4831, 535, 1295, 20, 19, 16, 3347, 535, 18672, 20, 21 ], "markers": [ 13, 19 ], "logits_raw": [ -3.17026948928833, 11.013753890991211 ], "probs_raw": [ 6.91762181759259e-07, 0.9999992847442627 ], "probs": [ 0.9999992847442627, 6.91762181759259e-07 ], "argmax_index": 0, "argmax": "true", "top2_margin": 0.9999985929820809, "near_tie": false, "gold": "true", "correct": true, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 2, "B": 3 } }, { "qid": "brightness", "type": "score", "K": 3, "calibrate": false, "temperature_key": null, "T": null, "prefix": 144, "n_ids": 34, "positions": 178, "bucket": 256, "ids": [ 1, 17, 18, 4023, 9899, 856, 779, 5362, 540, 19, 16, 2119, 730, 525, 535, 18003, 20, 19, 16, 2119, 730, 526, 535, 19640, 20, 19, 16, 2119, 730, 527, 535, 32989, 20, 21 ], "markers": [ 10, 18, 26 ], "logits_raw": [ -0.548412561416626, -6.016256332397461, -5.538527011871338 ], "probs_raw": [ 0.9890949130058289, 0.004174298141151667, 0.006730679422616959 ], "probs": [ 0.9890949130058289, 0.004174298141151667, 0.006730679422616959 ], "argmax_index": 0, "argmax": "0", "top2_margin": 0.9823642335832119, "near_tie": false, "gold": "2", "correct": false, "host_softmax_max_abs_dp": 4.656612873077393e-10, "batch": { "call": 0, "index": 1, "B": 3 } } ], "response": { "answers": { "room": { "type": "choice", "choice": "bedroom", "confidence": 0.9993657469749451, "probabilities": { "kitchen": 1.5310895662423718e-07, "bedroom": 0.9993657469749451, "office": 1.8647758679435356e-06, "bathroom": 0.0006322711706161499 } }, "suitcase": { "type": "noul", "noul": 0.9999992847442627 }, "brightness": { "type": "score", "score": 0.017635656986385584, "confidence": 0.9890949130058289, "probabilities": { "0": 0.9890949130058289, "1": 0.004174298141151667, "2": 0.006730679422616959 }, "legend": { "0": "Dark", "1": "Dim", "2": "Bright" } } }, "usage": { "input_tokens": 530, "output_tokens": 0 } }, "oracle_file": "records_ref_images.json" }, { "id": "img_01", "mode": "text", "source": "own_image", "native": false, "prefix": 0, "usage_input_tokens": 108, "questions": [ { "qid": "room", "type": "choice", "K": 4, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 39, "positions": 39, "bucket": 256, "ids": [ 1, 17, 18, 3493, 3746, 803, 5362, 856, 1033, 540, 19, 16, 15936, 535, 835, 15936, 20, 19, 16, 32874, 535, 835, 32874, 20, 19, 16, 6756, 535, 1592, 6756, 20, 19, 16, 32062, 535, 835, 32062, 20, 21 ], "markers": [ 11, 18, 25, 32 ], "logits_raw": [ -0.9235021471977234, 1.1862629652023315, -0.5316015481948853, 0.6049119234085083 ], "probs_raw": [ 0.0652020052075386, 0.5376754999160767, 0.09648530185222626, 0.3006371855735779 ], "probs": [ 0.1018717959523201, 0.4598037302494049, 0.13478277623653412, 0.3035416603088379 ], "argmax_index": 1, "argmax": "bedroom", "top2_margin": 0.15626206994056702, "near_tie": false, "gold": null, "correct": null, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 0, "B": 3 } }, { "qid": "suitcase", "type": "noul", "K": 2, "calibrate": true, "temperature_key": "noul:2", "T": 1.6663223505020142, "prefix": 0, "n_ids": 35, "positions": 35, "bucket": 256, "ids": [ 1, 17, 18, 4940, 1413, 768, 14503, 10511, 797, 779, 5362, 540, 19, 16, 4831, 535, 1295, 521, 779, 8294, 2177, 1014, 3965, 20, 19, 16, 3347, 535, 18672, 521, 779, 8294, 10995, 20, 21 ], "markers": [ 13, 25 ], "logits_raw": [ -0.22425425052642822, 1.6848607063293457 ], "probs_raw": [ 0.12908032536506653, 0.8709197044372559 ], "probs": [ 0.7587257027626038, 0.24127434194087982 ], "argmax_index": 0, "argmax": "true", "top2_margin": 0.5174513608217239, "near_tie": false, "gold": null, "correct": null, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 1, "B": 3 } }, { "qid": "brightness", "type": "score", "K": 3, "calibrate": true, "temperature_key": "score:3-5", "T": 1.7301132678985596, "prefix": 0, "n_ids": 34, "positions": 34, "bucket": 256, "ids": [ 1, 17, 18, 4023, 9899, 856, 779, 5362, 540, 19, 16, 2119, 730, 525, 535, 18003, 20, 19, 16, 2119, 730, 526, 535, 19640, 20, 19, 16, 2119, 730, 527, 535, 32989, 20, 21 ], "markers": [ 10, 18, 26 ], "logits_raw": [ -2.512366533279419, -0.007647328078746796, 1.206966519355774 ], "probs_raw": [ 0.01835636794567108, 0.2246842086315155, 0.7569594383239746 ], "probs": [ 0.07227402925491333, 0.30741074681282043, 0.6203151941299438 ], "argmax_index": 2, "argmax": "2", "top2_margin": 0.3129044473171234, "near_tie": false, "gold": null, "correct": null, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 2, "B": 3 } } ], "response": { "answers": { "room": { "type": "choice", "choice": "bedroom", "confidence": 0.4598037302494049, "probabilities": { "kitchen": 0.1018717959523201, "bedroom": 0.4598037302494049, "office": 0.13478277623653412, "bathroom": 0.3035416603088379 } }, "suitcase": { "type": "noul", "noul": 0.7587257027626038 }, "brightness": { "type": "score", "score": 1.5480411350727081, "confidence": 0.6203151941299438, "probabilities": { "0": 0.07227402925491333, "1": 0.30741074681282043, "2": 0.6203151941299438 }, "legend": { "0": "Dark", "1": "Dim", "2": "Bright" } } }, "usage": { "input_tokens": 108, "output_tokens": 0 } }, "oracle_file": "records_ref_images.json" }, { "id": "img_02", "mode": "image", "source": "own_image", "native": true, "prefix": 108, "usage_input_tokens": 437, "questions": [ { "qid": "room", "type": "choice", "K": 4, "calibrate": false, "temperature_key": null, "T": null, "prefix": 108, "n_ids": 45, "positions": 153, "bucket": 256, "ids": [ 1, 17, 26252, 988, 768, 5362, 3999, 523, 18, 3493, 3746, 803, 5362, 856, 1033, 540, 19, 16, 15936, 535, 835, 15936, 20, 19, 16, 32874, 535, 835, 32874, 20, 19, 16, 6756, 535, 1592, 6756, 20, 19, 16, 32062, 535, 835, 32062, 20, 21 ], "markers": [ 17, 24, 31, 38 ], "logits_raw": [ -13.370351791381836, 4.531220436096191, -10.323932647705078, -0.2855755388736725 ], "probs_raw": [ 1.6670359315185124e-08, 0.9919719696044922, 3.5074251059086237e-07, 0.008027708157896996 ], "probs": [ 1.6670359315185124e-08, 0.9919719696044922, 3.5074251059086237e-07, 0.008027708157896996 ], "argmax_index": 1, "argmax": "bedroom", "top2_margin": 0.9839442614465952, "near_tie": false, "gold": "bedroom", "correct": true, "host_softmax_max_abs_dp": 2.842170943040401e-14, "batch": { "call": 0, "index": 0, "B": 3 } }, { "qid": "lamp", "type": "noul", "K": 2, "calibrate": false, "temperature_key": null, "T": null, "prefix": 108, "n_ids": 28, "positions": 136, "bucket": 256, "ids": [ 1, 17, 26252, 988, 768, 5362, 3999, 523, 18, 4940, 768, 29114, 35589, 884, 540, 19, 16, 4831, 535, 1295, 20, 19, 16, 3347, 535, 18672, 20, 21 ], "markers": [ 16, 22 ], "logits_raw": [ 1.247429609298706, -0.5311287641525269 ], "probs_raw": [ 0.8555187582969666, 0.14448124170303345 ], "probs": [ 0.14448124170303345, 0.8555187582969666 ], "argmax_index": 1, "argmax": "false", "top2_margin": 0.7110375165939331, "near_tie": false, "gold": "true", "correct": false, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 2, "B": 3 } }, { "qid": "brightness", "type": "score", "K": 3, "calibrate": false, "temperature_key": null, "T": null, "prefix": 108, "n_ids": 40, "positions": 148, "bucket": 256, "ids": [ 1, 17, 26252, 988, 768, 5362, 3999, 523, 18, 4023, 9899, 856, 779, 5362, 540, 19, 16, 2119, 730, 525, 535, 18003, 20, 19, 16, 2119, 730, 526, 535, 19640, 20, 19, 16, 2119, 730, 527, 535, 32989, 20, 21 ], "markers": [ 16, 24, 32 ], "logits_raw": [ -2.5174143314361572, -0.759251594543457, -3.9761927127838135 ], "probs_raw": [ 0.1421608030796051, 0.824783980846405, 0.033055249601602554 ], "probs": [ 0.1421608030796051, 0.824783980846405, 0.033055249601602554 ], "argmax_index": 1, "argmax": "1", "top2_margin": 0.6826231777667999, "near_tie": false, "gold": "1", "correct": true, "host_softmax_max_abs_dp": 1.4901161193847656e-08, "batch": { "call": 0, "index": 1, "B": 3 } } ], "response": { "answers": { "room": { "type": "choice", "choice": "bedroom", "confidence": 0.9919719696044922, "probabilities": { "kitchen": 1.6670359315185124e-08, "bedroom": 0.9919719696044922, "office": 3.5074251059086237e-07, "bathroom": 0.008027708157896996 } }, "lamp": { "type": "noul", "noul": 0.14448124170303345 }, "brightness": { "type": "score", "score": 0.8908944800496101, "confidence": 0.824783980846405, "probabilities": { "0": 0.1421608030796051, "1": 0.824783980846405, "2": 0.033055249601602554 }, "legend": { "0": "Dark", "1": "Dim", "2": "Bright" } } }, "usage": { "input_tokens": 437, "output_tokens": 0 } }, "oracle_file": "records_ref_images.json" }, { "id": "img_02", "mode": "text", "source": "own_image", "native": false, "prefix": 0, "usage_input_tokens": 123, "questions": [ { "qid": "room", "type": "choice", "K": 4, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 45, "positions": 45, "bucket": 256, "ids": [ 1, 17, 26252, 988, 768, 5362, 3999, 523, 18, 3493, 3746, 803, 5362, 856, 1033, 540, 19, 16, 15936, 535, 835, 15936, 20, 19, 16, 32874, 535, 835, 32874, 20, 19, 16, 6756, 535, 1592, 6756, 20, 19, 16, 32062, 535, 835, 32062, 20, 21 ], "markers": [ 17, 24, 31, 38 ], "logits_raw": [ -2.2342491149902344, 1.2450541257858276, -1.7813855409622192, -1.1000628471374512 ], "probs_raw": [ 0.026233933866024017, 0.8509531617164612, 0.041260987520217896, 0.0815519317984581 ], "probs": [ 0.0601099468767643, 0.7216733694076538, 0.0830691009759903, 0.1351475715637207 ], "argmax_index": 1, "argmax": "bedroom", "top2_margin": 0.5865257978439331, "near_tie": false, "gold": null, "correct": null, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 0, "B": 3 } }, { "qid": "lamp", "type": "noul", "K": 2, "calibrate": true, "temperature_key": "noul:2", "T": 1.6663223505020142, "prefix": 0, "n_ids": 38, "positions": 38, "bucket": 256, "ids": [ 1, 17, 26252, 988, 768, 5362, 3999, 523, 18, 4940, 768, 29114, 35589, 884, 540, 19, 16, 4831, 535, 1295, 521, 779, 8294, 2177, 1014, 3965, 20, 19, 16, 3347, 535, 18672, 521, 779, 8294, 10995, 20, 21 ], "markers": [ 16, 28 ], "logits_raw": [ 1.744201421737671, -1.7224774360656738 ], "probs_raw": [ 0.9697246551513672, 0.030275333672761917 ], "probs": [ 0.11101281642913818, 0.8889872431755066 ], "argmax_index": 1, "argmax": "false", "top2_margin": 0.7779744267463684, "near_tie": false, "gold": null, "correct": null, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 2, "B": 3 } }, { "qid": "brightness", "type": "score", "K": 3, "calibrate": true, "temperature_key": "score:3-5", "T": 1.7301132678985596, "prefix": 0, "n_ids": 40, "positions": 40, "bucket": 256, "ids": [ 1, 17, 26252, 988, 768, 5362, 3999, 523, 18, 4023, 9899, 856, 779, 5362, 540, 19, 16, 2119, 730, 525, 535, 18003, 20, 19, 16, 2119, 730, 526, 535, 19640, 20, 19, 16, 2119, 730, 527, 535, 32989, 20, 21 ], "markers": [ 16, 24, 32 ], "logits_raw": [ -1.4011543989181519, 0.5800229907035828, 0.1053333580493927 ], "probs_raw": [ 0.0783567875623703, 0.5681866407394409, 0.3534564673900604 ], "probs": [ 0.15310442447662354, 0.48117661476135254, 0.3657189905643463 ], "argmax_index": 1, "argmax": "1", "top2_margin": 0.11545762419700623, "near_tie": false, "gold": null, "correct": null, "host_softmax_max_abs_dp": 2.9802322387695312e-08, "batch": { "call": 0, "index": 1, "B": 3 } } ], "response": { "answers": { "room": { "type": "choice", "choice": "bedroom", "confidence": 0.7216733694076538, "probabilities": { "kitchen": 0.0601099468767643, "bedroom": 0.7216733694076538, "office": 0.0830691009759903, "bathroom": 0.1351475715637207 } }, "lamp": { "type": "noul", "noul": 0.11101281642913818 }, "brightness": { "type": "score", "score": 1.2126145958900452, "confidence": 0.48117661476135254, "probabilities": { "0": 0.15310442447662354, "1": 0.48117661476135254, "2": 0.3657189905643463 }, "legend": { "0": "Dark", "1": "Dim", "2": "Bright" } } }, "usage": { "input_tokens": 123, "output_tokens": 0 } }, "oracle_file": "records_ref_images.json" }, { "id": "img_03", "mode": "image", "source": "own_image", "native": true, "prefix": 1770, "usage_input_tokens": 5411, "questions": [ { "qid": "beds", "type": "choice", "K": 3, "calibrate": false, "temperature_key": null, "T": null, "prefix": 1770, "n_ids": 35, "positions": 1805, "bucket": 2048, "ids": [ 1, 17, 18, 4023, 1697, 25893, 938, 797, 779, 5362, 540, 19, 16, 1235, 535, 3732, 6100, 20, 19, 16, 1547, 535, 8726, 25893, 20, 19, 16, 1209, 535, 13016, 933, 1209, 25893, 20, 21 ], "markers": [ 12, 19, 26 ], "logits_raw": [ 1.1830202341079712, 3.3079018592834473, -7.577269554138184 ], "probs_raw": [ 0.10670007765293121, 0.8932831883430481, 1.673474434937816e-05 ], "probs": [ 0.10670007765293121, 0.8932831883430481, 1.673474434937816e-05 ], "argmax_index": 1, "argmax": "two", "top2_margin": 0.7865831106901169, "near_tie": false, "gold": "two", "correct": true, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 1, "B": 3 } }, { "qid": "towel", "type": "noul", "K": 2, "calibrate": false, "temperature_key": null, "T": null, "prefix": 1770, "n_ids": 25, "positions": 1795, "bucket": 2048, "ids": [ 1, 17, 18, 4940, 1413, 768, 41478, 50624, 884, 768, 6100, 540, 19, 16, 4831, 535, 1295, 20, 19, 16, 3347, 535, 18672, 20, 21 ], "markers": [ 13, 19 ], "logits_raw": [ 3.5223562717437744, -2.9612913131713867 ], "probs_raw": [ 0.9984740614891052, 0.0015258945059031248 ], "probs": [ 0.0015258945059031248, 0.9984740614891052 ], "argmax_index": 1, "argmax": "false", "top2_margin": 0.9969481669832021, "near_tie": false, "gold": "true", "correct": false, "host_softmax_max_abs_dp": 1.1641532182693481e-10, "batch": { "call": 0, "index": 2, "B": 3 } }, { "qid": "tidy", "type": "score", "K": 3, "calibrate": false, "temperature_key": null, "T": null, "prefix": 1770, "n_ids": 41, "positions": 1811, "bucket": 2048, "ids": [ 1, 17, 18, 4023, 18584, 598, 856, 779, 5362, 540, 19, 16, 2119, 730, 525, 535, 25913, 1064, 11119, 791, 20, 19, 16, 2119, 730, 526, 535, 4260, 1064, 11119, 20, 19, 16, 2119, 730, 527, 535, 837, 48307, 20, 21 ], "markers": [ 11, 22, 32 ], "logits_raw": [ -4.784424304962158, -1.4155879020690918, 3.0489602088928223 ], "probs_raw": [ 0.00039161930908448994, 0.01137446891516447, 0.9882339239120483 ], "probs": [ 0.00039161930908448994, 0.01137446891516447, 0.9882339239120483 ], "argmax_index": 2, "argmax": "2", "top2_margin": 0.9768594549968839, "near_tie": false, "gold": "2", "correct": true, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 0, "B": 3 } } ], "response": { "answers": { "beds": { "type": "choice", "choice": "two", "confidence": 0.8932831883430481, "probabilities": { "one": 0.10670007765293121, "two": 0.8932831883430481, "more": 1.673474434937816e-05 } }, "towel": { "type": "noul", "noul": 0.0015258945059031248 }, "tidy": { "type": "score", "score": 1.9878423167392612, "confidence": 0.9882339239120483, "probabilities": { "0": 0.00039161930908448994, "1": 0.01137446891516447, "2": 0.9882339239120483 }, "legend": { "0": "Very cluttered", "1": "Some clutter", "2": "Tidy" } } }, "usage": { "input_tokens": 5411, "output_tokens": 0 } }, "oracle_file": "records_ref_images.json" }, { "id": "img_03", "mode": "text", "source": "own_image", "native": false, "prefix": 0, "usage_input_tokens": 111, "questions": [ { "qid": "beds", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 35, "positions": 35, "bucket": 256, "ids": [ 1, 17, 18, 4023, 1697, 25893, 938, 797, 779, 5362, 540, 19, 16, 1235, 535, 3732, 6100, 20, 19, 16, 1547, 535, 8726, 25893, 20, 19, 16, 1209, 535, 13016, 933, 1209, 25893, 20, 21 ], "markers": [ 12, 19, 26 ], "logits_raw": [ -0.25700244307518005, 0.2678792476654053, 1.0923843383789062 ], "probs_raw": [ 0.1527809202671051, 0.2582393288612366, 0.5889797806739807 ], "probs": [ 0.19697225093841553, 0.2865765392780304, 0.5164511799812317 ], "argmax_index": 2, "argmax": "more", "top2_margin": 0.2298746407032013, "near_tie": false, "gold": null, "correct": null, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 1, "B": 3 } }, { "qid": "towel", "type": "noul", "K": 2, "calibrate": true, "temperature_key": "noul:2", "T": 1.6663223505020142, "prefix": 0, "n_ids": 35, "positions": 35, "bucket": 256, "ids": [ 1, 17, 18, 4940, 1413, 768, 41478, 50624, 884, 768, 6100, 540, 19, 16, 4831, 535, 1295, 521, 779, 8294, 2177, 1014, 3965, 20, 19, 16, 3347, 535, 18672, 521, 779, 8294, 10995, 20, 21 ], "markers": [ 13, 25 ], "logits_raw": [ -0.3175913095474243, 7.188986301422119 ], "probs_raw": [ 0.0005491566262207925, 0.9994508624076843 ], "probs": [ 0.9890658855438232, 0.0109340725466609 ], "argmax_index": 0, "argmax": "true", "top2_margin": 0.9781318129971623, "near_tie": false, "gold": null, "correct": null, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 2, "B": 3 } }, { "qid": "tidy", "type": "score", "K": 3, "calibrate": true, "temperature_key": "score:3-5", "T": 1.7301132678985596, "prefix": 0, "n_ids": 41, "positions": 41, "bucket": 256, "ids": [ 1, 17, 18, 4023, 18584, 598, 856, 779, 5362, 540, 19, 16, 2119, 730, 525, 535, 25913, 1064, 11119, 791, 20, 19, 16, 2119, 730, 526, 535, 4260, 1064, 11119, 20, 19, 16, 2119, 730, 527, 535, 837, 48307, 20, 21 ], "markers": [ 11, 22, 32 ], "logits_raw": [ -3.0232226848602295, -0.7932671308517456, 3.672762393951416 ], "probs_raw": [ 0.0012203309452161193, 0.011348405852913857, 0.9874312877655029 ], "probs": [ 0.01901734061539173, 0.06901048868894577, 0.911972165107727 ], "argmax_index": 2, "argmax": "2", "top2_margin": 0.8429616764187813, "near_tie": false, "gold": null, "correct": null, "host_softmax_max_abs_dp": 1.862645149230957e-09, "batch": { "call": 0, "index": 0, "B": 3 } } ], "response": { "answers": { "beds": { "type": "choice", "choice": "more", "confidence": 0.5164511799812317, "probabilities": { "one": 0.19697225093841553, "two": 0.2865765392780304, "more": 0.5164511799812317 } }, "towel": { "type": "noul", "noul": 0.9890658855438232 }, "tidy": { "type": "score", "score": 1.8929548189043999, "confidence": 0.911972165107727, "probabilities": { "0": 0.01901734061539173, "1": 0.06901048868894577, "2": 0.911972165107727 }, "legend": { "0": "Very cluttered", "1": "Some clutter", "2": "Tidy" } } }, "usage": { "input_tokens": 111, "output_tokens": 0 } }, "oracle_file": "records_ref_images.json" }, { "id": "long_3400", "mode": "text", "source": "own_long_extended", "native": true, "prefix": 0, "usage_input_tokens": 13867, "questions": [ { "qid": "component", "type": "choice", "K": 5, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 3467, "positions": 3467, "bucket": 4096, "ids": [ 1, 17, 15991, 23707, 3613, 4425, 4857, 561, 33371, 12292, 38620, 1141, 61026, 17824, 730, 527, 521, 5008, 1985, 2647, 8411, 886, 826, 526, 521, 22694, 3446, 843, 1513, 31225, 1015, 27464, 535, 2960, 843, 6468, 535, 1056, 522, 1536, 535, 1056, 518, 767, 52574, 884, 15774, 535, 1267, 868, 2290, 767, 12003, 884, 1868, 535, 4582, 1055, 509, 6468, 535, 1056, 535, 1056, 7174, 17261, 730, 63688, 730, 10426, 527, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 526, 12079, 6468, 535, 6135, 535, 1056, 7174, 17261, 730, 63688, 730, 11233, 533, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 528, 12079, 6468, 535, 6468, 535, 1056, 7174, 17261, 730, 63688, 730, 10556, 526, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 530, 12079, 6468, 535, 4905, 535, 1056, 7174, 17261, 730, 63688, 730, 8380, 529, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 532, 12079, 6468, 535, 4905, 535, 4536, 7174, 17261, 730, 49574, 886, 889, 529, 41172, 3999, 7410, 843, 22500, 730, 525, 523, 3188, 8061, 1015, 6468, 535, 1374, 535, 1056, 7174, 17261, 730, 63688, 730, 10173, 534, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 526, 12079, 6468, 535, 1374, 535, 1417, 923, 32535, 730, 49883, 837, 527, 1295, 49321, 884, 61026, 730, 33118, 3375, 536, 61026, 3567, 811, 16687, 43497, 708, 6468, 535, 1386, 535, 1056, 7174, 17261, 730, 63688, 730, 10520, 525, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 528, 12079, 6468, 535, 1133, 535, 1056, 7174, 17261, 730, 63688, 730, 10426, 531, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 530, 12079, 6468, 535, 2197, 535, 1056, 7174, 17261, 730, 63688, 730, 8380, 527, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 532, 12079, 6468, 535, 1826, 535, 1056, 7174, 17261, 730, 63688, 730, 11233, 530, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 526, 12079, 6468, 535, 2142, 535, 1056, 7174, 17261, 730, 63688, 730, 10520, 532, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 528, 12079, 6468, 535, 2142, 535, 4536, 7174, 17261, 730, 49574, 886, 889, 529, 41172, 3999, 7410, 843, 22500, 730, 525, 523, 5489, 8061, 1015, 6468, 535, 1664, 535, 1056, 7174, 17261, 730, 63688, 730, 8380, 534, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 530, 12079, 6468, 535, 2625, 535, 1056, 7174, 17261, 730, 63688, 730, 10173, 528, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 532, 12079, 6468, 535, 2625, 535, 1417, 923, 32535, 730, 49883, 837, 527, 1295, 49321, 884, 61026, 730, 30664, 2108, 536, 61026, 3567, 811, 16687, 43497, 708, 6468, 535, 3024, 535, 1056, 7174, 17261, 730, 63688, 730, 10426, 533, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 526, 12079, 6468, 535, 3188, 535, 1056, 7174, 17261, 730, 63688, 730, 8380, 525, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 528, 12079, 6468, 535, 2108, 535, 1056, 7174, 17261, 730, 63688, 730, 11233, 529, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 528, 12079, 6468, 535, 4536, 535, 1374, 923, 32535, 730, 43497, 730, 532, 8323, 7925, 2835, 730, 533, 523, 534, 835, 843, 35436, 730, 533, 523, 525, 835, 518, 875, 730, 530, 775, 708, 6468, 535, 4536, 535, 5046, 923, 32535, 730, 43497, 730, 532, 8323, 7925, 2835, 730, 534, 523, 529, 835, 843, 35436, 730, 533, 523, 525, 835, 518, 875, 730, 1374, 775, 708, 6468, 535, 4344, 535, 2108, 923, 32535, 730, 43497, 730, 532, 23961, 6586, 730, 526, 523, 5370, 796, 5266, 843, 2390, 5723, 730, 526, 523, 2346, 796, 5266, 1015, 6468, 535, 4613, 535, 3654, 59687, 2850, 43497, 730, 532, 16017, 12365, 535, 11697, 38993, 578, 27493, 532, 522, 528, 26781, 875, 1209, 1399, 730, 528, 775, 708, 6468, 535, 4613, 535, 6468, 59687, 2850, 8411, 886, 826, 526, 13525, 968, 6155, 22694, 843, 7188, 535, 16017, 43497, 730, 532, 4032, 1113, 50027, 23993, 791, 708, 6468, 535, 4613, 535, 6468, 7174, 17261, 730, 1641, 13053, 19461, 1189, 60433, 536, 19331, 1706, 4908, 884, 24163, 57506, 708, 6468, 535, 4613, 535, 2108, 7174, 17261, 730, 16753, 28984, 968, 1267, 868, 2290, 963, 779, 2669, 26862, 708, 6468, 535, 3337, 535, 1386, 7174, 17261, 730, 12003, 1868, 9388, 811, 4582, 1055, 843, 589, 60088, 730, 526, 1015, 6468, 535, 4195, 535, 5843, 7174, 17261, 730, 49883, 837, 526, 810, 837, 527, 54126, 843, 2744, 19331, 1706, 4032, 2515, 29067, 7410, 708, 6468, 535, 3745, 535, 1664, 7174, 17261, 730, 24163, 57506, 963, 730, 2752, 514, 7614, 708, 6468, 535, 1838, 535, 1374, 7174, 17261, 730, 4582, 1055, 884, 3475, 963, 43497, 730, 532, 536, 11023, 17652, 6944, 811, 43497, 730, 532, 8323, 708, 6468, 535, 7058, 535, 3167, 7174, 17261, 730, 24163, 57506, 963, 730, 9731, 514, 7614, 536, 1641, 13053, 18483, 730, 528, 810, 730, 529, 8557, 811, 1361, 19082, 871, 708, 6468, 535, 3878, 535, 898, 7174, 17261, 730, 16017, 3360, 4468, 968, 4582, 1055, 535, 38169, 8227, 772, 28782, 4563, 1490, 779, 43497, 730, 532, 23961, 963, 779, 2341, 46129, 708, 6468, 535, 5046, 535, 4536, 7174, 17261, 730, 8227, 772, 8365, 988, 43497, 730, 532, 536, 46129, 909, 532, 522, 527, 2083, 7937, 521, 1295, 5896, 1833, 708, 4972, 535, 1056, 535, 3654, 7174, 17261, 730, 43497, 730, 532, 23961, 17828, 8744, 4174, 35036, 536, 32215, 4984, 708, 4972, 535, 6135, 535, 1374, 7174, 17261, 730, 43497, 730, 532, 23961, 17828, 23848, 536, 1927, 2685, 963, 2920, 6586, 730, 525, 523, 530, 796, 5266, 875, 730, 2198, 775, 7410, 708, 4972, 535, 6326, 535, 1698, 7174, 17261, 730, 11023, 17652, 8365, 988, 43497, 730, 532, 8323, 708, 4972, 535, 3654, 535, 1664, 7174, 17261, 730, 8411, 886, 826, 526, 2280, 33480, 968, 1267, 868, 2290, 536, 878, 4318, 1047, 730, 2019, 1750, 730, 1826, 775, 708, 4972, 535, 3654, 535, 2999, 7174, 17261, 730, 1641, 13053, 19461, 1189, 49826, 536, 1641, 13053, 18483, 730, 528, 810, 730, 529, 840, 59188, 708, 4972, 535, 6468, 535, 1056, 7174, 17261, 730, 43497, 730, 532, 8323, 7925, 2835, 730, 531, 523, 526, 835, 843, 35436, 730, 533, 523, 525, 835, 1015, 4972, 535, 4905, 535, 1056, 7174, 17261, 730, 63688, 730, 4787, 525, 19331, 1706, 10012, 843, 3321, 3600, 38689, 4032, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 525, 12079, 4972, 535, 1374, 535, 1056, 7174, 17261, 730, 63688, 730, 3818, 531, 19331, 1706, 10012, 843, 3321, 3600, 38689, 4032, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 528, 12079, 4972, 535, 1386, 535, 1056, 7174, 17261, 730, 63688, 730, 4420, 526, 19331, 1706, 10012, 843, 3321, 3600, 38689, 4032, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 531, 12079, 4972, 535, 1386, 535, 3337, 923, 32535, 730, 49883, 837, 526, 1295, 49321, 884, 61026, 730, 32928, 2476, 536, 61026, 3567, 811, 16687, 43497, 708, 4972, 535, 1133, 535, 1056, 7174, 17261, 730, 63688, 730, 6583, 527, 19331, 1706, 10012, 843, 3321, 3600, 38689, 4032, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 525, 12079, 4972, 535, 2197, 535, 1056, 7174, 17261, 730, 63688, 730, 9887, 533, 19331, 1706, 10012, 843, 3321, 3600, 38689, 4032, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 528, 12079, 4972, 535, 2197, 535, 4905, 7174, 17261, 730, 24163, 57506, 1892, 4174, 730, 1155, 514, 7614, 708, 4972, 535, 1826, 535, 1056, 7174, 17261, 730, 63688, 730, 10556, 526, 19331, 1706, 10012, 843, 3321, 3600, 38689, 4032, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 531, 12079, 4972, 535, 2142, 535, 1056, 7174, 17261, 730, 63688, 730, 10520, 527, 19331, 1706, 10012, 843, 3321, 3600, 38689, 4032, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 525, 12079, 4972, 535, 2142, 535, 1664, 7174, 17261, 730, 43497, 730, 532, 8323, 7925, 2835, 730, 530, 523, 534, 835, 843, 35436, 730, 533, 523, 525, 835, 4032, 23961, 6586, 730, 526, 523, 2346, 796, 5266, 708, 4972, 535, 1664, 535, 1056, 7174, 17261, 730, 63688, 730, 10426, 530, 19331, 1706, 10012, 843, 3321, 3600, 38689, 4032, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 528, 12079, 4972, 535, 2625, 535, 1056, 7174, 17261, 730, 8501, 6634, 968, 1267, 868, 2290, 535, 43497, 730, 532, 23961, 17828, 811, 874, 20425, 2030, 963, 730, 1155, 535, 1056, 536, 8227, 772, 42806, 811, 874, 5689, 1337, 8233, 20842, 708, 4972, 535, 2625, 535, 3654, 7174, 17261, 730, 8411, 886, 826, 526, 6840, 12465, 536, 1295, 5592, 16753, 592, 708, 4972, 535, 3024, 535, 1056, 7174, 17261, 730, 63688, 730, 8380, 533, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 525, 12079, 4972, 535, 2108, 535, 1056, 7174, 17261, 730, 63688, 730, 10426, 531, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 526, 12079, 4972, 535, 3337, 535, 1056, 7174, 17261, 730, 63688, 730, 11233, 534, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 527, 12079, 4972, 535, 3337, 535, 4536, 7174, 17261, 730, 49574, 886, 889, 529, 41172, 3999, 7410, 843, 22500, 730, 525, 523, 4613, 8061, 1015, 4972, 535, 3167, 535, 1056, 7174, 17261, 730, 63688, 730, 10520, 527, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 528, 12079, 4972, 535, 7058, 535, 1056, 7174, 17261, 730, 63688, 730, 8380, 529, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 529, 12079, 4972, 535, 7058, 535, 1417, 923, 32535, 730, 49883, 837, 526, 1295, 49321, 884, 61026, 730, 22218, 1532, 536, 61026, 3567, 811, 16687, 43497, 708, 4972, 535, 7568, 535, 1056, 7174, 17261, 730, 63688, 730, 10426, 526, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 530, 12079, 4528, 535, 1056, 535, 1056, 7174, 17261, 730, 63688, 730, 11233, 528, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 531, 12079, 4528, 535, 1056, 535, 1664, 7174, 17261, 730, 43497, 730, 532, 8323, 7925, 2835, 730, 530, 523, 533, 835, 843, 35436, 730, 533, 523, 525, 835, 4032, 23961, 6586, 730, 526, 523, 2346, 796, 5266, 708, 4528, 535, 6326, 535, 1056, 7174, 17261, 730, 63688, 730, 10426, 534, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 532, 12079, 4528, 535, 4528, 535, 1056, 7174, 17261, 730, 63688, 730, 8380, 527, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 525, 12079, 4528, 535, 4528, 535, 4536, 7174, 17261, 730, 49574, 886, 889, 529, 41172, 3999, 7410, 843, 22500, 730, 525, 523, 3756, 8061, 1015, 4528, 535, 1374, 535, 1056, 7174, 17261, 730, 63688, 730, 10426, 529, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 526, 12079, 4528, 535, 1374, 535, 1838, 7174, 17261, 730, 24163, 57506, 963, 730, 533, 514, 7614, 708, 4528, 535, 1443, 535, 1056, 7174, 17261, 730, 63688, 730, 11233, 532, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 527, 12079, 4528, 535, 898, 535, 1056, 7174, 17261, 730, 63688, 730, 10520, 530, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 528, 12079, 4528, 535, 898, 535, 1417, 923, 32535, 730, 49883, 837, 527, 1295, 49321, 884, 61026, 730, 28701, 1155, 536, 61026, 3567, 811, 16687, 43497, 708, 4528, 535, 1826, 535, 1056, 7174, 17261, 730, 63688, 730, 8380, 531, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 529, 12079, 4528, 535, 2138, 535, 1056, 7174, 17261, 730, 63688, 730, 11233, 525, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 530, 12079, 4528, 535, 2925, 535, 1056, 7174, 17261, 730, 63688, 730, 10426, 533, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 531, 12079, 4528, 535, 2925, 535, 4536, 7174, 17261, 730, 49574, 886, 889, 529, 41172, 3999, 7410, 843, 22500, 730, 525, 523, 3878, 8061, 1015, 4528, 535, 3024, 535, 1056, 7174, 17261, 730, 63688, 730, 8380, 526, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 532, 12079, 4528, 535, 2108, 535, 1056, 7174, 17261, 730, 63688, 730, 8380, 533, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 525, 12079, 4528, 535, 2108, 535, 1664, 7174, 17261, 730, 43497, 730, 532, 8323, 7925, 2835, 730, 530, 523, 534, 835, 843, 35436, 730, 533, 523, 525, 835, 4032, 23961, 6586, 730, 526, 523, 2346, 796, 5266, 708, 4528, 535, 3337, 535, 1056, 7174, 17261, 730, 63688, 730, 10426, 531, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 526, 12079, 4528, 535, 3167, 535, 1056, 7174, 17261, 730, 63688, 730, 11233, 534, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 527, 12079, 4528, 535, 3167, 535, 1417, 923, 32535, 730, 49883, 837, 526, 1295, 49321, 884, 61026, 730, 26465, 4972, 536, 61026, 3567, 811, 16687, 43497, 708, 4528, 535, 7058, 535, 1056, 7174, 17261, 730, 63688, 730, 10520, 527, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 528, 12079, 4528, 535, 7568, 535, 1056, 7174, 17261, 730, 63688, 730, 8380, 529, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 529, 12079, 4528, 535, 7568, 535, 4536, 7174, 17261, 730, 49574, 886, 889, 529, 41172, 3999, 7410, 843, 22500, 730, 525, 523, 2999, 8061, 1015, 4905, 535, 1056, 535, 1056, 7174, 17261, 730, 63688, 730, 10426, 526, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 530, 12079, 4905, 535, 1056, 535, 1838, 7174, 17261, 730, 24163, 57506, 963, 730, 530, 514, 7614, 708, 4905, 535, 6326, 535, 1056, 7174, 17261, 730, 63688, 730, 11233, 528, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 531, 12079, 4905, 535, 4528, 535, 1056, 7174, 17261, 730, 63688, 730, 10426, 534, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 532, 12079, 4905, 535, 1374, 535, 1056, 7174, 17261, 730, 63688, 730, 8380, 527, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 525, 12079, 4905, 535, 1443, 535, 1056, 7174, 17261, 730, 63688, 730, 10426, 529, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 526, 12079, 4905, 535, 1443, 535, 1417, 923, 32535, 730, 49883, 837, 527, 1295, 49321, 884, 61026, 730, 32522, 6326, 536, 61026, 3567, 811, 16687, 43497, 708, 4905, 535, 898, 535, 1056, 7174, 17261, 730, 63688, 730, 11233, 532, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 527, 12079, 4905, 535, 898, 535, 4536, 7174, 17261, 730, 49574, 886, 889, 529, 41172, 3999, 7410, 843, 22500, 730, 525, 523, 3092, 8061, 1015, 4905, 535, 898, 535, 1664, 7174, 17261, 730, 43497, 730, 532, 8323, 7925, 2835, 730, 531, 523, 525, 835, 843, 35436, 730, 533, 523, 525, 835, 4032, 23961, 6586, 730, 526, 523, 2346, 796, 5266, 708, 4905, 535, 1826, 535, 1056, 7174, 17261, 730, 63688, 730, 10520, 530, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 528, 12079, 4905, 535, 2138, 535, 1056, 7174, 17261, 730, 63688, 730, 8380, 531, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 529, 12079, 4905, 535, 2925, 535, 1056, 7174, 17261, 730, 63688, 730, 11233, 525, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 530, 12079, 4905, 535, 3024, 535, 1056, 7174, 17261, 730, 63688, 730, 10426, 533, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 531, 12079, 4905, 535, 2108, 535, 1056, 7174, 17261, 730, 63688, 730, 8380, 526, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 532, 12079, 4905, 535, 3337, 535, 1056, 7174, 17261, 730, 63688, 730, 8380, 533, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 525, 12079, 4905, 535, 3337, 535, 4536, 7174, 17261, 730, 49574, 886, 889, 529, 41172, 3999, 7410, 843, 22500, 730, 525, 523, 4613, 8061, 1015, 4905, 535, 3337, 535, 1417, 923, 32535, 730, 49883, 837, 526, 1295, 49321, 884, 61026, 730, 33231, 4541, 536, 61026, 3567, 811, 16687, 43497, 708, 4905, 535, 3167, 535, 1056, 7174, 17261, 730, 63688, 730, 10426, 531, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 526, 12079, 4905, 535, 3167, 535, 1838, 7174, 17261, 730, 24163, 57506, 963, 730, 532, 514, 7614, 708, 4905, 535, 7058, 535, 1056, 7174, 17261, 730, 63688, 730, 11233, 534, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 527, 12079, 4905, 535, 7568, 535, 1056, 7174, 17261, 730, 63688, 730, 10520, 527, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 528, 12079, 1155, 535, 1056, 535, 1056, 7174, 17261, 730, 63688, 730, 8380, 529, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 529, 12079, 1155, 535, 1056, 535, 1664, 7174, 17261, 730, 43497, 730, 532, 8323, 7925, 2835, 730, 530, 523, 533, 835, 843, 35436, 730, 533, 523, 525, 835, 4032, 23961, 6586, 730, 526, 523, 2346, 796, 5266, 708, 1155, 535, 1056, 535, 2108, 7174, 17261, 730, 43497, 730, 532, 23961, 17828, 20425, 968, 4582, 1055, 843, 802, 13573, 797, 779, 730, 4972, 535, 2625, 8501, 6634, 3271, 2638, 35036, 536, 1295, 32215, 4602, 708, 1155, 535, 6326, 535, 1056, 7174, 17261, 730, 63688, 730, 10426, 526, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 530, 12079, 1155, 535, 4528, 535, 1056, 7174, 17261, 730, 63688, 730, 11233, 528, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 531, 12079, 1155, 535, 4528, 535, 4536, 7174, 17261, 730, 49574, 886, 889, 529, 41172, 3999, 7410, 843, 22500, 730, 525, 523, 3756, 8061, 1015, 1155, 535, 1374, 535, 1056, 7174, 17261, 730, 63688, 730, 10426, 534, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 532, 12079, 1155, 535, 1374, 535, 1417, 923, 32535, 730, 49883, 837, 527, 1295, 49321, 884, 61026, 730, 24656, 3361, 536, 61026, 3567, 811, 16687, 43497, 708, 1155, 535, 1443, 535, 1056, 7174, 17261, 730, 63688, 730, 8380, 527, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 525, 12079, 1155, 535, 898, 535, 1056, 7174, 17261, 730, 63688, 730, 10426, 529, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 526, 12079, 1155, 535, 2197, 535, 3654, 7174, 17261, 730, 8411, 886, 826, 526, 6840, 12465, 536, 1295, 5592, 16753, 592, 18, 28828, 8150, 5457, 779, 878, 4318, 1047, 540, 19, 16, 49883, 535, 3492, 3955, 49883, 20, 19, 16, 43497, 572, 532, 535, 2020, 4674, 777, 43497, 730, 532, 20, 19, 16, 49574, 886, 535, 889, 1991, 886, 5716, 20, 19, 16, 2410, 535, 9642, 7018, 20, 19, 16, 5901, 535, 10478, 5901, 20, 21 ], "markers": [ 3422, 3430, 3443, 3453, 3460 ], "logits_raw": [ -2.666999101638794, -0.14830780029296875, 0.5828009843826294, -1.5615222454071045, -2.4719622135162354 ], "probs_raw": [ 0.023023677989840508, 0.2857777774333954, 0.5936698317527771, 0.06954680383205414, 0.027981961145997047 ], "probs": [ 0.048573244363069534, 0.29361486434936523, 0.4949849545955658, 0.10699225217103958, 0.05583466589450836 ], "argmax_index": 2, "argmax": "diverter", "top2_margin": 0.20137009024620056, "near_tie": false, "gold": "lane_7", "correct": false, "host_softmax_max_abs_dp": 3.725290298461914e-09, "batch": { "call": 0, "index": 1, "B": 4 } }, { "qid": "normal_at_end", "type": "noul", "K": 2, "calibrate": true, "temperature_key": "noul:2", "T": 1.6663223505020142, "prefix": 0, "n_ids": 3449, "positions": 3449, "bucket": 4096, "ids": [ 1, 17, 15991, 23707, 3613, 4425, 4857, 561, 33371, 12292, 38620, 1141, 61026, 17824, 730, 527, 521, 5008, 1985, 2647, 8411, 886, 826, 526, 521, 22694, 3446, 843, 1513, 31225, 1015, 27464, 535, 2960, 843, 6468, 535, 1056, 522, 1536, 535, 1056, 518, 767, 52574, 884, 15774, 535, 1267, 868, 2290, 767, 12003, 884, 1868, 535, 4582, 1055, 509, 6468, 535, 1056, 535, 1056, 7174, 17261, 730, 63688, 730, 10426, 527, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 526, 12079, 6468, 535, 6135, 535, 1056, 7174, 17261, 730, 63688, 730, 11233, 533, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 528, 12079, 6468, 535, 6468, 535, 1056, 7174, 17261, 730, 63688, 730, 10556, 526, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 530, 12079, 6468, 535, 4905, 535, 1056, 7174, 17261, 730, 63688, 730, 8380, 529, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 532, 12079, 6468, 535, 4905, 535, 4536, 7174, 17261, 730, 49574, 886, 889, 529, 41172, 3999, 7410, 843, 22500, 730, 525, 523, 3188, 8061, 1015, 6468, 535, 1374, 535, 1056, 7174, 17261, 730, 63688, 730, 10173, 534, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 526, 12079, 6468, 535, 1374, 535, 1417, 923, 32535, 730, 49883, 837, 527, 1295, 49321, 884, 61026, 730, 33118, 3375, 536, 61026, 3567, 811, 16687, 43497, 708, 6468, 535, 1386, 535, 1056, 7174, 17261, 730, 63688, 730, 10520, 525, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 528, 12079, 6468, 535, 1133, 535, 1056, 7174, 17261, 730, 63688, 730, 10426, 531, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 530, 12079, 6468, 535, 2197, 535, 1056, 7174, 17261, 730, 63688, 730, 8380, 527, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 532, 12079, 6468, 535, 1826, 535, 1056, 7174, 17261, 730, 63688, 730, 11233, 530, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 526, 12079, 6468, 535, 2142, 535, 1056, 7174, 17261, 730, 63688, 730, 10520, 532, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 528, 12079, 6468, 535, 2142, 535, 4536, 7174, 17261, 730, 49574, 886, 889, 529, 41172, 3999, 7410, 843, 22500, 730, 525, 523, 5489, 8061, 1015, 6468, 535, 1664, 535, 1056, 7174, 17261, 730, 63688, 730, 8380, 534, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 530, 12079, 6468, 535, 2625, 535, 1056, 7174, 17261, 730, 63688, 730, 10173, 528, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 532, 12079, 6468, 535, 2625, 535, 1417, 923, 32535, 730, 49883, 837, 527, 1295, 49321, 884, 61026, 730, 30664, 2108, 536, 61026, 3567, 811, 16687, 43497, 708, 6468, 535, 3024, 535, 1056, 7174, 17261, 730, 63688, 730, 10426, 533, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 526, 12079, 6468, 535, 3188, 535, 1056, 7174, 17261, 730, 63688, 730, 8380, 525, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 528, 12079, 6468, 535, 2108, 535, 1056, 7174, 17261, 730, 63688, 730, 11233, 529, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 528, 12079, 6468, 535, 4536, 535, 1374, 923, 32535, 730, 43497, 730, 532, 8323, 7925, 2835, 730, 533, 523, 534, 835, 843, 35436, 730, 533, 523, 525, 835, 518, 875, 730, 530, 775, 708, 6468, 535, 4536, 535, 5046, 923, 32535, 730, 43497, 730, 532, 8323, 7925, 2835, 730, 534, 523, 529, 835, 843, 35436, 730, 533, 523, 525, 835, 518, 875, 730, 1374, 775, 708, 6468, 535, 4344, 535, 2108, 923, 32535, 730, 43497, 730, 532, 23961, 6586, 730, 526, 523, 5370, 796, 5266, 843, 2390, 5723, 730, 526, 523, 2346, 796, 5266, 1015, 6468, 535, 4613, 535, 3654, 59687, 2850, 43497, 730, 532, 16017, 12365, 535, 11697, 38993, 578, 27493, 532, 522, 528, 26781, 875, 1209, 1399, 730, 528, 775, 708, 6468, 535, 4613, 535, 6468, 59687, 2850, 8411, 886, 826, 526, 13525, 968, 6155, 22694, 843, 7188, 535, 16017, 43497, 730, 532, 4032, 1113, 50027, 23993, 791, 708, 6468, 535, 4613, 535, 6468, 7174, 17261, 730, 1641, 13053, 19461, 1189, 60433, 536, 19331, 1706, 4908, 884, 24163, 57506, 708, 6468, 535, 4613, 535, 2108, 7174, 17261, 730, 16753, 28984, 968, 1267, 868, 2290, 963, 779, 2669, 26862, 708, 6468, 535, 3337, 535, 1386, 7174, 17261, 730, 12003, 1868, 9388, 811, 4582, 1055, 843, 589, 60088, 730, 526, 1015, 6468, 535, 4195, 535, 5843, 7174, 17261, 730, 49883, 837, 526, 810, 837, 527, 54126, 843, 2744, 19331, 1706, 4032, 2515, 29067, 7410, 708, 6468, 535, 3745, 535, 1664, 7174, 17261, 730, 24163, 57506, 963, 730, 2752, 514, 7614, 708, 6468, 535, 1838, 535, 1374, 7174, 17261, 730, 4582, 1055, 884, 3475, 963, 43497, 730, 532, 536, 11023, 17652, 6944, 811, 43497, 730, 532, 8323, 708, 6468, 535, 7058, 535, 3167, 7174, 17261, 730, 24163, 57506, 963, 730, 9731, 514, 7614, 536, 1641, 13053, 18483, 730, 528, 810, 730, 529, 8557, 811, 1361, 19082, 871, 708, 6468, 535, 3878, 535, 898, 7174, 17261, 730, 16017, 3360, 4468, 968, 4582, 1055, 535, 38169, 8227, 772, 28782, 4563, 1490, 779, 43497, 730, 532, 23961, 963, 779, 2341, 46129, 708, 6468, 535, 5046, 535, 4536, 7174, 17261, 730, 8227, 772, 8365, 988, 43497, 730, 532, 536, 46129, 909, 532, 522, 527, 2083, 7937, 521, 1295, 5896, 1833, 708, 4972, 535, 1056, 535, 3654, 7174, 17261, 730, 43497, 730, 532, 23961, 17828, 8744, 4174, 35036, 536, 32215, 4984, 708, 4972, 535, 6135, 535, 1374, 7174, 17261, 730, 43497, 730, 532, 23961, 17828, 23848, 536, 1927, 2685, 963, 2920, 6586, 730, 525, 523, 530, 796, 5266, 875, 730, 2198, 775, 7410, 708, 4972, 535, 6326, 535, 1698, 7174, 17261, 730, 11023, 17652, 8365, 988, 43497, 730, 532, 8323, 708, 4972, 535, 3654, 535, 1664, 7174, 17261, 730, 8411, 886, 826, 526, 2280, 33480, 968, 1267, 868, 2290, 536, 878, 4318, 1047, 730, 2019, 1750, 730, 1826, 775, 708, 4972, 535, 3654, 535, 2999, 7174, 17261, 730, 1641, 13053, 19461, 1189, 49826, 536, 1641, 13053, 18483, 730, 528, 810, 730, 529, 840, 59188, 708, 4972, 535, 6468, 535, 1056, 7174, 17261, 730, 43497, 730, 532, 8323, 7925, 2835, 730, 531, 523, 526, 835, 843, 35436, 730, 533, 523, 525, 835, 1015, 4972, 535, 4905, 535, 1056, 7174, 17261, 730, 63688, 730, 4787, 525, 19331, 1706, 10012, 843, 3321, 3600, 38689, 4032, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 525, 12079, 4972, 535, 1374, 535, 1056, 7174, 17261, 730, 63688, 730, 3818, 531, 19331, 1706, 10012, 843, 3321, 3600, 38689, 4032, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 528, 12079, 4972, 535, 1386, 535, 1056, 7174, 17261, 730, 63688, 730, 4420, 526, 19331, 1706, 10012, 843, 3321, 3600, 38689, 4032, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 531, 12079, 4972, 535, 1386, 535, 3337, 923, 32535, 730, 49883, 837, 526, 1295, 49321, 884, 61026, 730, 32928, 2476, 536, 61026, 3567, 811, 16687, 43497, 708, 4972, 535, 1133, 535, 1056, 7174, 17261, 730, 63688, 730, 6583, 527, 19331, 1706, 10012, 843, 3321, 3600, 38689, 4032, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 525, 12079, 4972, 535, 2197, 535, 1056, 7174, 17261, 730, 63688, 730, 9887, 533, 19331, 1706, 10012, 843, 3321, 3600, 38689, 4032, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 528, 12079, 4972, 535, 2197, 535, 4905, 7174, 17261, 730, 24163, 57506, 1892, 4174, 730, 1155, 514, 7614, 708, 4972, 535, 1826, 535, 1056, 7174, 17261, 730, 63688, 730, 10556, 526, 19331, 1706, 10012, 843, 3321, 3600, 38689, 4032, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 531, 12079, 4972, 535, 2142, 535, 1056, 7174, 17261, 730, 63688, 730, 10520, 527, 19331, 1706, 10012, 843, 3321, 3600, 38689, 4032, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 525, 12079, 4972, 535, 2142, 535, 1664, 7174, 17261, 730, 43497, 730, 532, 8323, 7925, 2835, 730, 530, 523, 534, 835, 843, 35436, 730, 533, 523, 525, 835, 4032, 23961, 6586, 730, 526, 523, 2346, 796, 5266, 708, 4972, 535, 1664, 535, 1056, 7174, 17261, 730, 63688, 730, 10426, 530, 19331, 1706, 10012, 843, 3321, 3600, 38689, 4032, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 528, 12079, 4972, 535, 2625, 535, 1056, 7174, 17261, 730, 8501, 6634, 968, 1267, 868, 2290, 535, 43497, 730, 532, 23961, 17828, 811, 874, 20425, 2030, 963, 730, 1155, 535, 1056, 536, 8227, 772, 42806, 811, 874, 5689, 1337, 8233, 20842, 708, 4972, 535, 2625, 535, 3654, 7174, 17261, 730, 8411, 886, 826, 526, 6840, 12465, 536, 1295, 5592, 16753, 592, 708, 4972, 535, 3024, 535, 1056, 7174, 17261, 730, 63688, 730, 8380, 533, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 525, 12079, 4972, 535, 2108, 535, 1056, 7174, 17261, 730, 63688, 730, 10426, 531, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 526, 12079, 4972, 535, 3337, 535, 1056, 7174, 17261, 730, 63688, 730, 11233, 534, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 527, 12079, 4972, 535, 3337, 535, 4536, 7174, 17261, 730, 49574, 886, 889, 529, 41172, 3999, 7410, 843, 22500, 730, 525, 523, 4613, 8061, 1015, 4972, 535, 3167, 535, 1056, 7174, 17261, 730, 63688, 730, 10520, 527, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 528, 12079, 4972, 535, 7058, 535, 1056, 7174, 17261, 730, 63688, 730, 8380, 529, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 529, 12079, 4972, 535, 7058, 535, 1417, 923, 32535, 730, 49883, 837, 526, 1295, 49321, 884, 61026, 730, 22218, 1532, 536, 61026, 3567, 811, 16687, 43497, 708, 4972, 535, 7568, 535, 1056, 7174, 17261, 730, 63688, 730, 10426, 526, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 530, 12079, 4528, 535, 1056, 535, 1056, 7174, 17261, 730, 63688, 730, 11233, 528, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 531, 12079, 4528, 535, 1056, 535, 1664, 7174, 17261, 730, 43497, 730, 532, 8323, 7925, 2835, 730, 530, 523, 533, 835, 843, 35436, 730, 533, 523, 525, 835, 4032, 23961, 6586, 730, 526, 523, 2346, 796, 5266, 708, 4528, 535, 6326, 535, 1056, 7174, 17261, 730, 63688, 730, 10426, 534, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 532, 12079, 4528, 535, 4528, 535, 1056, 7174, 17261, 730, 63688, 730, 8380, 527, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 525, 12079, 4528, 535, 4528, 535, 4536, 7174, 17261, 730, 49574, 886, 889, 529, 41172, 3999, 7410, 843, 22500, 730, 525, 523, 3756, 8061, 1015, 4528, 535, 1374, 535, 1056, 7174, 17261, 730, 63688, 730, 10426, 529, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 526, 12079, 4528, 535, 1374, 535, 1838, 7174, 17261, 730, 24163, 57506, 963, 730, 533, 514, 7614, 708, 4528, 535, 1443, 535, 1056, 7174, 17261, 730, 63688, 730, 11233, 532, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 527, 12079, 4528, 535, 898, 535, 1056, 7174, 17261, 730, 63688, 730, 10520, 530, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 528, 12079, 4528, 535, 898, 535, 1417, 923, 32535, 730, 49883, 837, 527, 1295, 49321, 884, 61026, 730, 28701, 1155, 536, 61026, 3567, 811, 16687, 43497, 708, 4528, 535, 1826, 535, 1056, 7174, 17261, 730, 63688, 730, 8380, 531, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 529, 12079, 4528, 535, 2138, 535, 1056, 7174, 17261, 730, 63688, 730, 11233, 525, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 530, 12079, 4528, 535, 2925, 535, 1056, 7174, 17261, 730, 63688, 730, 10426, 533, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 531, 12079, 4528, 535, 2925, 535, 4536, 7174, 17261, 730, 49574, 886, 889, 529, 41172, 3999, 7410, 843, 22500, 730, 525, 523, 3878, 8061, 1015, 4528, 535, 3024, 535, 1056, 7174, 17261, 730, 63688, 730, 8380, 526, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 532, 12079, 4528, 535, 2108, 535, 1056, 7174, 17261, 730, 63688, 730, 8380, 533, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 525, 12079, 4528, 535, 2108, 535, 1664, 7174, 17261, 730, 43497, 730, 532, 8323, 7925, 2835, 730, 530, 523, 534, 835, 843, 35436, 730, 533, 523, 525, 835, 4032, 23961, 6586, 730, 526, 523, 2346, 796, 5266, 708, 4528, 535, 3337, 535, 1056, 7174, 17261, 730, 63688, 730, 10426, 531, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 526, 12079, 4528, 535, 3167, 535, 1056, 7174, 17261, 730, 63688, 730, 11233, 534, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 527, 12079, 4528, 535, 3167, 535, 1417, 923, 32535, 730, 49883, 837, 526, 1295, 49321, 884, 61026, 730, 26465, 4972, 536, 61026, 3567, 811, 16687, 43497, 708, 4528, 535, 7058, 535, 1056, 7174, 17261, 730, 63688, 730, 10520, 527, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 528, 12079, 4528, 535, 7568, 535, 1056, 7174, 17261, 730, 63688, 730, 8380, 529, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 529, 12079, 4528, 535, 7568, 535, 4536, 7174, 17261, 730, 49574, 886, 889, 529, 41172, 3999, 7410, 843, 22500, 730, 525, 523, 2999, 8061, 1015, 4905, 535, 1056, 535, 1056, 7174, 17261, 730, 63688, 730, 10426, 526, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 530, 12079, 4905, 535, 1056, 535, 1838, 7174, 17261, 730, 24163, 57506, 963, 730, 530, 514, 7614, 708, 4905, 535, 6326, 535, 1056, 7174, 17261, 730, 63688, 730, 11233, 528, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 531, 12079, 4905, 535, 4528, 535, 1056, 7174, 17261, 730, 63688, 730, 10426, 534, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 532, 12079, 4905, 535, 1374, 535, 1056, 7174, 17261, 730, 63688, 730, 8380, 527, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 525, 12079, 4905, 535, 1443, 535, 1056, 7174, 17261, 730, 63688, 730, 10426, 529, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 526, 12079, 4905, 535, 1443, 535, 1417, 923, 32535, 730, 49883, 837, 527, 1295, 49321, 884, 61026, 730, 32522, 6326, 536, 61026, 3567, 811, 16687, 43497, 708, 4905, 535, 898, 535, 1056, 7174, 17261, 730, 63688, 730, 11233, 532, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 527, 12079, 4905, 535, 898, 535, 4536, 7174, 17261, 730, 49574, 886, 889, 529, 41172, 3999, 7410, 843, 22500, 730, 525, 523, 3092, 8061, 1015, 4905, 535, 898, 535, 1664, 7174, 17261, 730, 43497, 730, 532, 8323, 7925, 2835, 730, 531, 523, 525, 835, 843, 35436, 730, 533, 523, 525, 835, 4032, 23961, 6586, 730, 526, 523, 2346, 796, 5266, 708, 4905, 535, 1826, 535, 1056, 7174, 17261, 730, 63688, 730, 10520, 530, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 528, 12079, 4905, 535, 2138, 535, 1056, 7174, 17261, 730, 63688, 730, 8380, 531, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 529, 12079, 4905, 535, 2925, 535, 1056, 7174, 17261, 730, 63688, 730, 11233, 525, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 530, 12079, 4905, 535, 3024, 535, 1056, 7174, 17261, 730, 63688, 730, 10426, 533, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 531, 12079, 4905, 535, 2108, 535, 1056, 7174, 17261, 730, 63688, 730, 8380, 526, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 532, 12079, 4905, 535, 3337, 535, 1056, 7174, 17261, 730, 63688, 730, 8380, 533, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 525, 12079, 4905, 535, 3337, 535, 4536, 7174, 17261, 730, 49574, 886, 889, 529, 41172, 3999, 7410, 843, 22500, 730, 525, 523, 4613, 8061, 1015, 4905, 535, 3337, 535, 1417, 923, 32535, 730, 49883, 837, 526, 1295, 49321, 884, 61026, 730, 33231, 4541, 536, 61026, 3567, 811, 16687, 43497, 708, 4905, 535, 3167, 535, 1056, 7174, 17261, 730, 63688, 730, 10426, 531, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 526, 12079, 4905, 535, 3167, 535, 1838, 7174, 17261, 730, 24163, 57506, 963, 730, 532, 514, 7614, 708, 4905, 535, 7058, 535, 1056, 7174, 17261, 730, 63688, 730, 11233, 534, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 527, 12079, 4905, 535, 7568, 535, 1056, 7174, 17261, 730, 63688, 730, 10520, 527, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 528, 12079, 1155, 535, 1056, 535, 1056, 7174, 17261, 730, 63688, 730, 8380, 529, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 529, 12079, 1155, 535, 1056, 535, 1664, 7174, 17261, 730, 43497, 730, 532, 8323, 7925, 2835, 730, 530, 523, 533, 835, 843, 35436, 730, 533, 523, 525, 835, 4032, 23961, 6586, 730, 526, 523, 2346, 796, 5266, 708, 1155, 535, 1056, 535, 2108, 7174, 17261, 730, 43497, 730, 532, 23961, 17828, 20425, 968, 4582, 1055, 843, 802, 13573, 797, 779, 730, 4972, 535, 2625, 8501, 6634, 3271, 2638, 35036, 536, 1295, 32215, 4602, 708, 1155, 535, 6326, 535, 1056, 7174, 17261, 730, 63688, 730, 10426, 526, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 530, 12079, 1155, 535, 4528, 535, 1056, 7174, 17261, 730, 63688, 730, 11233, 528, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 531, 12079, 1155, 535, 4528, 535, 4536, 7174, 17261, 730, 49574, 886, 889, 529, 41172, 3999, 7410, 843, 22500, 730, 525, 523, 3756, 8061, 1015, 1155, 535, 1374, 535, 1056, 7174, 17261, 730, 63688, 730, 10426, 534, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 532, 12079, 1155, 535, 1374, 535, 1417, 923, 32535, 730, 49883, 837, 527, 1295, 49321, 884, 61026, 730, 24656, 3361, 536, 61026, 3567, 811, 16687, 43497, 708, 1155, 535, 1443, 535, 1056, 7174, 17261, 730, 63688, 730, 8380, 527, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 525, 12079, 1155, 535, 898, 535, 1056, 7174, 17261, 730, 63688, 730, 10426, 529, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 526, 12079, 1155, 535, 2197, 535, 3654, 7174, 17261, 730, 8411, 886, 826, 526, 6840, 12465, 536, 1295, 5592, 16753, 592, 18, 28519, 779, 8411, 886, 6840, 12465, 963, 779, 1764, 803, 779, 3446, 540, 19, 16, 4831, 535, 1295, 521, 779, 8294, 2177, 1014, 3965, 20, 19, 16, 3347, 535, 18672, 521, 779, 8294, 10995, 20, 21 ], "markers": [ 3427, 3439 ], "logits_raw": [ 0.9065542221069336, 1.8131377696990967 ], "probs_raw": [ 0.28769946098327637, 0.7123005390167236 ], "probs": [ 0.6327569484710693, 0.36724305152893066 ], "argmax_index": 0, "argmax": "true", "top2_margin": 0.26551389694213867, "near_tie": false, "gold": "true", "correct": true, "host_softmax_max_abs_dp": 2.9802322387695312e-08, "batch": { "call": 0, "index": 3, "B": 4 } }, { "qid": "stoppage", "type": "score", "K": 6, "calibrate": true, "temperature_key": "score:6-10", "T": 1.0, "prefix": 0, "n_ids": 3497, "positions": 3497, "bucket": 4096, "ids": [ 1, 17, 15991, 23707, 3613, 4425, 4857, 561, 33371, 12292, 38620, 1141, 61026, 17824, 730, 527, 521, 5008, 1985, 2647, 8411, 886, 826, 526, 521, 22694, 3446, 843, 1513, 31225, 1015, 27464, 535, 2960, 843, 6468, 535, 1056, 522, 1536, 535, 1056, 518, 767, 52574, 884, 15774, 535, 1267, 868, 2290, 767, 12003, 884, 1868, 535, 4582, 1055, 509, 6468, 535, 1056, 535, 1056, 7174, 17261, 730, 63688, 730, 10426, 527, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 526, 12079, 6468, 535, 6135, 535, 1056, 7174, 17261, 730, 63688, 730, 11233, 533, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 528, 12079, 6468, 535, 6468, 535, 1056, 7174, 17261, 730, 63688, 730, 10556, 526, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 530, 12079, 6468, 535, 4905, 535, 1056, 7174, 17261, 730, 63688, 730, 8380, 529, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 532, 12079, 6468, 535, 4905, 535, 4536, 7174, 17261, 730, 49574, 886, 889, 529, 41172, 3999, 7410, 843, 22500, 730, 525, 523, 3188, 8061, 1015, 6468, 535, 1374, 535, 1056, 7174, 17261, 730, 63688, 730, 10173, 534, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 526, 12079, 6468, 535, 1374, 535, 1417, 923, 32535, 730, 49883, 837, 527, 1295, 49321, 884, 61026, 730, 33118, 3375, 536, 61026, 3567, 811, 16687, 43497, 708, 6468, 535, 1386, 535, 1056, 7174, 17261, 730, 63688, 730, 10520, 525, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 528, 12079, 6468, 535, 1133, 535, 1056, 7174, 17261, 730, 63688, 730, 10426, 531, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 530, 12079, 6468, 535, 2197, 535, 1056, 7174, 17261, 730, 63688, 730, 8380, 527, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 532, 12079, 6468, 535, 1826, 535, 1056, 7174, 17261, 730, 63688, 730, 11233, 530, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 526, 12079, 6468, 535, 2142, 535, 1056, 7174, 17261, 730, 63688, 730, 10520, 532, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 528, 12079, 6468, 535, 2142, 535, 4536, 7174, 17261, 730, 49574, 886, 889, 529, 41172, 3999, 7410, 843, 22500, 730, 525, 523, 5489, 8061, 1015, 6468, 535, 1664, 535, 1056, 7174, 17261, 730, 63688, 730, 8380, 534, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 530, 12079, 6468, 535, 2625, 535, 1056, 7174, 17261, 730, 63688, 730, 10173, 528, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 532, 12079, 6468, 535, 2625, 535, 1417, 923, 32535, 730, 49883, 837, 527, 1295, 49321, 884, 61026, 730, 30664, 2108, 536, 61026, 3567, 811, 16687, 43497, 708, 6468, 535, 3024, 535, 1056, 7174, 17261, 730, 63688, 730, 10426, 533, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 526, 12079, 6468, 535, 3188, 535, 1056, 7174, 17261, 730, 63688, 730, 8380, 525, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 528, 12079, 6468, 535, 2108, 535, 1056, 7174, 17261, 730, 63688, 730, 11233, 529, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 528, 12079, 6468, 535, 4536, 535, 1374, 923, 32535, 730, 43497, 730, 532, 8323, 7925, 2835, 730, 533, 523, 534, 835, 843, 35436, 730, 533, 523, 525, 835, 518, 875, 730, 530, 775, 708, 6468, 535, 4536, 535, 5046, 923, 32535, 730, 43497, 730, 532, 8323, 7925, 2835, 730, 534, 523, 529, 835, 843, 35436, 730, 533, 523, 525, 835, 518, 875, 730, 1374, 775, 708, 6468, 535, 4344, 535, 2108, 923, 32535, 730, 43497, 730, 532, 23961, 6586, 730, 526, 523, 5370, 796, 5266, 843, 2390, 5723, 730, 526, 523, 2346, 796, 5266, 1015, 6468, 535, 4613, 535, 3654, 59687, 2850, 43497, 730, 532, 16017, 12365, 535, 11697, 38993, 578, 27493, 532, 522, 528, 26781, 875, 1209, 1399, 730, 528, 775, 708, 6468, 535, 4613, 535, 6468, 59687, 2850, 8411, 886, 826, 526, 13525, 968, 6155, 22694, 843, 7188, 535, 16017, 43497, 730, 532, 4032, 1113, 50027, 23993, 791, 708, 6468, 535, 4613, 535, 6468, 7174, 17261, 730, 1641, 13053, 19461, 1189, 60433, 536, 19331, 1706, 4908, 884, 24163, 57506, 708, 6468, 535, 4613, 535, 2108, 7174, 17261, 730, 16753, 28984, 968, 1267, 868, 2290, 963, 779, 2669, 26862, 708, 6468, 535, 3337, 535, 1386, 7174, 17261, 730, 12003, 1868, 9388, 811, 4582, 1055, 843, 589, 60088, 730, 526, 1015, 6468, 535, 4195, 535, 5843, 7174, 17261, 730, 49883, 837, 526, 810, 837, 527, 54126, 843, 2744, 19331, 1706, 4032, 2515, 29067, 7410, 708, 6468, 535, 3745, 535, 1664, 7174, 17261, 730, 24163, 57506, 963, 730, 2752, 514, 7614, 708, 6468, 535, 1838, 535, 1374, 7174, 17261, 730, 4582, 1055, 884, 3475, 963, 43497, 730, 532, 536, 11023, 17652, 6944, 811, 43497, 730, 532, 8323, 708, 6468, 535, 7058, 535, 3167, 7174, 17261, 730, 24163, 57506, 963, 730, 9731, 514, 7614, 536, 1641, 13053, 18483, 730, 528, 810, 730, 529, 8557, 811, 1361, 19082, 871, 708, 6468, 535, 3878, 535, 898, 7174, 17261, 730, 16017, 3360, 4468, 968, 4582, 1055, 535, 38169, 8227, 772, 28782, 4563, 1490, 779, 43497, 730, 532, 23961, 963, 779, 2341, 46129, 708, 6468, 535, 5046, 535, 4536, 7174, 17261, 730, 8227, 772, 8365, 988, 43497, 730, 532, 536, 46129, 909, 532, 522, 527, 2083, 7937, 521, 1295, 5896, 1833, 708, 4972, 535, 1056, 535, 3654, 7174, 17261, 730, 43497, 730, 532, 23961, 17828, 8744, 4174, 35036, 536, 32215, 4984, 708, 4972, 535, 6135, 535, 1374, 7174, 17261, 730, 43497, 730, 532, 23961, 17828, 23848, 536, 1927, 2685, 963, 2920, 6586, 730, 525, 523, 530, 796, 5266, 875, 730, 2198, 775, 7410, 708, 4972, 535, 6326, 535, 1698, 7174, 17261, 730, 11023, 17652, 8365, 988, 43497, 730, 532, 8323, 708, 4972, 535, 3654, 535, 1664, 7174, 17261, 730, 8411, 886, 826, 526, 2280, 33480, 968, 1267, 868, 2290, 536, 878, 4318, 1047, 730, 2019, 1750, 730, 1826, 775, 708, 4972, 535, 3654, 535, 2999, 7174, 17261, 730, 1641, 13053, 19461, 1189, 49826, 536, 1641, 13053, 18483, 730, 528, 810, 730, 529, 840, 59188, 708, 4972, 535, 6468, 535, 1056, 7174, 17261, 730, 43497, 730, 532, 8323, 7925, 2835, 730, 531, 523, 526, 835, 843, 35436, 730, 533, 523, 525, 835, 1015, 4972, 535, 4905, 535, 1056, 7174, 17261, 730, 63688, 730, 4787, 525, 19331, 1706, 10012, 843, 3321, 3600, 38689, 4032, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 525, 12079, 4972, 535, 1374, 535, 1056, 7174, 17261, 730, 63688, 730, 3818, 531, 19331, 1706, 10012, 843, 3321, 3600, 38689, 4032, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 528, 12079, 4972, 535, 1386, 535, 1056, 7174, 17261, 730, 63688, 730, 4420, 526, 19331, 1706, 10012, 843, 3321, 3600, 38689, 4032, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 531, 12079, 4972, 535, 1386, 535, 3337, 923, 32535, 730, 49883, 837, 526, 1295, 49321, 884, 61026, 730, 32928, 2476, 536, 61026, 3567, 811, 16687, 43497, 708, 4972, 535, 1133, 535, 1056, 7174, 17261, 730, 63688, 730, 6583, 527, 19331, 1706, 10012, 843, 3321, 3600, 38689, 4032, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 525, 12079, 4972, 535, 2197, 535, 1056, 7174, 17261, 730, 63688, 730, 9887, 533, 19331, 1706, 10012, 843, 3321, 3600, 38689, 4032, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 528, 12079, 4972, 535, 2197, 535, 4905, 7174, 17261, 730, 24163, 57506, 1892, 4174, 730, 1155, 514, 7614, 708, 4972, 535, 1826, 535, 1056, 7174, 17261, 730, 63688, 730, 10556, 526, 19331, 1706, 10012, 843, 3321, 3600, 38689, 4032, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 531, 12079, 4972, 535, 2142, 535, 1056, 7174, 17261, 730, 63688, 730, 10520, 527, 19331, 1706, 10012, 843, 3321, 3600, 38689, 4032, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 525, 12079, 4972, 535, 2142, 535, 1664, 7174, 17261, 730, 43497, 730, 532, 8323, 7925, 2835, 730, 530, 523, 534, 835, 843, 35436, 730, 533, 523, 525, 835, 4032, 23961, 6586, 730, 526, 523, 2346, 796, 5266, 708, 4972, 535, 1664, 535, 1056, 7174, 17261, 730, 63688, 730, 10426, 530, 19331, 1706, 10012, 843, 3321, 3600, 38689, 4032, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 528, 12079, 4972, 535, 2625, 535, 1056, 7174, 17261, 730, 8501, 6634, 968, 1267, 868, 2290, 535, 43497, 730, 532, 23961, 17828, 811, 874, 20425, 2030, 963, 730, 1155, 535, 1056, 536, 8227, 772, 42806, 811, 874, 5689, 1337, 8233, 20842, 708, 4972, 535, 2625, 535, 3654, 7174, 17261, 730, 8411, 886, 826, 526, 6840, 12465, 536, 1295, 5592, 16753, 592, 708, 4972, 535, 3024, 535, 1056, 7174, 17261, 730, 63688, 730, 8380, 533, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 525, 12079, 4972, 535, 2108, 535, 1056, 7174, 17261, 730, 63688, 730, 10426, 531, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 526, 12079, 4972, 535, 3337, 535, 1056, 7174, 17261, 730, 63688, 730, 11233, 534, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 527, 12079, 4972, 535, 3337, 535, 4536, 7174, 17261, 730, 49574, 886, 889, 529, 41172, 3999, 7410, 843, 22500, 730, 525, 523, 4613, 8061, 1015, 4972, 535, 3167, 535, 1056, 7174, 17261, 730, 63688, 730, 10520, 527, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 528, 12079, 4972, 535, 7058, 535, 1056, 7174, 17261, 730, 63688, 730, 8380, 529, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 529, 12079, 4972, 535, 7058, 535, 1417, 923, 32535, 730, 49883, 837, 526, 1295, 49321, 884, 61026, 730, 22218, 1532, 536, 61026, 3567, 811, 16687, 43497, 708, 4972, 535, 7568, 535, 1056, 7174, 17261, 730, 63688, 730, 10426, 526, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 530, 12079, 4528, 535, 1056, 535, 1056, 7174, 17261, 730, 63688, 730, 11233, 528, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 531, 12079, 4528, 535, 1056, 535, 1664, 7174, 17261, 730, 43497, 730, 532, 8323, 7925, 2835, 730, 530, 523, 533, 835, 843, 35436, 730, 533, 523, 525, 835, 4032, 23961, 6586, 730, 526, 523, 2346, 796, 5266, 708, 4528, 535, 6326, 535, 1056, 7174, 17261, 730, 63688, 730, 10426, 534, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 532, 12079, 4528, 535, 4528, 535, 1056, 7174, 17261, 730, 63688, 730, 8380, 527, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 525, 12079, 4528, 535, 4528, 535, 4536, 7174, 17261, 730, 49574, 886, 889, 529, 41172, 3999, 7410, 843, 22500, 730, 525, 523, 3756, 8061, 1015, 4528, 535, 1374, 535, 1056, 7174, 17261, 730, 63688, 730, 10426, 529, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 526, 12079, 4528, 535, 1374, 535, 1838, 7174, 17261, 730, 24163, 57506, 963, 730, 533, 514, 7614, 708, 4528, 535, 1443, 535, 1056, 7174, 17261, 730, 63688, 730, 11233, 532, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 527, 12079, 4528, 535, 898, 535, 1056, 7174, 17261, 730, 63688, 730, 10520, 530, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 528, 12079, 4528, 535, 898, 535, 1417, 923, 32535, 730, 49883, 837, 527, 1295, 49321, 884, 61026, 730, 28701, 1155, 536, 61026, 3567, 811, 16687, 43497, 708, 4528, 535, 1826, 535, 1056, 7174, 17261, 730, 63688, 730, 8380, 531, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 529, 12079, 4528, 535, 2138, 535, 1056, 7174, 17261, 730, 63688, 730, 11233, 525, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 530, 12079, 4528, 535, 2925, 535, 1056, 7174, 17261, 730, 63688, 730, 10426, 533, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 531, 12079, 4528, 535, 2925, 535, 4536, 7174, 17261, 730, 49574, 886, 889, 529, 41172, 3999, 7410, 843, 22500, 730, 525, 523, 3878, 8061, 1015, 4528, 535, 3024, 535, 1056, 7174, 17261, 730, 63688, 730, 8380, 526, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 532, 12079, 4528, 535, 2108, 535, 1056, 7174, 17261, 730, 63688, 730, 8380, 533, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 525, 12079, 4528, 535, 2108, 535, 1664, 7174, 17261, 730, 43497, 730, 532, 8323, 7925, 2835, 730, 530, 523, 534, 835, 843, 35436, 730, 533, 523, 525, 835, 4032, 23961, 6586, 730, 526, 523, 2346, 796, 5266, 708, 4528, 535, 3337, 535, 1056, 7174, 17261, 730, 63688, 730, 10426, 531, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 526, 12079, 4528, 535, 3167, 535, 1056, 7174, 17261, 730, 63688, 730, 11233, 534, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 527, 12079, 4528, 535, 3167, 535, 1417, 923, 32535, 730, 49883, 837, 526, 1295, 49321, 884, 61026, 730, 26465, 4972, 536, 61026, 3567, 811, 16687, 43497, 708, 4528, 535, 7058, 535, 1056, 7174, 17261, 730, 63688, 730, 10520, 527, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 528, 12079, 4528, 535, 7568, 535, 1056, 7174, 17261, 730, 63688, 730, 8380, 529, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 529, 12079, 4528, 535, 7568, 535, 4536, 7174, 17261, 730, 49574, 886, 889, 529, 41172, 3999, 7410, 843, 22500, 730, 525, 523, 2999, 8061, 1015, 4905, 535, 1056, 535, 1056, 7174, 17261, 730, 63688, 730, 10426, 526, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 530, 12079, 4905, 535, 1056, 535, 1838, 7174, 17261, 730, 24163, 57506, 963, 730, 530, 514, 7614, 708, 4905, 535, 6326, 535, 1056, 7174, 17261, 730, 63688, 730, 11233, 528, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 531, 12079, 4905, 535, 4528, 535, 1056, 7174, 17261, 730, 63688, 730, 10426, 534, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 532, 12079, 4905, 535, 1374, 535, 1056, 7174, 17261, 730, 63688, 730, 8380, 527, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 525, 12079, 4905, 535, 1443, 535, 1056, 7174, 17261, 730, 63688, 730, 10426, 529, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 526, 12079, 4905, 535, 1443, 535, 1417, 923, 32535, 730, 49883, 837, 527, 1295, 49321, 884, 61026, 730, 32522, 6326, 536, 61026, 3567, 811, 16687, 43497, 708, 4905, 535, 898, 535, 1056, 7174, 17261, 730, 63688, 730, 11233, 532, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 527, 12079, 4905, 535, 898, 535, 4536, 7174, 17261, 730, 49574, 886, 889, 529, 41172, 3999, 7410, 843, 22500, 730, 525, 523, 3092, 8061, 1015, 4905, 535, 898, 535, 1664, 7174, 17261, 730, 43497, 730, 532, 8323, 7925, 2835, 730, 531, 523, 525, 835, 843, 35436, 730, 533, 523, 525, 835, 4032, 23961, 6586, 730, 526, 523, 2346, 796, 5266, 708, 4905, 535, 1826, 535, 1056, 7174, 17261, 730, 63688, 730, 10520, 530, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 528, 12079, 4905, 535, 2138, 535, 1056, 7174, 17261, 730, 63688, 730, 8380, 531, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 529, 12079, 4905, 535, 2925, 535, 1056, 7174, 17261, 730, 63688, 730, 11233, 525, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 530, 12079, 4905, 535, 3024, 535, 1056, 7174, 17261, 730, 63688, 730, 10426, 533, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 531, 12079, 4905, 535, 2108, 535, 1056, 7174, 17261, 730, 63688, 730, 8380, 526, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 532, 12079, 4905, 535, 3337, 535, 1056, 7174, 17261, 730, 63688, 730, 8380, 533, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 525, 12079, 4905, 535, 3337, 535, 4536, 7174, 17261, 730, 49574, 886, 889, 529, 41172, 3999, 7410, 843, 22500, 730, 525, 523, 4613, 8061, 1015, 4905, 535, 3337, 535, 1417, 923, 32535, 730, 49883, 837, 526, 1295, 49321, 884, 61026, 730, 33231, 4541, 536, 61026, 3567, 811, 16687, 43497, 708, 4905, 535, 3167, 535, 1056, 7174, 17261, 730, 63688, 730, 10426, 531, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 526, 12079, 4905, 535, 3167, 535, 1838, 7174, 17261, 730, 24163, 57506, 963, 730, 532, 514, 7614, 708, 4905, 535, 7058, 535, 1056, 7174, 17261, 730, 63688, 730, 11233, 534, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 527, 12079, 4905, 535, 7568, 535, 1056, 7174, 17261, 730, 63688, 730, 10520, 527, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 528, 12079, 1155, 535, 1056, 535, 1056, 7174, 17261, 730, 63688, 730, 8380, 529, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 529, 12079, 1155, 535, 1056, 535, 1664, 7174, 17261, 730, 43497, 730, 532, 8323, 7925, 2835, 730, 530, 523, 533, 835, 843, 35436, 730, 533, 523, 525, 835, 4032, 23961, 6586, 730, 526, 523, 2346, 796, 5266, 708, 1155, 535, 1056, 535, 2108, 7174, 17261, 730, 43497, 730, 532, 23961, 17828, 20425, 968, 4582, 1055, 843, 802, 13573, 797, 779, 730, 4972, 535, 2625, 8501, 6634, 3271, 2638, 35036, 536, 1295, 32215, 4602, 708, 1155, 535, 6326, 535, 1056, 7174, 17261, 730, 63688, 730, 10426, 526, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 530, 12079, 1155, 535, 4528, 535, 1056, 7174, 17261, 730, 63688, 730, 11233, 528, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 531, 12079, 1155, 535, 4528, 535, 4536, 7174, 17261, 730, 49574, 886, 889, 529, 41172, 3999, 7410, 843, 22500, 730, 525, 523, 3756, 8061, 1015, 1155, 535, 1374, 535, 1056, 7174, 17261, 730, 63688, 730, 10426, 534, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 532, 12079, 1155, 535, 1374, 535, 1417, 923, 32535, 730, 49883, 837, 527, 1295, 49321, 884, 61026, 730, 24656, 3361, 536, 61026, 3567, 811, 16687, 43497, 708, 1155, 535, 1443, 535, 1056, 7174, 17261, 730, 63688, 730, 8380, 527, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 525, 12079, 1155, 535, 898, 535, 1056, 7174, 17261, 730, 63688, 730, 10426, 529, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 526, 12079, 1155, 535, 2197, 535, 3654, 7174, 17261, 730, 8411, 886, 826, 526, 6840, 12465, 536, 1295, 5592, 16753, 592, 18, 4023, 1787, 953, 779, 8411, 886, 7408, 13525, 540, 19, 16, 2119, 730, 525, 535, 1490, 730, 527, 5112, 20, 19, 16, 2119, 730, 526, 535, 730, 527, 811, 730, 530, 5112, 20, 19, 16, 2119, 730, 527, 535, 730, 530, 811, 730, 1155, 5112, 20, 19, 16, 2119, 730, 528, 535, 730, 1155, 811, 730, 898, 5112, 20, 19, 16, 2119, 730, 529, 535, 730, 898, 811, 730, 2108, 5112, 20, 19, 16, 2119, 730, 530, 535, 1391, 730, 2108, 5112, 20, 21 ], "markers": [ 3423, 3434, 3447, 3460, 3473, 3486 ], "logits_raw": [ -5.387547016143799, -3.613497495651245, -2.67159366607666, -1.864549160003662, 0.024332530796527863, -3.945512294769287 ], "probs_raw": [ 0.00351888220757246, 0.020742669701576233, 0.05320204421877861, 0.11924036592245102, 0.7884137034416199, 0.014882382936775684 ], "probs": [ 0.00351888220757246, 0.020742669701576233, 0.05320204421877861, 0.11924036592245102, 0.7884137034416199, 0.014882382936775684 ], "argmax_index": 4, "argmax": "4", "top2_margin": 0.6691733375191689, "near_tie": false, "gold": "4", "correct": true, "host_softmax_max_abs_dp": 7.450580596923828e-09, "batch": { "call": 0, "index": 0, "B": 4 } }, { "qid": "tension_10", "type": "noul", "K": 2, "calibrate": true, "temperature_key": "noul:2", "T": 1.6663223505020142, "prefix": 0, "n_ids": 3454, "positions": 3454, "bucket": 4096, "ids": [ 1, 17, 15991, 23707, 3613, 4425, 4857, 561, 33371, 12292, 38620, 1141, 61026, 17824, 730, 527, 521, 5008, 1985, 2647, 8411, 886, 826, 526, 521, 22694, 3446, 843, 1513, 31225, 1015, 27464, 535, 2960, 843, 6468, 535, 1056, 522, 1536, 535, 1056, 518, 767, 52574, 884, 15774, 535, 1267, 868, 2290, 767, 12003, 884, 1868, 535, 4582, 1055, 509, 6468, 535, 1056, 535, 1056, 7174, 17261, 730, 63688, 730, 10426, 527, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 526, 12079, 6468, 535, 6135, 535, 1056, 7174, 17261, 730, 63688, 730, 11233, 533, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 528, 12079, 6468, 535, 6468, 535, 1056, 7174, 17261, 730, 63688, 730, 10556, 526, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 530, 12079, 6468, 535, 4905, 535, 1056, 7174, 17261, 730, 63688, 730, 8380, 529, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 532, 12079, 6468, 535, 4905, 535, 4536, 7174, 17261, 730, 49574, 886, 889, 529, 41172, 3999, 7410, 843, 22500, 730, 525, 523, 3188, 8061, 1015, 6468, 535, 1374, 535, 1056, 7174, 17261, 730, 63688, 730, 10173, 534, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 526, 12079, 6468, 535, 1374, 535, 1417, 923, 32535, 730, 49883, 837, 527, 1295, 49321, 884, 61026, 730, 33118, 3375, 536, 61026, 3567, 811, 16687, 43497, 708, 6468, 535, 1386, 535, 1056, 7174, 17261, 730, 63688, 730, 10520, 525, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 528, 12079, 6468, 535, 1133, 535, 1056, 7174, 17261, 730, 63688, 730, 10426, 531, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 530, 12079, 6468, 535, 2197, 535, 1056, 7174, 17261, 730, 63688, 730, 8380, 527, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 532, 12079, 6468, 535, 1826, 535, 1056, 7174, 17261, 730, 63688, 730, 11233, 530, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 526, 12079, 6468, 535, 2142, 535, 1056, 7174, 17261, 730, 63688, 730, 10520, 532, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 528, 12079, 6468, 535, 2142, 535, 4536, 7174, 17261, 730, 49574, 886, 889, 529, 41172, 3999, 7410, 843, 22500, 730, 525, 523, 5489, 8061, 1015, 6468, 535, 1664, 535, 1056, 7174, 17261, 730, 63688, 730, 8380, 534, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 530, 12079, 6468, 535, 2625, 535, 1056, 7174, 17261, 730, 63688, 730, 10173, 528, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 532, 12079, 6468, 535, 2625, 535, 1417, 923, 32535, 730, 49883, 837, 527, 1295, 49321, 884, 61026, 730, 30664, 2108, 536, 61026, 3567, 811, 16687, 43497, 708, 6468, 535, 3024, 535, 1056, 7174, 17261, 730, 63688, 730, 10426, 533, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 526, 12079, 6468, 535, 3188, 535, 1056, 7174, 17261, 730, 63688, 730, 8380, 525, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 528, 12079, 6468, 535, 2108, 535, 1056, 7174, 17261, 730, 63688, 730, 11233, 529, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 528, 12079, 6468, 535, 4536, 535, 1374, 923, 32535, 730, 43497, 730, 532, 8323, 7925, 2835, 730, 533, 523, 534, 835, 843, 35436, 730, 533, 523, 525, 835, 518, 875, 730, 530, 775, 708, 6468, 535, 4536, 535, 5046, 923, 32535, 730, 43497, 730, 532, 8323, 7925, 2835, 730, 534, 523, 529, 835, 843, 35436, 730, 533, 523, 525, 835, 518, 875, 730, 1374, 775, 708, 6468, 535, 4344, 535, 2108, 923, 32535, 730, 43497, 730, 532, 23961, 6586, 730, 526, 523, 5370, 796, 5266, 843, 2390, 5723, 730, 526, 523, 2346, 796, 5266, 1015, 6468, 535, 4613, 535, 3654, 59687, 2850, 43497, 730, 532, 16017, 12365, 535, 11697, 38993, 578, 27493, 532, 522, 528, 26781, 875, 1209, 1399, 730, 528, 775, 708, 6468, 535, 4613, 535, 6468, 59687, 2850, 8411, 886, 826, 526, 13525, 968, 6155, 22694, 843, 7188, 535, 16017, 43497, 730, 532, 4032, 1113, 50027, 23993, 791, 708, 6468, 535, 4613, 535, 6468, 7174, 17261, 730, 1641, 13053, 19461, 1189, 60433, 536, 19331, 1706, 4908, 884, 24163, 57506, 708, 6468, 535, 4613, 535, 2108, 7174, 17261, 730, 16753, 28984, 968, 1267, 868, 2290, 963, 779, 2669, 26862, 708, 6468, 535, 3337, 535, 1386, 7174, 17261, 730, 12003, 1868, 9388, 811, 4582, 1055, 843, 589, 60088, 730, 526, 1015, 6468, 535, 4195, 535, 5843, 7174, 17261, 730, 49883, 837, 526, 810, 837, 527, 54126, 843, 2744, 19331, 1706, 4032, 2515, 29067, 7410, 708, 6468, 535, 3745, 535, 1664, 7174, 17261, 730, 24163, 57506, 963, 730, 2752, 514, 7614, 708, 6468, 535, 1838, 535, 1374, 7174, 17261, 730, 4582, 1055, 884, 3475, 963, 43497, 730, 532, 536, 11023, 17652, 6944, 811, 43497, 730, 532, 8323, 708, 6468, 535, 7058, 535, 3167, 7174, 17261, 730, 24163, 57506, 963, 730, 9731, 514, 7614, 536, 1641, 13053, 18483, 730, 528, 810, 730, 529, 8557, 811, 1361, 19082, 871, 708, 6468, 535, 3878, 535, 898, 7174, 17261, 730, 16017, 3360, 4468, 968, 4582, 1055, 535, 38169, 8227, 772, 28782, 4563, 1490, 779, 43497, 730, 532, 23961, 963, 779, 2341, 46129, 708, 6468, 535, 5046, 535, 4536, 7174, 17261, 730, 8227, 772, 8365, 988, 43497, 730, 532, 536, 46129, 909, 532, 522, 527, 2083, 7937, 521, 1295, 5896, 1833, 708, 4972, 535, 1056, 535, 3654, 7174, 17261, 730, 43497, 730, 532, 23961, 17828, 8744, 4174, 35036, 536, 32215, 4984, 708, 4972, 535, 6135, 535, 1374, 7174, 17261, 730, 43497, 730, 532, 23961, 17828, 23848, 536, 1927, 2685, 963, 2920, 6586, 730, 525, 523, 530, 796, 5266, 875, 730, 2198, 775, 7410, 708, 4972, 535, 6326, 535, 1698, 7174, 17261, 730, 11023, 17652, 8365, 988, 43497, 730, 532, 8323, 708, 4972, 535, 3654, 535, 1664, 7174, 17261, 730, 8411, 886, 826, 526, 2280, 33480, 968, 1267, 868, 2290, 536, 878, 4318, 1047, 730, 2019, 1750, 730, 1826, 775, 708, 4972, 535, 3654, 535, 2999, 7174, 17261, 730, 1641, 13053, 19461, 1189, 49826, 536, 1641, 13053, 18483, 730, 528, 810, 730, 529, 840, 59188, 708, 4972, 535, 6468, 535, 1056, 7174, 17261, 730, 43497, 730, 532, 8323, 7925, 2835, 730, 531, 523, 526, 835, 843, 35436, 730, 533, 523, 525, 835, 1015, 4972, 535, 4905, 535, 1056, 7174, 17261, 730, 63688, 730, 4787, 525, 19331, 1706, 10012, 843, 3321, 3600, 38689, 4032, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 525, 12079, 4972, 535, 1374, 535, 1056, 7174, 17261, 730, 63688, 730, 3818, 531, 19331, 1706, 10012, 843, 3321, 3600, 38689, 4032, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 528, 12079, 4972, 535, 1386, 535, 1056, 7174, 17261, 730, 63688, 730, 4420, 526, 19331, 1706, 10012, 843, 3321, 3600, 38689, 4032, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 531, 12079, 4972, 535, 1386, 535, 3337, 923, 32535, 730, 49883, 837, 526, 1295, 49321, 884, 61026, 730, 32928, 2476, 536, 61026, 3567, 811, 16687, 43497, 708, 4972, 535, 1133, 535, 1056, 7174, 17261, 730, 63688, 730, 6583, 527, 19331, 1706, 10012, 843, 3321, 3600, 38689, 4032, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 525, 12079, 4972, 535, 2197, 535, 1056, 7174, 17261, 730, 63688, 730, 9887, 533, 19331, 1706, 10012, 843, 3321, 3600, 38689, 4032, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 528, 12079, 4972, 535, 2197, 535, 4905, 7174, 17261, 730, 24163, 57506, 1892, 4174, 730, 1155, 514, 7614, 708, 4972, 535, 1826, 535, 1056, 7174, 17261, 730, 63688, 730, 10556, 526, 19331, 1706, 10012, 843, 3321, 3600, 38689, 4032, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 531, 12079, 4972, 535, 2142, 535, 1056, 7174, 17261, 730, 63688, 730, 10520, 527, 19331, 1706, 10012, 843, 3321, 3600, 38689, 4032, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 525, 12079, 4972, 535, 2142, 535, 1664, 7174, 17261, 730, 43497, 730, 532, 8323, 7925, 2835, 730, 530, 523, 534, 835, 843, 35436, 730, 533, 523, 525, 835, 4032, 23961, 6586, 730, 526, 523, 2346, 796, 5266, 708, 4972, 535, 1664, 535, 1056, 7174, 17261, 730, 63688, 730, 10426, 530, 19331, 1706, 10012, 843, 3321, 3600, 38689, 4032, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 528, 12079, 4972, 535, 2625, 535, 1056, 7174, 17261, 730, 8501, 6634, 968, 1267, 868, 2290, 535, 43497, 730, 532, 23961, 17828, 811, 874, 20425, 2030, 963, 730, 1155, 535, 1056, 536, 8227, 772, 42806, 811, 874, 5689, 1337, 8233, 20842, 708, 4972, 535, 2625, 535, 3654, 7174, 17261, 730, 8411, 886, 826, 526, 6840, 12465, 536, 1295, 5592, 16753, 592, 708, 4972, 535, 3024, 535, 1056, 7174, 17261, 730, 63688, 730, 8380, 533, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 525, 12079, 4972, 535, 2108, 535, 1056, 7174, 17261, 730, 63688, 730, 10426, 531, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 526, 12079, 4972, 535, 3337, 535, 1056, 7174, 17261, 730, 63688, 730, 11233, 534, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 527, 12079, 4972, 535, 3337, 535, 4536, 7174, 17261, 730, 49574, 886, 889, 529, 41172, 3999, 7410, 843, 22500, 730, 525, 523, 4613, 8061, 1015, 4972, 535, 3167, 535, 1056, 7174, 17261, 730, 63688, 730, 10520, 527, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 528, 12079, 4972, 535, 7058, 535, 1056, 7174, 17261, 730, 63688, 730, 8380, 529, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 529, 12079, 4972, 535, 7058, 535, 1417, 923, 32535, 730, 49883, 837, 526, 1295, 49321, 884, 61026, 730, 22218, 1532, 536, 61026, 3567, 811, 16687, 43497, 708, 4972, 535, 7568, 535, 1056, 7174, 17261, 730, 63688, 730, 10426, 526, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 530, 12079, 4528, 535, 1056, 535, 1056, 7174, 17261, 730, 63688, 730, 11233, 528, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 531, 12079, 4528, 535, 1056, 535, 1664, 7174, 17261, 730, 43497, 730, 532, 8323, 7925, 2835, 730, 530, 523, 533, 835, 843, 35436, 730, 533, 523, 525, 835, 4032, 23961, 6586, 730, 526, 523, 2346, 796, 5266, 708, 4528, 535, 6326, 535, 1056, 7174, 17261, 730, 63688, 730, 10426, 534, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 532, 12079, 4528, 535, 4528, 535, 1056, 7174, 17261, 730, 63688, 730, 8380, 527, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 525, 12079, 4528, 535, 4528, 535, 4536, 7174, 17261, 730, 49574, 886, 889, 529, 41172, 3999, 7410, 843, 22500, 730, 525, 523, 3756, 8061, 1015, 4528, 535, 1374, 535, 1056, 7174, 17261, 730, 63688, 730, 10426, 529, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 526, 12079, 4528, 535, 1374, 535, 1838, 7174, 17261, 730, 24163, 57506, 963, 730, 533, 514, 7614, 708, 4528, 535, 1443, 535, 1056, 7174, 17261, 730, 63688, 730, 11233, 532, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 527, 12079, 4528, 535, 898, 535, 1056, 7174, 17261, 730, 63688, 730, 10520, 530, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 528, 12079, 4528, 535, 898, 535, 1417, 923, 32535, 730, 49883, 837, 527, 1295, 49321, 884, 61026, 730, 28701, 1155, 536, 61026, 3567, 811, 16687, 43497, 708, 4528, 535, 1826, 535, 1056, 7174, 17261, 730, 63688, 730, 8380, 531, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 529, 12079, 4528, 535, 2138, 535, 1056, 7174, 17261, 730, 63688, 730, 11233, 525, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 530, 12079, 4528, 535, 2925, 535, 1056, 7174, 17261, 730, 63688, 730, 10426, 533, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 531, 12079, 4528, 535, 2925, 535, 4536, 7174, 17261, 730, 49574, 886, 889, 529, 41172, 3999, 7410, 843, 22500, 730, 525, 523, 3878, 8061, 1015, 4528, 535, 3024, 535, 1056, 7174, 17261, 730, 63688, 730, 8380, 526, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 532, 12079, 4528, 535, 2108, 535, 1056, 7174, 17261, 730, 63688, 730, 8380, 533, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 525, 12079, 4528, 535, 2108, 535, 1664, 7174, 17261, 730, 43497, 730, 532, 8323, 7925, 2835, 730, 530, 523, 534, 835, 843, 35436, 730, 533, 523, 525, 835, 4032, 23961, 6586, 730, 526, 523, 2346, 796, 5266, 708, 4528, 535, 3337, 535, 1056, 7174, 17261, 730, 63688, 730, 10426, 531, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 526, 12079, 4528, 535, 3167, 535, 1056, 7174, 17261, 730, 63688, 730, 11233, 534, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 527, 12079, 4528, 535, 3167, 535, 1417, 923, 32535, 730, 49883, 837, 526, 1295, 49321, 884, 61026, 730, 26465, 4972, 536, 61026, 3567, 811, 16687, 43497, 708, 4528, 535, 7058, 535, 1056, 7174, 17261, 730, 63688, 730, 10520, 527, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 528, 12079, 4528, 535, 7568, 535, 1056, 7174, 17261, 730, 63688, 730, 8380, 529, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 529, 12079, 4528, 535, 7568, 535, 4536, 7174, 17261, 730, 49574, 886, 889, 529, 41172, 3999, 7410, 843, 22500, 730, 525, 523, 2999, 8061, 1015, 4905, 535, 1056, 535, 1056, 7174, 17261, 730, 63688, 730, 10426, 526, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 530, 12079, 4905, 535, 1056, 535, 1838, 7174, 17261, 730, 24163, 57506, 963, 730, 530, 514, 7614, 708, 4905, 535, 6326, 535, 1056, 7174, 17261, 730, 63688, 730, 11233, 528, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 531, 12079, 4905, 535, 4528, 535, 1056, 7174, 17261, 730, 63688, 730, 10426, 534, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 532, 12079, 4905, 535, 1374, 535, 1056, 7174, 17261, 730, 63688, 730, 8380, 527, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 525, 12079, 4905, 535, 1443, 535, 1056, 7174, 17261, 730, 63688, 730, 10426, 529, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 526, 12079, 4905, 535, 1443, 535, 1417, 923, 32535, 730, 49883, 837, 527, 1295, 49321, 884, 61026, 730, 32522, 6326, 536, 61026, 3567, 811, 16687, 43497, 708, 4905, 535, 898, 535, 1056, 7174, 17261, 730, 63688, 730, 11233, 532, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 527, 12079, 4905, 535, 898, 535, 4536, 7174, 17261, 730, 49574, 886, 889, 529, 41172, 3999, 7410, 843, 22500, 730, 525, 523, 3092, 8061, 1015, 4905, 535, 898, 535, 1664, 7174, 17261, 730, 43497, 730, 532, 8323, 7925, 2835, 730, 531, 523, 525, 835, 843, 35436, 730, 533, 523, 525, 835, 4032, 23961, 6586, 730, 526, 523, 2346, 796, 5266, 708, 4905, 535, 1826, 535, 1056, 7174, 17261, 730, 63688, 730, 10520, 530, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 528, 12079, 4905, 535, 2138, 535, 1056, 7174, 17261, 730, 63688, 730, 8380, 531, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 529, 12079, 4905, 535, 2925, 535, 1056, 7174, 17261, 730, 63688, 730, 11233, 525, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 530, 12079, 4905, 535, 3024, 535, 1056, 7174, 17261, 730, 63688, 730, 10426, 533, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 531, 12079, 4905, 535, 2108, 535, 1056, 7174, 17261, 730, 63688, 730, 8380, 526, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 532, 12079, 4905, 535, 3337, 535, 1056, 7174, 17261, 730, 63688, 730, 8380, 533, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 525, 12079, 4905, 535, 3337, 535, 4536, 7174, 17261, 730, 49574, 886, 889, 529, 41172, 3999, 7410, 843, 22500, 730, 525, 523, 4613, 8061, 1015, 4905, 535, 3337, 535, 1417, 923, 32535, 730, 49883, 837, 526, 1295, 49321, 884, 61026, 730, 33231, 4541, 536, 61026, 3567, 811, 16687, 43497, 708, 4905, 535, 3167, 535, 1056, 7174, 17261, 730, 63688, 730, 10426, 531, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 526, 12079, 4905, 535, 3167, 535, 1838, 7174, 17261, 730, 24163, 57506, 963, 730, 532, 514, 7614, 708, 4905, 535, 7058, 535, 1056, 7174, 17261, 730, 63688, 730, 11233, 534, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 527, 12079, 4905, 535, 7568, 535, 1056, 7174, 17261, 730, 63688, 730, 10520, 527, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 528, 12079, 1155, 535, 1056, 535, 1056, 7174, 17261, 730, 63688, 730, 8380, 529, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 529, 12079, 1155, 535, 1056, 535, 1664, 7174, 17261, 730, 43497, 730, 532, 8323, 7925, 2835, 730, 530, 523, 533, 835, 843, 35436, 730, 533, 523, 525, 835, 4032, 23961, 6586, 730, 526, 523, 2346, 796, 5266, 708, 1155, 535, 1056, 535, 2108, 7174, 17261, 730, 43497, 730, 532, 23961, 17828, 20425, 968, 4582, 1055, 843, 802, 13573, 797, 779, 730, 4972, 535, 2625, 8501, 6634, 3271, 2638, 35036, 536, 1295, 32215, 4602, 708, 1155, 535, 6326, 535, 1056, 7174, 17261, 730, 63688, 730, 10426, 526, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 530, 12079, 1155, 535, 4528, 535, 1056, 7174, 17261, 730, 63688, 730, 11233, 528, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 531, 12079, 1155, 535, 4528, 535, 4536, 7174, 17261, 730, 49574, 886, 889, 529, 41172, 3999, 7410, 843, 22500, 730, 525, 523, 3756, 8061, 1015, 1155, 535, 1374, 535, 1056, 7174, 17261, 730, 63688, 730, 10426, 534, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 532, 12079, 1155, 535, 1374, 535, 1417, 923, 32535, 730, 49883, 837, 527, 1295, 49321, 884, 61026, 730, 24656, 3361, 536, 61026, 3567, 811, 16687, 43497, 708, 1155, 535, 1443, 535, 1056, 7174, 17261, 730, 63688, 730, 8380, 527, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 525, 12079, 1155, 535, 898, 535, 1056, 7174, 17261, 730, 63688, 730, 10426, 529, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 526, 12079, 1155, 535, 2197, 535, 3654, 7174, 17261, 730, 8411, 886, 826, 526, 6840, 12465, 536, 1295, 5592, 16753, 592, 18, 5257, 779, 730, 1155, 535, 1056, 3999, 521, 953, 779, 43497, 730, 532, 23961, 17828, 2638, 35036, 540, 19, 16, 4831, 535, 1295, 521, 779, 8294, 2177, 1014, 3965, 20, 19, 16, 3347, 535, 18672, 521, 779, 8294, 10995, 20, 21 ], "markers": [ 3432, 3444 ], "logits_raw": [ 0.22075581550598145, 0.5800595879554749 ], "probs_raw": [ 0.41112810373306274, 0.5888718366622925 ], "probs": [ 0.5536988377571106, 0.4463012218475342 ], "argmax_index": 0, "argmax": "true", "top2_margin": 0.10739761590957642, "near_tie": false, "gold": "true", "correct": true, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 2, "B": 4 } } ], "response": { "answers": { "component": { "type": "choice", "choice": "diverter", "confidence": 0.4949849545955658, "probabilities": { "scanner": 0.048573244363069534, "lane_7": 0.29361486434936523, "diverter": 0.4949849545955658, "power": 0.10699225217103958, "software": 0.05583466589450836 } }, "normal_at_end": { "type": "noul", "noul": 0.6327569484710693 }, "stoppage": { "type": "score", "score": 3.7129345843568444, "confidence": 0.7884137034416199, "probabilities": { "0": 0.00351888220757246, "1": 0.020742669701576233, "2": 0.05320204421877861, "3": 0.11924036592245102, "4": 0.7884137034416199, "5": 0.014882382936775684 }, "legend": { "0": "under 2 minutes", "1": "2 to 5 minutes", "2": "5 to 10 minutes", "3": "10 to 20 minutes", "4": "20 to 40 minutes", "5": "over 40 minutes" } }, "tension_10": { "type": "noul", "noul": 0.5536988377571106 } }, "usage": { "input_tokens": 13867, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "own_L01", "mode": "text", "source": "own_long", "native": true, "prefix": 0, "usage_input_tokens": 4830, "questions": [ { "qid": "component", "type": "choice", "K": 5, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 1606, "positions": 1606, "bucket": 2048, "ids": [ 1, 17, 15991, 23707, 3613, 4425, 4857, 561, 33371, 12292, 38620, 1141, 61026, 17824, 730, 527, 521, 5008, 1985, 2647, 8411, 886, 826, 526, 521, 22694, 3446, 843, 1513, 31225, 1015, 27464, 535, 2960, 843, 6468, 535, 1056, 522, 1536, 535, 1056, 518, 767, 52574, 884, 15774, 535, 1267, 868, 2290, 767, 12003, 884, 1868, 535, 4582, 1055, 509, 6468, 535, 1056, 535, 1056, 7174, 17261, 730, 63688, 730, 10426, 527, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 526, 12079, 6468, 535, 6135, 535, 1056, 7174, 17261, 730, 63688, 730, 11233, 533, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 528, 12079, 6468, 535, 6468, 535, 1056, 7174, 17261, 730, 63688, 730, 10556, 526, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 530, 12079, 6468, 535, 4905, 535, 1056, 7174, 17261, 730, 63688, 730, 8380, 529, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 532, 12079, 6468, 535, 4905, 535, 4536, 7174, 17261, 730, 49574, 886, 889, 529, 41172, 3999, 7410, 843, 22500, 730, 525, 523, 3188, 8061, 1015, 6468, 535, 1374, 535, 1056, 7174, 17261, 730, 63688, 730, 10173, 534, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 526, 12079, 6468, 535, 1374, 535, 1417, 923, 32535, 730, 49883, 837, 527, 1295, 49321, 884, 61026, 730, 33118, 3375, 536, 61026, 3567, 811, 16687, 43497, 708, 6468, 535, 1386, 535, 1056, 7174, 17261, 730, 63688, 730, 10520, 525, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 528, 12079, 6468, 535, 1133, 535, 1056, 7174, 17261, 730, 63688, 730, 10426, 531, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 530, 12079, 6468, 535, 2197, 535, 1056, 7174, 17261, 730, 63688, 730, 8380, 527, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 532, 12079, 6468, 535, 1826, 535, 1056, 7174, 17261, 730, 63688, 730, 11233, 530, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 526, 12079, 6468, 535, 2142, 535, 1056, 7174, 17261, 730, 63688, 730, 10520, 532, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 528, 12079, 6468, 535, 2142, 535, 4536, 7174, 17261, 730, 49574, 886, 889, 529, 41172, 3999, 7410, 843, 22500, 730, 525, 523, 5489, 8061, 1015, 6468, 535, 1664, 535, 1056, 7174, 17261, 730, 63688, 730, 8380, 534, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 530, 12079, 6468, 535, 2625, 535, 1056, 7174, 17261, 730, 63688, 730, 10173, 528, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 532, 12079, 6468, 535, 2625, 535, 1417, 923, 32535, 730, 49883, 837, 527, 1295, 49321, 884, 61026, 730, 30664, 2108, 536, 61026, 3567, 811, 16687, 43497, 708, 6468, 535, 3024, 535, 1056, 7174, 17261, 730, 63688, 730, 10426, 533, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 526, 12079, 6468, 535, 3188, 535, 1056, 7174, 17261, 730, 63688, 730, 8380, 525, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 528, 12079, 6468, 535, 2108, 535, 1056, 7174, 17261, 730, 63688, 730, 11233, 529, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 528, 12079, 6468, 535, 4536, 535, 1374, 923, 32535, 730, 43497, 730, 532, 8323, 7925, 2835, 730, 533, 523, 534, 835, 843, 35436, 730, 533, 523, 525, 835, 518, 875, 730, 530, 775, 708, 6468, 535, 4536, 535, 5046, 923, 32535, 730, 43497, 730, 532, 8323, 7925, 2835, 730, 534, 523, 529, 835, 843, 35436, 730, 533, 523, 525, 835, 518, 875, 730, 1374, 775, 708, 6468, 535, 4344, 535, 2108, 923, 32535, 730, 43497, 730, 532, 23961, 6586, 730, 526, 523, 5370, 796, 5266, 843, 2390, 5723, 730, 526, 523, 2346, 796, 5266, 1015, 6468, 535, 4613, 535, 3654, 59687, 2850, 43497, 730, 532, 16017, 12365, 535, 11697, 38993, 578, 27493, 532, 522, 528, 26781, 875, 1209, 1399, 730, 528, 775, 708, 6468, 535, 4613, 535, 6468, 59687, 2850, 8411, 886, 826, 526, 13525, 968, 6155, 22694, 843, 7188, 535, 16017, 43497, 730, 532, 4032, 1113, 50027, 23993, 791, 708, 6468, 535, 4613, 535, 6468, 7174, 17261, 730, 1641, 13053, 19461, 1189, 60433, 536, 19331, 1706, 4908, 884, 24163, 57506, 708, 6468, 535, 4613, 535, 2108, 7174, 17261, 730, 16753, 28984, 968, 1267, 868, 2290, 963, 779, 2669, 26862, 708, 6468, 535, 3337, 535, 1386, 7174, 17261, 730, 12003, 1868, 9388, 811, 4582, 1055, 843, 589, 60088, 730, 526, 1015, 6468, 535, 4195, 535, 5843, 7174, 17261, 730, 49883, 837, 526, 810, 837, 527, 54126, 843, 2744, 19331, 1706, 4032, 2515, 29067, 7410, 708, 6468, 535, 3745, 535, 1664, 7174, 17261, 730, 24163, 57506, 963, 730, 2752, 514, 7614, 708, 6468, 535, 1838, 535, 1374, 7174, 17261, 730, 4582, 1055, 884, 3475, 963, 43497, 730, 532, 536, 11023, 17652, 6944, 811, 43497, 730, 532, 8323, 708, 6468, 535, 7058, 535, 3167, 7174, 17261, 730, 24163, 57506, 963, 730, 9731, 514, 7614, 536, 1641, 13053, 18483, 730, 528, 810, 730, 529, 8557, 811, 1361, 19082, 871, 708, 6468, 535, 3878, 535, 898, 7174, 17261, 730, 16017, 3360, 4468, 968, 4582, 1055, 535, 38169, 8227, 772, 28782, 4563, 1490, 779, 43497, 730, 532, 23961, 963, 779, 2341, 46129, 708, 6468, 535, 5046, 535, 4536, 7174, 17261, 730, 8227, 772, 8365, 988, 43497, 730, 532, 536, 46129, 909, 532, 522, 527, 2083, 7937, 521, 1295, 5896, 1833, 708, 4972, 535, 1056, 535, 3654, 7174, 17261, 730, 43497, 730, 532, 23961, 17828, 8744, 4174, 35036, 536, 32215, 4984, 708, 4972, 535, 6135, 535, 1374, 7174, 17261, 730, 43497, 730, 532, 23961, 17828, 23848, 536, 1927, 2685, 963, 2920, 6586, 730, 525, 523, 530, 796, 5266, 875, 730, 2198, 775, 7410, 708, 4972, 535, 6326, 535, 1698, 7174, 17261, 730, 11023, 17652, 8365, 988, 43497, 730, 532, 8323, 708, 4972, 535, 3654, 535, 1664, 7174, 17261, 730, 8411, 886, 826, 526, 2280, 33480, 968, 1267, 868, 2290, 536, 878, 4318, 1047, 730, 2019, 1750, 730, 1826, 775, 708, 4972, 535, 3654, 535, 2999, 7174, 17261, 730, 1641, 13053, 19461, 1189, 49826, 536, 1641, 13053, 18483, 730, 528, 810, 730, 529, 840, 59188, 708, 4972, 535, 6468, 535, 1056, 7174, 17261, 730, 43497, 730, 532, 8323, 7925, 2835, 730, 531, 523, 526, 835, 843, 35436, 730, 533, 523, 525, 835, 1015, 4972, 535, 4905, 535, 1056, 7174, 17261, 730, 63688, 730, 4787, 525, 19331, 1706, 10012, 843, 3321, 3600, 38689, 4032, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 525, 12079, 4972, 535, 1374, 535, 1056, 7174, 17261, 730, 63688, 730, 3818, 531, 19331, 1706, 10012, 843, 3321, 3600, 38689, 4032, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 528, 12079, 4972, 535, 1386, 535, 1056, 7174, 17261, 730, 63688, 730, 4420, 526, 19331, 1706, 10012, 843, 3321, 3600, 38689, 4032, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 531, 12079, 4972, 535, 1386, 535, 3337, 923, 32535, 730, 49883, 837, 526, 1295, 49321, 884, 61026, 730, 32928, 2476, 536, 61026, 3567, 811, 16687, 43497, 708, 4972, 535, 1133, 535, 1056, 7174, 17261, 730, 63688, 730, 6583, 527, 19331, 1706, 10012, 843, 3321, 3600, 38689, 4032, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 525, 12079, 4972, 535, 2197, 535, 1056, 7174, 17261, 730, 63688, 730, 9887, 533, 19331, 1706, 10012, 843, 3321, 3600, 38689, 4032, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 528, 12079, 4972, 535, 2197, 535, 4905, 7174, 17261, 730, 24163, 57506, 1892, 4174, 730, 1155, 514, 7614, 708, 4972, 535, 1826, 535, 1056, 7174, 17261, 730, 63688, 730, 10556, 526, 19331, 1706, 10012, 843, 3321, 3600, 38689, 4032, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 531, 12079, 4972, 535, 2142, 535, 1056, 7174, 17261, 730, 63688, 730, 10520, 527, 19331, 1706, 10012, 843, 3321, 3600, 38689, 4032, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 525, 12079, 4972, 535, 2142, 535, 1664, 7174, 17261, 730, 43497, 730, 532, 8323, 7925, 2835, 730, 530, 523, 534, 835, 843, 35436, 730, 533, 523, 525, 835, 4032, 23961, 6586, 730, 526, 523, 2346, 796, 5266, 708, 4972, 535, 1664, 535, 1056, 7174, 17261, 730, 63688, 730, 10426, 530, 19331, 1706, 10012, 843, 3321, 3600, 38689, 4032, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 528, 12079, 4972, 535, 2625, 535, 1056, 7174, 17261, 730, 8501, 6634, 968, 1267, 868, 2290, 535, 43497, 730, 532, 23961, 17828, 811, 874, 20425, 2030, 963, 730, 1155, 535, 1056, 536, 8227, 772, 42806, 811, 874, 5689, 1337, 8233, 20842, 708, 4972, 535, 2625, 535, 3654, 7174, 17261, 730, 8411, 886, 826, 526, 6840, 12465, 536, 1295, 5592, 16753, 592, 18, 28828, 8150, 5457, 779, 878, 4318, 1047, 540, 19, 16, 49883, 535, 3492, 3955, 49883, 20, 19, 16, 43497, 572, 532, 535, 2020, 4674, 777, 43497, 730, 532, 20, 19, 16, 49574, 886, 535, 889, 1991, 886, 5716, 20, 19, 16, 2410, 535, 9642, 7018, 20, 19, 16, 5901, 535, 10478, 5901, 20, 21 ], "markers": [ 1561, 1569, 1582, 1592, 1599 ], "logits_raw": [ -0.840394139289856, 1.6203656196594238, 0.25618529319763184, -2.4982593059539795, -3.887787103652954 ], "probs_raw": [ 0.06271307170391083, 0.7346024513244629, 0.18775710463523865, 0.011949683539569378, 0.0029777754098176956 ], "probs": [ 0.10629419982433319, 0.6164789795875549, 0.23265084624290466, 0.032522596418857574, 0.012053357437252998 ], "argmax_index": 1, "argmax": "lane_7", "top2_margin": 0.38382813334465027, "near_tie": false, "gold": "lane_7", "correct": true, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 1, "B": 3 } }, { "qid": "normal_at_end", "type": "noul", "K": 2, "calibrate": true, "temperature_key": "noul:2", "T": 1.6663223505020142, "prefix": 0, "n_ids": 1588, "positions": 1588, "bucket": 2048, "ids": [ 1, 17, 15991, 23707, 3613, 4425, 4857, 561, 33371, 12292, 38620, 1141, 61026, 17824, 730, 527, 521, 5008, 1985, 2647, 8411, 886, 826, 526, 521, 22694, 3446, 843, 1513, 31225, 1015, 27464, 535, 2960, 843, 6468, 535, 1056, 522, 1536, 535, 1056, 518, 767, 52574, 884, 15774, 535, 1267, 868, 2290, 767, 12003, 884, 1868, 535, 4582, 1055, 509, 6468, 535, 1056, 535, 1056, 7174, 17261, 730, 63688, 730, 10426, 527, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 526, 12079, 6468, 535, 6135, 535, 1056, 7174, 17261, 730, 63688, 730, 11233, 533, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 528, 12079, 6468, 535, 6468, 535, 1056, 7174, 17261, 730, 63688, 730, 10556, 526, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 530, 12079, 6468, 535, 4905, 535, 1056, 7174, 17261, 730, 63688, 730, 8380, 529, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 532, 12079, 6468, 535, 4905, 535, 4536, 7174, 17261, 730, 49574, 886, 889, 529, 41172, 3999, 7410, 843, 22500, 730, 525, 523, 3188, 8061, 1015, 6468, 535, 1374, 535, 1056, 7174, 17261, 730, 63688, 730, 10173, 534, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 526, 12079, 6468, 535, 1374, 535, 1417, 923, 32535, 730, 49883, 837, 527, 1295, 49321, 884, 61026, 730, 33118, 3375, 536, 61026, 3567, 811, 16687, 43497, 708, 6468, 535, 1386, 535, 1056, 7174, 17261, 730, 63688, 730, 10520, 525, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 528, 12079, 6468, 535, 1133, 535, 1056, 7174, 17261, 730, 63688, 730, 10426, 531, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 530, 12079, 6468, 535, 2197, 535, 1056, 7174, 17261, 730, 63688, 730, 8380, 527, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 532, 12079, 6468, 535, 1826, 535, 1056, 7174, 17261, 730, 63688, 730, 11233, 530, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 526, 12079, 6468, 535, 2142, 535, 1056, 7174, 17261, 730, 63688, 730, 10520, 532, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 528, 12079, 6468, 535, 2142, 535, 4536, 7174, 17261, 730, 49574, 886, 889, 529, 41172, 3999, 7410, 843, 22500, 730, 525, 523, 5489, 8061, 1015, 6468, 535, 1664, 535, 1056, 7174, 17261, 730, 63688, 730, 8380, 534, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 530, 12079, 6468, 535, 2625, 535, 1056, 7174, 17261, 730, 63688, 730, 10173, 528, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 532, 12079, 6468, 535, 2625, 535, 1417, 923, 32535, 730, 49883, 837, 527, 1295, 49321, 884, 61026, 730, 30664, 2108, 536, 61026, 3567, 811, 16687, 43497, 708, 6468, 535, 3024, 535, 1056, 7174, 17261, 730, 63688, 730, 10426, 533, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 526, 12079, 6468, 535, 3188, 535, 1056, 7174, 17261, 730, 63688, 730, 8380, 525, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 528, 12079, 6468, 535, 2108, 535, 1056, 7174, 17261, 730, 63688, 730, 11233, 529, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 528, 12079, 6468, 535, 4536, 535, 1374, 923, 32535, 730, 43497, 730, 532, 8323, 7925, 2835, 730, 533, 523, 534, 835, 843, 35436, 730, 533, 523, 525, 835, 518, 875, 730, 530, 775, 708, 6468, 535, 4536, 535, 5046, 923, 32535, 730, 43497, 730, 532, 8323, 7925, 2835, 730, 534, 523, 529, 835, 843, 35436, 730, 533, 523, 525, 835, 518, 875, 730, 1374, 775, 708, 6468, 535, 4344, 535, 2108, 923, 32535, 730, 43497, 730, 532, 23961, 6586, 730, 526, 523, 5370, 796, 5266, 843, 2390, 5723, 730, 526, 523, 2346, 796, 5266, 1015, 6468, 535, 4613, 535, 3654, 59687, 2850, 43497, 730, 532, 16017, 12365, 535, 11697, 38993, 578, 27493, 532, 522, 528, 26781, 875, 1209, 1399, 730, 528, 775, 708, 6468, 535, 4613, 535, 6468, 59687, 2850, 8411, 886, 826, 526, 13525, 968, 6155, 22694, 843, 7188, 535, 16017, 43497, 730, 532, 4032, 1113, 50027, 23993, 791, 708, 6468, 535, 4613, 535, 6468, 7174, 17261, 730, 1641, 13053, 19461, 1189, 60433, 536, 19331, 1706, 4908, 884, 24163, 57506, 708, 6468, 535, 4613, 535, 2108, 7174, 17261, 730, 16753, 28984, 968, 1267, 868, 2290, 963, 779, 2669, 26862, 708, 6468, 535, 3337, 535, 1386, 7174, 17261, 730, 12003, 1868, 9388, 811, 4582, 1055, 843, 589, 60088, 730, 526, 1015, 6468, 535, 4195, 535, 5843, 7174, 17261, 730, 49883, 837, 526, 810, 837, 527, 54126, 843, 2744, 19331, 1706, 4032, 2515, 29067, 7410, 708, 6468, 535, 3745, 535, 1664, 7174, 17261, 730, 24163, 57506, 963, 730, 2752, 514, 7614, 708, 6468, 535, 1838, 535, 1374, 7174, 17261, 730, 4582, 1055, 884, 3475, 963, 43497, 730, 532, 536, 11023, 17652, 6944, 811, 43497, 730, 532, 8323, 708, 6468, 535, 7058, 535, 3167, 7174, 17261, 730, 24163, 57506, 963, 730, 9731, 514, 7614, 536, 1641, 13053, 18483, 730, 528, 810, 730, 529, 8557, 811, 1361, 19082, 871, 708, 6468, 535, 3878, 535, 898, 7174, 17261, 730, 16017, 3360, 4468, 968, 4582, 1055, 535, 38169, 8227, 772, 28782, 4563, 1490, 779, 43497, 730, 532, 23961, 963, 779, 2341, 46129, 708, 6468, 535, 5046, 535, 4536, 7174, 17261, 730, 8227, 772, 8365, 988, 43497, 730, 532, 536, 46129, 909, 532, 522, 527, 2083, 7937, 521, 1295, 5896, 1833, 708, 4972, 535, 1056, 535, 3654, 7174, 17261, 730, 43497, 730, 532, 23961, 17828, 8744, 4174, 35036, 536, 32215, 4984, 708, 4972, 535, 6135, 535, 1374, 7174, 17261, 730, 43497, 730, 532, 23961, 17828, 23848, 536, 1927, 2685, 963, 2920, 6586, 730, 525, 523, 530, 796, 5266, 875, 730, 2198, 775, 7410, 708, 4972, 535, 6326, 535, 1698, 7174, 17261, 730, 11023, 17652, 8365, 988, 43497, 730, 532, 8323, 708, 4972, 535, 3654, 535, 1664, 7174, 17261, 730, 8411, 886, 826, 526, 2280, 33480, 968, 1267, 868, 2290, 536, 878, 4318, 1047, 730, 2019, 1750, 730, 1826, 775, 708, 4972, 535, 3654, 535, 2999, 7174, 17261, 730, 1641, 13053, 19461, 1189, 49826, 536, 1641, 13053, 18483, 730, 528, 810, 730, 529, 840, 59188, 708, 4972, 535, 6468, 535, 1056, 7174, 17261, 730, 43497, 730, 532, 8323, 7925, 2835, 730, 531, 523, 526, 835, 843, 35436, 730, 533, 523, 525, 835, 1015, 4972, 535, 4905, 535, 1056, 7174, 17261, 730, 63688, 730, 4787, 525, 19331, 1706, 10012, 843, 3321, 3600, 38689, 4032, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 525, 12079, 4972, 535, 1374, 535, 1056, 7174, 17261, 730, 63688, 730, 3818, 531, 19331, 1706, 10012, 843, 3321, 3600, 38689, 4032, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 528, 12079, 4972, 535, 1386, 535, 1056, 7174, 17261, 730, 63688, 730, 4420, 526, 19331, 1706, 10012, 843, 3321, 3600, 38689, 4032, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 531, 12079, 4972, 535, 1386, 535, 3337, 923, 32535, 730, 49883, 837, 526, 1295, 49321, 884, 61026, 730, 32928, 2476, 536, 61026, 3567, 811, 16687, 43497, 708, 4972, 535, 1133, 535, 1056, 7174, 17261, 730, 63688, 730, 6583, 527, 19331, 1706, 10012, 843, 3321, 3600, 38689, 4032, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 525, 12079, 4972, 535, 2197, 535, 1056, 7174, 17261, 730, 63688, 730, 9887, 533, 19331, 1706, 10012, 843, 3321, 3600, 38689, 4032, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 528, 12079, 4972, 535, 2197, 535, 4905, 7174, 17261, 730, 24163, 57506, 1892, 4174, 730, 1155, 514, 7614, 708, 4972, 535, 1826, 535, 1056, 7174, 17261, 730, 63688, 730, 10556, 526, 19331, 1706, 10012, 843, 3321, 3600, 38689, 4032, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 531, 12079, 4972, 535, 2142, 535, 1056, 7174, 17261, 730, 63688, 730, 10520, 527, 19331, 1706, 10012, 843, 3321, 3600, 38689, 4032, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 525, 12079, 4972, 535, 2142, 535, 1664, 7174, 17261, 730, 43497, 730, 532, 8323, 7925, 2835, 730, 530, 523, 534, 835, 843, 35436, 730, 533, 523, 525, 835, 4032, 23961, 6586, 730, 526, 523, 2346, 796, 5266, 708, 4972, 535, 1664, 535, 1056, 7174, 17261, 730, 63688, 730, 10426, 530, 19331, 1706, 10012, 843, 3321, 3600, 38689, 4032, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 528, 12079, 4972, 535, 2625, 535, 1056, 7174, 17261, 730, 8501, 6634, 968, 1267, 868, 2290, 535, 43497, 730, 532, 23961, 17828, 811, 874, 20425, 2030, 963, 730, 1155, 535, 1056, 536, 8227, 772, 42806, 811, 874, 5689, 1337, 8233, 20842, 708, 4972, 535, 2625, 535, 3654, 7174, 17261, 730, 8411, 886, 826, 526, 6840, 12465, 536, 1295, 5592, 16753, 592, 18, 28519, 779, 8411, 886, 6840, 12465, 963, 779, 1764, 803, 779, 3446, 540, 19, 16, 4831, 535, 1295, 521, 779, 8294, 2177, 1014, 3965, 20, 19, 16, 3347, 535, 18672, 521, 779, 8294, 10995, 20, 21 ], "markers": [ 1566, 1578 ], "logits_raw": [ 0.6751614809036255, 1.5158710479736328 ], "probs_raw": [ 0.3013853430747986, 0.6986145973205566 ], "probs": [ 0.6235232949256897, 0.3764767050743103 ], "argmax_index": 0, "argmax": "true", "top2_margin": 0.2470465898513794, "near_tie": false, "gold": "true", "correct": true, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 2, "B": 3 } }, { "qid": "stoppage", "type": "score", "K": 6, "calibrate": true, "temperature_key": "score:6-10", "T": 1.0, "prefix": 0, "n_ids": 1636, "positions": 1636, "bucket": 2048, "ids": [ 1, 17, 15991, 23707, 3613, 4425, 4857, 561, 33371, 12292, 38620, 1141, 61026, 17824, 730, 527, 521, 5008, 1985, 2647, 8411, 886, 826, 526, 521, 22694, 3446, 843, 1513, 31225, 1015, 27464, 535, 2960, 843, 6468, 535, 1056, 522, 1536, 535, 1056, 518, 767, 52574, 884, 15774, 535, 1267, 868, 2290, 767, 12003, 884, 1868, 535, 4582, 1055, 509, 6468, 535, 1056, 535, 1056, 7174, 17261, 730, 63688, 730, 10426, 527, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 526, 12079, 6468, 535, 6135, 535, 1056, 7174, 17261, 730, 63688, 730, 11233, 533, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 528, 12079, 6468, 535, 6468, 535, 1056, 7174, 17261, 730, 63688, 730, 10556, 526, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 530, 12079, 6468, 535, 4905, 535, 1056, 7174, 17261, 730, 63688, 730, 8380, 529, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 532, 12079, 6468, 535, 4905, 535, 4536, 7174, 17261, 730, 49574, 886, 889, 529, 41172, 3999, 7410, 843, 22500, 730, 525, 523, 3188, 8061, 1015, 6468, 535, 1374, 535, 1056, 7174, 17261, 730, 63688, 730, 10173, 534, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 526, 12079, 6468, 535, 1374, 535, 1417, 923, 32535, 730, 49883, 837, 527, 1295, 49321, 884, 61026, 730, 33118, 3375, 536, 61026, 3567, 811, 16687, 43497, 708, 6468, 535, 1386, 535, 1056, 7174, 17261, 730, 63688, 730, 10520, 525, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 528, 12079, 6468, 535, 1133, 535, 1056, 7174, 17261, 730, 63688, 730, 10426, 531, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 530, 12079, 6468, 535, 2197, 535, 1056, 7174, 17261, 730, 63688, 730, 8380, 527, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 532, 12079, 6468, 535, 1826, 535, 1056, 7174, 17261, 730, 63688, 730, 11233, 530, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 526, 12079, 6468, 535, 2142, 535, 1056, 7174, 17261, 730, 63688, 730, 10520, 532, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 528, 12079, 6468, 535, 2142, 535, 4536, 7174, 17261, 730, 49574, 886, 889, 529, 41172, 3999, 7410, 843, 22500, 730, 525, 523, 5489, 8061, 1015, 6468, 535, 1664, 535, 1056, 7174, 17261, 730, 63688, 730, 8380, 534, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 530, 12079, 6468, 535, 2625, 535, 1056, 7174, 17261, 730, 63688, 730, 10173, 528, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 532, 12079, 6468, 535, 2625, 535, 1417, 923, 32535, 730, 49883, 837, 527, 1295, 49321, 884, 61026, 730, 30664, 2108, 536, 61026, 3567, 811, 16687, 43497, 708, 6468, 535, 3024, 535, 1056, 7174, 17261, 730, 63688, 730, 10426, 533, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 526, 12079, 6468, 535, 3188, 535, 1056, 7174, 17261, 730, 63688, 730, 8380, 525, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 528, 12079, 6468, 535, 2108, 535, 1056, 7174, 17261, 730, 63688, 730, 11233, 529, 19331, 1706, 10012, 536, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 528, 12079, 6468, 535, 4536, 535, 1374, 923, 32535, 730, 43497, 730, 532, 8323, 7925, 2835, 730, 533, 523, 534, 835, 843, 35436, 730, 533, 523, 525, 835, 518, 875, 730, 530, 775, 708, 6468, 535, 4536, 535, 5046, 923, 32535, 730, 43497, 730, 532, 8323, 7925, 2835, 730, 534, 523, 529, 835, 843, 35436, 730, 533, 523, 525, 835, 518, 875, 730, 1374, 775, 708, 6468, 535, 4344, 535, 2108, 923, 32535, 730, 43497, 730, 532, 23961, 6586, 730, 526, 523, 5370, 796, 5266, 843, 2390, 5723, 730, 526, 523, 2346, 796, 5266, 1015, 6468, 535, 4613, 535, 3654, 59687, 2850, 43497, 730, 532, 16017, 12365, 535, 11697, 38993, 578, 27493, 532, 522, 528, 26781, 875, 1209, 1399, 730, 528, 775, 708, 6468, 535, 4613, 535, 6468, 59687, 2850, 8411, 886, 826, 526, 13525, 968, 6155, 22694, 843, 7188, 535, 16017, 43497, 730, 532, 4032, 1113, 50027, 23993, 791, 708, 6468, 535, 4613, 535, 6468, 7174, 17261, 730, 1641, 13053, 19461, 1189, 60433, 536, 19331, 1706, 4908, 884, 24163, 57506, 708, 6468, 535, 4613, 535, 2108, 7174, 17261, 730, 16753, 28984, 968, 1267, 868, 2290, 963, 779, 2669, 26862, 708, 6468, 535, 3337, 535, 1386, 7174, 17261, 730, 12003, 1868, 9388, 811, 4582, 1055, 843, 589, 60088, 730, 526, 1015, 6468, 535, 4195, 535, 5843, 7174, 17261, 730, 49883, 837, 526, 810, 837, 527, 54126, 843, 2744, 19331, 1706, 4032, 2515, 29067, 7410, 708, 6468, 535, 3745, 535, 1664, 7174, 17261, 730, 24163, 57506, 963, 730, 2752, 514, 7614, 708, 6468, 535, 1838, 535, 1374, 7174, 17261, 730, 4582, 1055, 884, 3475, 963, 43497, 730, 532, 536, 11023, 17652, 6944, 811, 43497, 730, 532, 8323, 708, 6468, 535, 7058, 535, 3167, 7174, 17261, 730, 24163, 57506, 963, 730, 9731, 514, 7614, 536, 1641, 13053, 18483, 730, 528, 810, 730, 529, 8557, 811, 1361, 19082, 871, 708, 6468, 535, 3878, 535, 898, 7174, 17261, 730, 16017, 3360, 4468, 968, 4582, 1055, 535, 38169, 8227, 772, 28782, 4563, 1490, 779, 43497, 730, 532, 23961, 963, 779, 2341, 46129, 708, 6468, 535, 5046, 535, 4536, 7174, 17261, 730, 8227, 772, 8365, 988, 43497, 730, 532, 536, 46129, 909, 532, 522, 527, 2083, 7937, 521, 1295, 5896, 1833, 708, 4972, 535, 1056, 535, 3654, 7174, 17261, 730, 43497, 730, 532, 23961, 17828, 8744, 4174, 35036, 536, 32215, 4984, 708, 4972, 535, 6135, 535, 1374, 7174, 17261, 730, 43497, 730, 532, 23961, 17828, 23848, 536, 1927, 2685, 963, 2920, 6586, 730, 525, 523, 530, 796, 5266, 875, 730, 2198, 775, 7410, 708, 4972, 535, 6326, 535, 1698, 7174, 17261, 730, 11023, 17652, 8365, 988, 43497, 730, 532, 8323, 708, 4972, 535, 3654, 535, 1664, 7174, 17261, 730, 8411, 886, 826, 526, 2280, 33480, 968, 1267, 868, 2290, 536, 878, 4318, 1047, 730, 2019, 1750, 730, 1826, 775, 708, 4972, 535, 3654, 535, 2999, 7174, 17261, 730, 1641, 13053, 19461, 1189, 49826, 536, 1641, 13053, 18483, 730, 528, 810, 730, 529, 840, 59188, 708, 4972, 535, 6468, 535, 1056, 7174, 17261, 730, 43497, 730, 532, 8323, 7925, 2835, 730, 531, 523, 526, 835, 843, 35436, 730, 533, 523, 525, 835, 1015, 4972, 535, 4905, 535, 1056, 7174, 17261, 730, 63688, 730, 4787, 525, 19331, 1706, 10012, 843, 3321, 3600, 38689, 4032, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 525, 12079, 4972, 535, 1374, 535, 1056, 7174, 17261, 730, 63688, 730, 3818, 531, 19331, 1706, 10012, 843, 3321, 3600, 38689, 4032, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 528, 12079, 4972, 535, 1386, 535, 1056, 7174, 17261, 730, 63688, 730, 4420, 526, 19331, 1706, 10012, 843, 3321, 3600, 38689, 4032, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 531, 12079, 4972, 535, 1386, 535, 3337, 923, 32535, 730, 49883, 837, 526, 1295, 49321, 884, 61026, 730, 32928, 2476, 536, 61026, 3567, 811, 16687, 43497, 708, 4972, 535, 1133, 535, 1056, 7174, 17261, 730, 63688, 730, 6583, 527, 19331, 1706, 10012, 843, 3321, 3600, 38689, 4032, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 525, 12079, 4972, 535, 2197, 535, 1056, 7174, 17261, 730, 63688, 730, 9887, 533, 19331, 1706, 10012, 843, 3321, 3600, 38689, 4032, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 528, 12079, 4972, 535, 2197, 535, 4905, 7174, 17261, 730, 24163, 57506, 1892, 4174, 730, 1155, 514, 7614, 708, 4972, 535, 1826, 535, 1056, 7174, 17261, 730, 63688, 730, 10556, 526, 19331, 1706, 10012, 843, 3321, 3600, 38689, 4032, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 531, 12079, 4972, 535, 2142, 535, 1056, 7174, 17261, 730, 63688, 730, 10520, 527, 19331, 1706, 10012, 843, 3321, 3600, 38689, 4032, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 525, 12079, 4972, 535, 2142, 535, 1664, 7174, 17261, 730, 43497, 730, 532, 8323, 7925, 2835, 730, 530, 523, 534, 835, 843, 35436, 730, 533, 523, 525, 835, 4032, 23961, 6586, 730, 526, 523, 2346, 796, 5266, 708, 4972, 535, 1664, 535, 1056, 7174, 17261, 730, 63688, 730, 10426, 530, 19331, 1706, 10012, 843, 3321, 3600, 38689, 4032, 50027, 730, 526, 522, 1374, 5592, 536, 49883, 2062, 4002, 730, 2962, 523, 528, 12079, 4972, 535, 2625, 535, 1056, 7174, 17261, 730, 8501, 6634, 968, 1267, 868, 2290, 535, 43497, 730, 532, 23961, 17828, 811, 874, 20425, 2030, 963, 730, 1155, 535, 1056, 536, 8227, 772, 42806, 811, 874, 5689, 1337, 8233, 20842, 708, 4972, 535, 2625, 535, 3654, 7174, 17261, 730, 8411, 886, 826, 526, 6840, 12465, 536, 1295, 5592, 16753, 592, 18, 4023, 1787, 953, 779, 8411, 886, 7408, 13525, 540, 19, 16, 2119, 730, 525, 535, 1490, 730, 527, 5112, 20, 19, 16, 2119, 730, 526, 535, 730, 527, 811, 730, 530, 5112, 20, 19, 16, 2119, 730, 527, 535, 730, 530, 811, 730, 1155, 5112, 20, 19, 16, 2119, 730, 528, 535, 730, 1155, 811, 730, 898, 5112, 20, 19, 16, 2119, 730, 529, 535, 730, 898, 811, 730, 2108, 5112, 20, 19, 16, 2119, 730, 530, 535, 1391, 730, 2108, 5112, 20, 21 ], "markers": [ 1562, 1573, 1586, 1599, 1612, 1625 ], "logits_raw": [ -6.378607273101807, -3.3338048458099365, -2.862976551055908, -1.375028133392334, 0.9950432181358337, -0.7478825449943542 ], "probs_raw": [ 0.00048148856149055064, 0.010114093311131, 0.01619589701294899, 0.07171545922756195, 0.7672231793403625, 0.13426990807056427 ], "probs": [ 0.00048148856149055064, 0.010114093311131, 0.01619589701294899, 0.07171545922756195, 0.7672231793403625, 0.13426990807056427 ], "argmax_index": 4, "argmax": "4", "top2_margin": 0.6329532712697983, "near_tie": false, "gold": "4", "correct": true, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 0, "B": 3 } } ], "response": { "answers": { "component": { "type": "choice", "choice": "lane_7", "confidence": 0.6164789795875549, "probabilities": { "scanner": 0.10629419982433319, "lane_7": 0.6164789795875549, "diverter": 0.23265084624290466, "power": 0.032522596418857574, "software": 0.012053357437252998 } }, "normal_at_end": { "type": "noul", "noul": 0.6235232949256897 }, "stoppage": { "type": "score", "score": 3.9978945227339864, "confidence": 0.7672231793403625, "probabilities": { "0": 0.00048148856149055064, "1": 0.010114093311131, "2": 0.01619589701294899, "3": 0.07171545922756195, "4": 0.7672231793403625, "5": 0.13426990807056427 }, "legend": { "0": "under 2 minutes", "1": "2 to 5 minutes", "2": "5 to 10 minutes", "3": "10 to 20 minutes", "4": "20 to 40 minutes", "5": "over 40 minutes" } } }, "usage": { "input_tokens": 4830, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "own_L03", "mode": "text", "source": "own_long", "native": true, "prefix": 0, "usage_input_tokens": 5815, "questions": [ { "qid": "fee_motion", "type": "choice", "K": 4, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 1463, "positions": 1463, "bucket": 2048, "ids": [ 1, 17, 564, 17769, 3613, 4425, 12411, 34632, 2730, 41891, 516, 31289, 7272, 550, 10877, 708, 13635, 3962, 803, 779, 27845, 5525, 35590, 4908, 884, 22045, 730, 1374, 4690, 730, 1718, 531, 963, 730, 913, 535, 1664, 797, 33864, 54898, 2176, 24579, 7329, 509, 41678, 535, 730, 3745, 3967, 843, 1615, 959, 1143, 13975, 12930, 518, 810, 730, 528, 24753, 521, 3136, 1377, 779, 29780, 12548, 50804, 777, 521, 2100, 3852, 1543, 523, 1030, 779, 12377, 535, 918, 54069, 523, 3617, 3962, 535, 9413, 598, 19718, 843, 52709, 1126, 1635, 10412, 6650, 535, 730, 1632, 3967, 3567, 1149, 1514, 6650, 523, 509, 526, 523, 56975, 810, 967, 24339, 708, 1098, 12377, 41975, 3967, 810, 24753, 810, 14954, 791, 779, 11917, 15671, 875, 779, 1517, 803, 779, 5362, 523, 941, 12778, 6835, 730, 1664, 3967, 811, 874, 2251, 875, 768, 3177, 8324, 811, 874, 967, 777, 931, 523, 3877, 730, 3745, 3967, 2251, 521, 779, 8324, 953, 13096, 967, 777, 931, 523, 509, 527, 523, 3617, 3962, 803, 779, 2659, 8900, 3177, 8324, 708, 1098, 5112, 803, 779, 8324, 4908, 884, 730, 1532, 4690, 730, 1718, 530, 1166, 12043, 906, 768, 3347, 3084, 521, 7885, 968, 9213, 833, 810, 2341, 791, 968, 28682, 523, 2479, 1166, 1295, 14332, 35517, 896, 1166, 1014, 4001, 884, 779, 26168, 523, 509, 528, 523, 21640, 1090, 2312, 708, 548, 54069, 17422, 2851, 779, 1423, 523, 941, 10039, 9657, 779, 6844, 7405, 521, 1144, 9388, 730, 527, 521, 9300, 875, 779, 11917, 12562, 3549, 521, 810, 35130, 1547, 27297, 45546, 3256, 779, 8008, 3509, 521, 1895, 916, 1209, 1399, 11583, 20470, 523, 941, 7982, 875, 768, 55097, 22212, 884, 8995, 27784, 6806, 521, 810, 779, 17677, 1178, 2166, 11860, 811, 6715, 1230, 768, 11013, 8153, 797, 3354, 523, 49427, 5313, 14118, 988, 730, 18833, 811, 730, 22989, 16671, 523, 1405, 14954, 791, 779, 15671, 810, 779, 20470, 1316, 5514, 779, 38666, 811, 1973, 3622, 2759, 3217, 768, 1423, 523, 509, 529, 523, 9312, 802, 9233, 1090, 2312, 708, 1098, 1230, 9932, 3185, 802, 9233, 521, 2345, 586, 521, 6892, 779, 12324, 875, 779, 1423, 523, 53813, 953, 730, 531, 521, 21957, 535, 22398, 23023, 730, 527, 521, 26547, 521, 779, 6844, 7405, 730, 527, 521, 9300, 810, 41634, 730, 526, 521, 30542, 523, 2315, 2124, 11761, 953, 730, 529, 521, 25803, 521, 8787, 38666, 17236, 843, 526, 521, 21405, 1109, 12235, 843, 23617, 1109, 11917, 33265, 843, 7537, 518, 810, 7494, 875, 779, 8008, 3509, 5158, 6688, 843, 526, 521, 26668, 1191, 941, 1423, 8557, 916, 768, 41938, 803, 730, 526, 521, 32647, 810, 768, 7493, 8386, 803, 730, 534, 521, 19222, 523, 941, 12324, 1320, 1381, 20425, 968, 902, 7054, 1924, 5725, 521, 1316, 1833, 1295, 4662, 523, 32826, 12043, 779, 12324, 968, 768, 1914, 803, 6108, 521, 916, 1295, 20402, 2723, 523, 509, 530, 523, 32964, 988, 779, 3822, 3892, 708, 26691, 2514, 843, 549, 37986, 3271, 779, 2514, 21440, 730, 5411, 8267, 35304, 884, 779, 1375, 977, 5329, 779, 2254, 810, 50398, 1547, 803, 779, 2759, 5854, 20842, 884, 779, 5253, 523, 1311, 1499, 1672, 811, 8153, 779, 8008, 3509, 875, 21275, 4476, 797, 779, 6844, 810, 856, 5400, 875, 1547, 1209, 20470, 819, 30193, 2514, 843, 545, 6310, 793, 3271, 29504, 779, 6844, 7405, 521, 779, 2514, 9657, 768, 30780, 4915, 797, 4843, 896, 17322, 1919, 523, 835, 3316, 4915, 797, 779, 11917, 856, 13573, 875, 4471, 1288, 768, 35563, 1011, 874, 20560, 963, 768, 16336, 7264, 819, 35798, 4977, 810, 5600, 2514, 843, 549, 831, 833, 3271, 10325, 988, 779, 22212, 7982, 521, 779, 2514, 4468, 2204, 15073, 11513, 15489, 810, 768, 26781, 15761, 884, 13878, 9690, 811, 779, 17677, 536, 1547, 803, 779, 15489, 1052, 1381, 50398, 523, 509, 531, 523, 7329, 12562, 3549, 708, 559, 4182, 4468, 875, 779, 11917, 15671, 896, 779, 12562, 12940, 1178, 1381, 29878, 963, 730, 1133, 521, 6019, 523, 3877, 779, 730, 527, 521, 9300, 988, 779, 6844, 7405, 521, 779, 3549, 2166, 12973, 963, 730, 1632, 521, 14671, 523, 941, 15671, 1178, 6944, 811, 1547, 56302, 8810, 592, 875, 779, 2280, 810, 45777, 1149, 13041, 968, 779, 1764, 803, 3354, 523, 6162, 4584, 2155, 2076, 4526, 1929, 30541, 1232, 1000, 797, 779, 12562, 3494, 2013, 779, 6844, 521, 810, 902, 8535, 8923, 1188, 3999, 779, 12562, 2076, 1523, 1336, 10309, 523, 32826, 5751, 3984, 779, 10039, 1499, 3644, 1523, 3290, 8372, 536, 28682, 1845, 896, 779, 15671, 1499, 2743, 1892, 811, 768, 3177, 8324, 2076, 54526, 1523, 803, 779, 10039, 1090, 4636, 523, 3732, 5484, 9303, 811, 49904, 768, 32673, 5043, 875, 779, 3549, 797, 4963, 521, 810, 2993, 3967, 12547, 3501, 811, 1801, 523, 509, 532, 523, 53776, 803, 12900, 708, 548, 54069, 11860, 811, 2472, 2030, 906, 12377, 810, 953, 840, 5508, 3800, 946, 920, 4346, 523, 9413, 598, 19718, 953, 840, 5508, 3800, 906, 27246, 946, 920, 4346, 523, 2345, 586, 13682, 2208, 906, 3185, 802, 9233, 1720, 4250, 1779, 523, 8726, 3967, 1166, 33256, 875, 779, 2375, 535, 10593, 3102, 810, 15642, 523, 4162, 3290, 15450, 988, 2059, 16397, 521, 768, 7155, 48506, 953, 4908, 523, 10593, 3102, 5068, 730, 2235, 20402, 810, 15642, 730, 1443, 521, 916, 730, 527, 20069, 9244, 12452, 523, 10593, 3102, 953, 13096, 11510, 906, 3185, 802, 9233, 523, 941, 8324, 14954, 791, 2345, 586, 6248, 876, 875, 1210, 1779, 803, 3714, 523, 509, 533, 523, 41453, 730, 526, 535, 39391, 1465, 14122, 884, 779, 7314, 5253, 708, 2364, 4346, 968, 1030, 774, 521, 2341, 791, 968, 15821, 793, 535, 997, 7464, 779, 10039, 47882, 1287, 811, 730, 526, 521, 1086, 988, 1352, 25511, 884, 2759, 39391, 1465, 14122, 963, 779, 9028, 803, 779, 7314, 5253, 521, 3910, 811, 17677, 15001, 2927, 1030, 774, 10201, 896, 28285, 5419, 938, 6654, 979, 3213, 811, 779, 8746, 1327, 521, 1144, 30809, 779, 5799, 810, 12785, 779, 3789, 8034, 875, 10415, 824, 5569, 523, 3732, 5484, 5751, 3984, 779, 1465, 14122, 1499, 8524, 38883, 536, 1030, 774, 25723, 896, 1186, 1499, 2472, 797, 3160, 803, 779, 15565, 1090, 11404, 523, 1019, 2176, 535, 730, 2999, 875, 521, 730, 534, 2723, 521, 730, 532, 41754, 806, 1114, 523, 941, 9557, 953, 7818, 523, 509, 534, 523, 41453, 730, 527, 535, 8900, 22398, 22819, 708, 2364, 4346, 968, 779, 15671, 535, 997, 7464, 779, 8900, 22819, 27870, 988, 730, 1374, 811, 730, 1386, 1094, 8461, 988, 730, 526, 4471, 730, 1718, 531, 2927, 941, 3185, 802, 9233, 59339, 1845, 896, 1033, 1423, 1090, 41938, 4201, 902, 3761, 27759, 875, 2166, 523, 17857, 3967, 1845, 896, 768, 3585, 22819, 1499, 3151, 1612, 1361, 16671, 521, 4155, 797, 779, 1361, 1434, 1872, 884, 13878, 9690, 523, 1019, 2176, 535, 730, 1133, 875, 521, 730, 2117, 2723, 521, 730, 528, 41754, 806, 1114, 523, 941, 9557, 953, 17602, 521, 810, 779, 22819, 32660, 963, 730, 1374, 523, 509, 1155, 523, 41453, 730, 528, 535, 3681, 9041, 708, 2364, 4346, 968, 38992, 1068, 521, 2341, 791, 968, 43842, 535, 997, 7464, 779, 10039, 12881, 768, 3681, 9041, 884, 779, 52352, 13646, 5598, 779, 11917, 2927, 32826, 7951, 779, 4588, 797, 11618, 1203, 5751, 1316, 1499, 2144, 1720, 779, 9041, 810, 3984, 779, 11917, 15671, 40917, 779, 13646, 523, 1311, 953, 11860, 811, 2021, 1287, 768, 3822, 2514, 803, 38992, 1068, 521, 43842, 810, 15642, 811, 2312, 811, 779, 3280, 3177, 8324, 523, 3253, 11386, 953, 4281, 523, 509, 1632, 523, 10972, 1312, 3436, 708, 13792, 518, 7223, 4729, 884, 8995, 27784, 535, 2993, 3967, 4468, 11242, 10978, 1714, 3537, 779, 5630, 963, 2254, 3217, 523, 941, 27246, 1188, 5773, 811, 779, 17677, 14065, 875, 779, 3354, 8153, 811, 2957, 6586, 1818, 819, 10388, 518, 54894, 535, 20470, 938, 4602, 811, 5514, 884, 779, 9168, 3392, 803, 779, 7314, 819, 10120, 518, 941, 29780, 12548, 50804, 777, 1845, 896, 779, 17677, 1090, 8825, 14018, 875, 3681, 7092, 21981, 797, 4963, 810, 896, 779, 9041, 2724, 33460, 819, 9285, 518, 1416, 48053, 535, 779, 17677, 1188, 5840, 7980, 23878, 988, 5105, 536, 779, 6111, 3378, 1188, 874, 16022, 797, 779, 3280, 38666, 819, 11634, 518, 8528, 3711, 937, 1888, 535, 9213, 833, 4468, 896, 779, 1629, 3711, 937, 1888, 5329, 779, 15565, 953, 20425, 797, 6060, 810, 896, 1352, 46967, 2716, 1046, 797, 5277, 536, 779, 15671, 1188, 2565, 8297, 1667, 523, 509, 1374, 523, 13887, 803, 779, 3280, 8324, 708, 1098, 3280, 3177, 8324, 1188, 874, 4908, 797, 4526, 536, 779, 5562, 1188, 874, 9139, 797, 779, 6844, 38666, 523, 941, 8324, 8557, 963, 730, 2197, 535, 1155, 523, 18, 3493, 9862, 811, 779, 9557, 884, 779, 8900, 22819, 540, 19, 16, 12043, 20, 19, 16, 20691, 20, 19, 16, 825, 25745, 535, 6953, 2939, 791, 811, 768, 2919, 8324, 933, 768, 3822, 2514, 20, 19, 16, 47927, 535, 3877, 577, 29058, 2076, 779, 11386, 20, 21 ], "markers": [ 1427, 1431, 1435, 1452 ], "logits_raw": [ -1.615821123123169, -2.7980968952178955, 0.5164243578910828, -2.294139862060547 ], "probs_raw": [ 0.09758160263299942, 0.029916591942310333, 0.8229821920394897, 0.04951968789100647 ], "probs": [ 0.15077616274356842, 0.0647965595126152, 0.6915529370307922, 0.09287437796592712 ], "argmax_index": 2, "argmax": "deferred", "top2_margin": 0.5407767742872238, "near_tie": false, "gold": "rejected", "correct": false, "host_softmax_max_abs_dp": 7.450580596923828e-09, "batch": { "call": 0, "index": 1, "B": 4 } }, { "qid": "quorate", "type": "noul", "K": 2, "calibrate": true, "temperature_key": "noul:2", "T": 1.6663223505020142, "prefix": 0, "n_ids": 1446, "positions": 1446, "bucket": 2048, "ids": [ 1, 17, 564, 17769, 3613, 4425, 12411, 34632, 2730, 41891, 516, 31289, 7272, 550, 10877, 708, 13635, 3962, 803, 779, 27845, 5525, 35590, 4908, 884, 22045, 730, 1374, 4690, 730, 1718, 531, 963, 730, 913, 535, 1664, 797, 33864, 54898, 2176, 24579, 7329, 509, 41678, 535, 730, 3745, 3967, 843, 1615, 959, 1143, 13975, 12930, 518, 810, 730, 528, 24753, 521, 3136, 1377, 779, 29780, 12548, 50804, 777, 521, 2100, 3852, 1543, 523, 1030, 779, 12377, 535, 918, 54069, 523, 3617, 3962, 535, 9413, 598, 19718, 843, 52709, 1126, 1635, 10412, 6650, 535, 730, 1632, 3967, 3567, 1149, 1514, 6650, 523, 509, 526, 523, 56975, 810, 967, 24339, 708, 1098, 12377, 41975, 3967, 810, 24753, 810, 14954, 791, 779, 11917, 15671, 875, 779, 1517, 803, 779, 5362, 523, 941, 12778, 6835, 730, 1664, 3967, 811, 874, 2251, 875, 768, 3177, 8324, 811, 874, 967, 777, 931, 523, 3877, 730, 3745, 3967, 2251, 521, 779, 8324, 953, 13096, 967, 777, 931, 523, 509, 527, 523, 3617, 3962, 803, 779, 2659, 8900, 3177, 8324, 708, 1098, 5112, 803, 779, 8324, 4908, 884, 730, 1532, 4690, 730, 1718, 530, 1166, 12043, 906, 768, 3347, 3084, 521, 7885, 968, 9213, 833, 810, 2341, 791, 968, 28682, 523, 2479, 1166, 1295, 14332, 35517, 896, 1166, 1014, 4001, 884, 779, 26168, 523, 509, 528, 523, 21640, 1090, 2312, 708, 548, 54069, 17422, 2851, 779, 1423, 523, 941, 10039, 9657, 779, 6844, 7405, 521, 1144, 9388, 730, 527, 521, 9300, 875, 779, 11917, 12562, 3549, 521, 810, 35130, 1547, 27297, 45546, 3256, 779, 8008, 3509, 521, 1895, 916, 1209, 1399, 11583, 20470, 523, 941, 7982, 875, 768, 55097, 22212, 884, 8995, 27784, 6806, 521, 810, 779, 17677, 1178, 2166, 11860, 811, 6715, 1230, 768, 11013, 8153, 797, 3354, 523, 49427, 5313, 14118, 988, 730, 18833, 811, 730, 22989, 16671, 523, 1405, 14954, 791, 779, 15671, 810, 779, 20470, 1316, 5514, 779, 38666, 811, 1973, 3622, 2759, 3217, 768, 1423, 523, 509, 529, 523, 9312, 802, 9233, 1090, 2312, 708, 1098, 1230, 9932, 3185, 802, 9233, 521, 2345, 586, 521, 6892, 779, 12324, 875, 779, 1423, 523, 53813, 953, 730, 531, 521, 21957, 535, 22398, 23023, 730, 527, 521, 26547, 521, 779, 6844, 7405, 730, 527, 521, 9300, 810, 41634, 730, 526, 521, 30542, 523, 2315, 2124, 11761, 953, 730, 529, 521, 25803, 521, 8787, 38666, 17236, 843, 526, 521, 21405, 1109, 12235, 843, 23617, 1109, 11917, 33265, 843, 7537, 518, 810, 7494, 875, 779, 8008, 3509, 5158, 6688, 843, 526, 521, 26668, 1191, 941, 1423, 8557, 916, 768, 41938, 803, 730, 526, 521, 32647, 810, 768, 7493, 8386, 803, 730, 534, 521, 19222, 523, 941, 12324, 1320, 1381, 20425, 968, 902, 7054, 1924, 5725, 521, 1316, 1833, 1295, 4662, 523, 32826, 12043, 779, 12324, 968, 768, 1914, 803, 6108, 521, 916, 1295, 20402, 2723, 523, 509, 530, 523, 32964, 988, 779, 3822, 3892, 708, 26691, 2514, 843, 549, 37986, 3271, 779, 2514, 21440, 730, 5411, 8267, 35304, 884, 779, 1375, 977, 5329, 779, 2254, 810, 50398, 1547, 803, 779, 2759, 5854, 20842, 884, 779, 5253, 523, 1311, 1499, 1672, 811, 8153, 779, 8008, 3509, 875, 21275, 4476, 797, 779, 6844, 810, 856, 5400, 875, 1547, 1209, 20470, 819, 30193, 2514, 843, 545, 6310, 793, 3271, 29504, 779, 6844, 7405, 521, 779, 2514, 9657, 768, 30780, 4915, 797, 4843, 896, 17322, 1919, 523, 835, 3316, 4915, 797, 779, 11917, 856, 13573, 875, 4471, 1288, 768, 35563, 1011, 874, 20560, 963, 768, 16336, 7264, 819, 35798, 4977, 810, 5600, 2514, 843, 549, 831, 833, 3271, 10325, 988, 779, 22212, 7982, 521, 779, 2514, 4468, 2204, 15073, 11513, 15489, 810, 768, 26781, 15761, 884, 13878, 9690, 811, 779, 17677, 536, 1547, 803, 779, 15489, 1052, 1381, 50398, 523, 509, 531, 523, 7329, 12562, 3549, 708, 559, 4182, 4468, 875, 779, 11917, 15671, 896, 779, 12562, 12940, 1178, 1381, 29878, 963, 730, 1133, 521, 6019, 523, 3877, 779, 730, 527, 521, 9300, 988, 779, 6844, 7405, 521, 779, 3549, 2166, 12973, 963, 730, 1632, 521, 14671, 523, 941, 15671, 1178, 6944, 811, 1547, 56302, 8810, 592, 875, 779, 2280, 810, 45777, 1149, 13041, 968, 779, 1764, 803, 3354, 523, 6162, 4584, 2155, 2076, 4526, 1929, 30541, 1232, 1000, 797, 779, 12562, 3494, 2013, 779, 6844, 521, 810, 902, 8535, 8923, 1188, 3999, 779, 12562, 2076, 1523, 1336, 10309, 523, 32826, 5751, 3984, 779, 10039, 1499, 3644, 1523, 3290, 8372, 536, 28682, 1845, 896, 779, 15671, 1499, 2743, 1892, 811, 768, 3177, 8324, 2076, 54526, 1523, 803, 779, 10039, 1090, 4636, 523, 3732, 5484, 9303, 811, 49904, 768, 32673, 5043, 875, 779, 3549, 797, 4963, 521, 810, 2993, 3967, 12547, 3501, 811, 1801, 523, 509, 532, 523, 53776, 803, 12900, 708, 548, 54069, 11860, 811, 2472, 2030, 906, 12377, 810, 953, 840, 5508, 3800, 946, 920, 4346, 523, 9413, 598, 19718, 953, 840, 5508, 3800, 906, 27246, 946, 920, 4346, 523, 2345, 586, 13682, 2208, 906, 3185, 802, 9233, 1720, 4250, 1779, 523, 8726, 3967, 1166, 33256, 875, 779, 2375, 535, 10593, 3102, 810, 15642, 523, 4162, 3290, 15450, 988, 2059, 16397, 521, 768, 7155, 48506, 953, 4908, 523, 10593, 3102, 5068, 730, 2235, 20402, 810, 15642, 730, 1443, 521, 916, 730, 527, 20069, 9244, 12452, 523, 10593, 3102, 953, 13096, 11510, 906, 3185, 802, 9233, 523, 941, 8324, 14954, 791, 2345, 586, 6248, 876, 875, 1210, 1779, 803, 3714, 523, 509, 533, 523, 41453, 730, 526, 535, 39391, 1465, 14122, 884, 779, 7314, 5253, 708, 2364, 4346, 968, 1030, 774, 521, 2341, 791, 968, 15821, 793, 535, 997, 7464, 779, 10039, 47882, 1287, 811, 730, 526, 521, 1086, 988, 1352, 25511, 884, 2759, 39391, 1465, 14122, 963, 779, 9028, 803, 779, 7314, 5253, 521, 3910, 811, 17677, 15001, 2927, 1030, 774, 10201, 896, 28285, 5419, 938, 6654, 979, 3213, 811, 779, 8746, 1327, 521, 1144, 30809, 779, 5799, 810, 12785, 779, 3789, 8034, 875, 10415, 824, 5569, 523, 3732, 5484, 5751, 3984, 779, 1465, 14122, 1499, 8524, 38883, 536, 1030, 774, 25723, 896, 1186, 1499, 2472, 797, 3160, 803, 779, 15565, 1090, 11404, 523, 1019, 2176, 535, 730, 2999, 875, 521, 730, 534, 2723, 521, 730, 532, 41754, 806, 1114, 523, 941, 9557, 953, 7818, 523, 509, 534, 523, 41453, 730, 527, 535, 8900, 22398, 22819, 708, 2364, 4346, 968, 779, 15671, 535, 997, 7464, 779, 8900, 22819, 27870, 988, 730, 1374, 811, 730, 1386, 1094, 8461, 988, 730, 526, 4471, 730, 1718, 531, 2927, 941, 3185, 802, 9233, 59339, 1845, 896, 1033, 1423, 1090, 41938, 4201, 902, 3761, 27759, 875, 2166, 523, 17857, 3967, 1845, 896, 768, 3585, 22819, 1499, 3151, 1612, 1361, 16671, 521, 4155, 797, 779, 1361, 1434, 1872, 884, 13878, 9690, 523, 1019, 2176, 535, 730, 1133, 875, 521, 730, 2117, 2723, 521, 730, 528, 41754, 806, 1114, 523, 941, 9557, 953, 17602, 521, 810, 779, 22819, 32660, 963, 730, 1374, 523, 509, 1155, 523, 41453, 730, 528, 535, 3681, 9041, 708, 2364, 4346, 968, 38992, 1068, 521, 2341, 791, 968, 43842, 535, 997, 7464, 779, 10039, 12881, 768, 3681, 9041, 884, 779, 52352, 13646, 5598, 779, 11917, 2927, 32826, 7951, 779, 4588, 797, 11618, 1203, 5751, 1316, 1499, 2144, 1720, 779, 9041, 810, 3984, 779, 11917, 15671, 40917, 779, 13646, 523, 1311, 953, 11860, 811, 2021, 1287, 768, 3822, 2514, 803, 38992, 1068, 521, 43842, 810, 15642, 811, 2312, 811, 779, 3280, 3177, 8324, 523, 3253, 11386, 953, 4281, 523, 509, 1632, 523, 10972, 1312, 3436, 708, 13792, 518, 7223, 4729, 884, 8995, 27784, 535, 2993, 3967, 4468, 11242, 10978, 1714, 3537, 779, 5630, 963, 2254, 3217, 523, 941, 27246, 1188, 5773, 811, 779, 17677, 14065, 875, 779, 3354, 8153, 811, 2957, 6586, 1818, 819, 10388, 518, 54894, 535, 20470, 938, 4602, 811, 5514, 884, 779, 9168, 3392, 803, 779, 7314, 819, 10120, 518, 941, 29780, 12548, 50804, 777, 1845, 896, 779, 17677, 1090, 8825, 14018, 875, 3681, 7092, 21981, 797, 4963, 810, 896, 779, 9041, 2724, 33460, 819, 9285, 518, 1416, 48053, 535, 779, 17677, 1188, 5840, 7980, 23878, 988, 5105, 536, 779, 6111, 3378, 1188, 874, 16022, 797, 779, 3280, 38666, 819, 11634, 518, 8528, 3711, 937, 1888, 535, 9213, 833, 4468, 896, 779, 1629, 3711, 937, 1888, 5329, 779, 15565, 953, 20425, 797, 6060, 810, 896, 1352, 46967, 2716, 1046, 797, 5277, 536, 779, 15671, 1188, 2565, 8297, 1667, 523, 509, 1374, 523, 13887, 803, 779, 3280, 8324, 708, 1098, 3280, 3177, 8324, 1188, 874, 4908, 797, 4526, 536, 779, 5562, 1188, 874, 9139, 797, 779, 6844, 38666, 523, 941, 8324, 8557, 963, 730, 2197, 535, 1155, 523, 18, 28519, 779, 8324, 967, 777, 931, 540, 19, 16, 4831, 535, 1295, 521, 779, 8294, 2177, 1014, 3965, 20, 19, 16, 3347, 535, 18672, 521, 779, 8294, 10995, 20, 21 ], "markers": [ 1424, 1436 ], "logits_raw": [ 0.2696681618690491, 5.161313533782959 ], "probs_raw": [ 0.007453090511262417, 0.9925468564033508 ], "probs": [ 0.9495782256126404, 0.0504218228161335 ], "argmax_index": 0, "argmax": "true", "top2_margin": 0.8991564027965069, "near_tie": false, "gold": "true", "correct": true, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 2, "B": 4 } }, { "qid": "treasurer", "type": "choice", "K": 2, "calibrate": true, "temperature_key": "choice:2", "T": 1.7465145587921143, "prefix": 0, "n_ids": 1440, "positions": 1440, "bucket": 2048, "ids": [ 1, 17, 564, 17769, 3613, 4425, 12411, 34632, 2730, 41891, 516, 31289, 7272, 550, 10877, 708, 13635, 3962, 803, 779, 27845, 5525, 35590, 4908, 884, 22045, 730, 1374, 4690, 730, 1718, 531, 963, 730, 913, 535, 1664, 797, 33864, 54898, 2176, 24579, 7329, 509, 41678, 535, 730, 3745, 3967, 843, 1615, 959, 1143, 13975, 12930, 518, 810, 730, 528, 24753, 521, 3136, 1377, 779, 29780, 12548, 50804, 777, 521, 2100, 3852, 1543, 523, 1030, 779, 12377, 535, 918, 54069, 523, 3617, 3962, 535, 9413, 598, 19718, 843, 52709, 1126, 1635, 10412, 6650, 535, 730, 1632, 3967, 3567, 1149, 1514, 6650, 523, 509, 526, 523, 56975, 810, 967, 24339, 708, 1098, 12377, 41975, 3967, 810, 24753, 810, 14954, 791, 779, 11917, 15671, 875, 779, 1517, 803, 779, 5362, 523, 941, 12778, 6835, 730, 1664, 3967, 811, 874, 2251, 875, 768, 3177, 8324, 811, 874, 967, 777, 931, 523, 3877, 730, 3745, 3967, 2251, 521, 779, 8324, 953, 13096, 967, 777, 931, 523, 509, 527, 523, 3617, 3962, 803, 779, 2659, 8900, 3177, 8324, 708, 1098, 5112, 803, 779, 8324, 4908, 884, 730, 1532, 4690, 730, 1718, 530, 1166, 12043, 906, 768, 3347, 3084, 521, 7885, 968, 9213, 833, 810, 2341, 791, 968, 28682, 523, 2479, 1166, 1295, 14332, 35517, 896, 1166, 1014, 4001, 884, 779, 26168, 523, 509, 528, 523, 21640, 1090, 2312, 708, 548, 54069, 17422, 2851, 779, 1423, 523, 941, 10039, 9657, 779, 6844, 7405, 521, 1144, 9388, 730, 527, 521, 9300, 875, 779, 11917, 12562, 3549, 521, 810, 35130, 1547, 27297, 45546, 3256, 779, 8008, 3509, 521, 1895, 916, 1209, 1399, 11583, 20470, 523, 941, 7982, 875, 768, 55097, 22212, 884, 8995, 27784, 6806, 521, 810, 779, 17677, 1178, 2166, 11860, 811, 6715, 1230, 768, 11013, 8153, 797, 3354, 523, 49427, 5313, 14118, 988, 730, 18833, 811, 730, 22989, 16671, 523, 1405, 14954, 791, 779, 15671, 810, 779, 20470, 1316, 5514, 779, 38666, 811, 1973, 3622, 2759, 3217, 768, 1423, 523, 509, 529, 523, 9312, 802, 9233, 1090, 2312, 708, 1098, 1230, 9932, 3185, 802, 9233, 521, 2345, 586, 521, 6892, 779, 12324, 875, 779, 1423, 523, 53813, 953, 730, 531, 521, 21957, 535, 22398, 23023, 730, 527, 521, 26547, 521, 779, 6844, 7405, 730, 527, 521, 9300, 810, 41634, 730, 526, 521, 30542, 523, 2315, 2124, 11761, 953, 730, 529, 521, 25803, 521, 8787, 38666, 17236, 843, 526, 521, 21405, 1109, 12235, 843, 23617, 1109, 11917, 33265, 843, 7537, 518, 810, 7494, 875, 779, 8008, 3509, 5158, 6688, 843, 526, 521, 26668, 1191, 941, 1423, 8557, 916, 768, 41938, 803, 730, 526, 521, 32647, 810, 768, 7493, 8386, 803, 730, 534, 521, 19222, 523, 941, 12324, 1320, 1381, 20425, 968, 902, 7054, 1924, 5725, 521, 1316, 1833, 1295, 4662, 523, 32826, 12043, 779, 12324, 968, 768, 1914, 803, 6108, 521, 916, 1295, 20402, 2723, 523, 509, 530, 523, 32964, 988, 779, 3822, 3892, 708, 26691, 2514, 843, 549, 37986, 3271, 779, 2514, 21440, 730, 5411, 8267, 35304, 884, 779, 1375, 977, 5329, 779, 2254, 810, 50398, 1547, 803, 779, 2759, 5854, 20842, 884, 779, 5253, 523, 1311, 1499, 1672, 811, 8153, 779, 8008, 3509, 875, 21275, 4476, 797, 779, 6844, 810, 856, 5400, 875, 1547, 1209, 20470, 819, 30193, 2514, 843, 545, 6310, 793, 3271, 29504, 779, 6844, 7405, 521, 779, 2514, 9657, 768, 30780, 4915, 797, 4843, 896, 17322, 1919, 523, 835, 3316, 4915, 797, 779, 11917, 856, 13573, 875, 4471, 1288, 768, 35563, 1011, 874, 20560, 963, 768, 16336, 7264, 819, 35798, 4977, 810, 5600, 2514, 843, 549, 831, 833, 3271, 10325, 988, 779, 22212, 7982, 521, 779, 2514, 4468, 2204, 15073, 11513, 15489, 810, 768, 26781, 15761, 884, 13878, 9690, 811, 779, 17677, 536, 1547, 803, 779, 15489, 1052, 1381, 50398, 523, 509, 531, 523, 7329, 12562, 3549, 708, 559, 4182, 4468, 875, 779, 11917, 15671, 896, 779, 12562, 12940, 1178, 1381, 29878, 963, 730, 1133, 521, 6019, 523, 3877, 779, 730, 527, 521, 9300, 988, 779, 6844, 7405, 521, 779, 3549, 2166, 12973, 963, 730, 1632, 521, 14671, 523, 941, 15671, 1178, 6944, 811, 1547, 56302, 8810, 592, 875, 779, 2280, 810, 45777, 1149, 13041, 968, 779, 1764, 803, 3354, 523, 6162, 4584, 2155, 2076, 4526, 1929, 30541, 1232, 1000, 797, 779, 12562, 3494, 2013, 779, 6844, 521, 810, 902, 8535, 8923, 1188, 3999, 779, 12562, 2076, 1523, 1336, 10309, 523, 32826, 5751, 3984, 779, 10039, 1499, 3644, 1523, 3290, 8372, 536, 28682, 1845, 896, 779, 15671, 1499, 2743, 1892, 811, 768, 3177, 8324, 2076, 54526, 1523, 803, 779, 10039, 1090, 4636, 523, 3732, 5484, 9303, 811, 49904, 768, 32673, 5043, 875, 779, 3549, 797, 4963, 521, 810, 2993, 3967, 12547, 3501, 811, 1801, 523, 509, 532, 523, 53776, 803, 12900, 708, 548, 54069, 11860, 811, 2472, 2030, 906, 12377, 810, 953, 840, 5508, 3800, 946, 920, 4346, 523, 9413, 598, 19718, 953, 840, 5508, 3800, 906, 27246, 946, 920, 4346, 523, 2345, 586, 13682, 2208, 906, 3185, 802, 9233, 1720, 4250, 1779, 523, 8726, 3967, 1166, 33256, 875, 779, 2375, 535, 10593, 3102, 810, 15642, 523, 4162, 3290, 15450, 988, 2059, 16397, 521, 768, 7155, 48506, 953, 4908, 523, 10593, 3102, 5068, 730, 2235, 20402, 810, 15642, 730, 1443, 521, 916, 730, 527, 20069, 9244, 12452, 523, 10593, 3102, 953, 13096, 11510, 906, 3185, 802, 9233, 523, 941, 8324, 14954, 791, 2345, 586, 6248, 876, 875, 1210, 1779, 803, 3714, 523, 509, 533, 523, 41453, 730, 526, 535, 39391, 1465, 14122, 884, 779, 7314, 5253, 708, 2364, 4346, 968, 1030, 774, 521, 2341, 791, 968, 15821, 793, 535, 997, 7464, 779, 10039, 47882, 1287, 811, 730, 526, 521, 1086, 988, 1352, 25511, 884, 2759, 39391, 1465, 14122, 963, 779, 9028, 803, 779, 7314, 5253, 521, 3910, 811, 17677, 15001, 2927, 1030, 774, 10201, 896, 28285, 5419, 938, 6654, 979, 3213, 811, 779, 8746, 1327, 521, 1144, 30809, 779, 5799, 810, 12785, 779, 3789, 8034, 875, 10415, 824, 5569, 523, 3732, 5484, 5751, 3984, 779, 1465, 14122, 1499, 8524, 38883, 536, 1030, 774, 25723, 896, 1186, 1499, 2472, 797, 3160, 803, 779, 15565, 1090, 11404, 523, 1019, 2176, 535, 730, 2999, 875, 521, 730, 534, 2723, 521, 730, 532, 41754, 806, 1114, 523, 941, 9557, 953, 7818, 523, 509, 534, 523, 41453, 730, 527, 535, 8900, 22398, 22819, 708, 2364, 4346, 968, 779, 15671, 535, 997, 7464, 779, 8900, 22819, 27870, 988, 730, 1374, 811, 730, 1386, 1094, 8461, 988, 730, 526, 4471, 730, 1718, 531, 2927, 941, 3185, 802, 9233, 59339, 1845, 896, 1033, 1423, 1090, 41938, 4201, 902, 3761, 27759, 875, 2166, 523, 17857, 3967, 1845, 896, 768, 3585, 22819, 1499, 3151, 1612, 1361, 16671, 521, 4155, 797, 779, 1361, 1434, 1872, 884, 13878, 9690, 523, 1019, 2176, 535, 730, 1133, 875, 521, 730, 2117, 2723, 521, 730, 528, 41754, 806, 1114, 523, 941, 9557, 953, 17602, 521, 810, 779, 22819, 32660, 963, 730, 1374, 523, 509, 1155, 523, 41453, 730, 528, 535, 3681, 9041, 708, 2364, 4346, 968, 38992, 1068, 521, 2341, 791, 968, 43842, 535, 997, 7464, 779, 10039, 12881, 768, 3681, 9041, 884, 779, 52352, 13646, 5598, 779, 11917, 2927, 32826, 7951, 779, 4588, 797, 11618, 1203, 5751, 1316, 1499, 2144, 1720, 779, 9041, 810, 3984, 779, 11917, 15671, 40917, 779, 13646, 523, 1311, 953, 11860, 811, 2021, 1287, 768, 3822, 2514, 803, 38992, 1068, 521, 43842, 810, 15642, 811, 2312, 811, 779, 3280, 3177, 8324, 523, 3253, 11386, 953, 4281, 523, 509, 1632, 523, 10972, 1312, 3436, 708, 13792, 518, 7223, 4729, 884, 8995, 27784, 535, 2993, 3967, 4468, 11242, 10978, 1714, 3537, 779, 5630, 963, 2254, 3217, 523, 941, 27246, 1188, 5773, 811, 779, 17677, 14065, 875, 779, 3354, 8153, 811, 2957, 6586, 1818, 819, 10388, 518, 54894, 535, 20470, 938, 4602, 811, 5514, 884, 779, 9168, 3392, 803, 779, 7314, 819, 10120, 518, 941, 29780, 12548, 50804, 777, 1845, 896, 779, 17677, 1090, 8825, 14018, 875, 3681, 7092, 21981, 797, 4963, 810, 896, 779, 9041, 2724, 33460, 819, 9285, 518, 1416, 48053, 535, 779, 17677, 1188, 5840, 7980, 23878, 988, 5105, 536, 779, 6111, 3378, 1188, 874, 16022, 797, 779, 3280, 38666, 819, 11634, 518, 8528, 3711, 937, 1888, 535, 9213, 833, 4468, 896, 779, 1629, 3711, 937, 1888, 5329, 779, 15565, 953, 20425, 797, 6060, 810, 896, 1352, 46967, 2716, 1046, 797, 5277, 536, 779, 15671, 1188, 2565, 8297, 1667, 523, 509, 1374, 523, 13887, 803, 779, 3280, 8324, 708, 1098, 3280, 3177, 8324, 1188, 874, 4908, 797, 4526, 536, 779, 5562, 1188, 874, 9139, 797, 779, 6844, 38666, 523, 941, 8324, 8557, 963, 730, 2197, 535, 1155, 523, 18, 17408, 953, 11510, 3185, 802, 9233, 540, 19, 16, 3352, 3102, 535, 10593, 3102, 20, 19, 16, 945, 596, 783, 535, 15642, 20, 21 ], "markers": [ 1424, 1432 ], "logits_raw": [ -0.2319762408733368, -2.207134246826172 ], "probs_raw": [ 0.8781641125679016, 0.12183595448732376 ], "probs": [ 0.7560074925422668, 0.24399243295192719 ], "argmax_index": 0, "argmax": "odette", "top2_margin": 0.5120150595903397, "near_tie": false, "gold": "odette", "correct": true, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 3, "B": 4 } }, { "qid": "racks_support", "type": "score", "K": 4, "calibrate": true, "temperature_key": "score:3-5", "T": 1.7301132678985596, "prefix": 0, "n_ids": 1466, "positions": 1466, "bucket": 2048, "ids": [ 1, 17, 564, 17769, 3613, 4425, 12411, 34632, 2730, 41891, 516, 31289, 7272, 550, 10877, 708, 13635, 3962, 803, 779, 27845, 5525, 35590, 4908, 884, 22045, 730, 1374, 4690, 730, 1718, 531, 963, 730, 913, 535, 1664, 797, 33864, 54898, 2176, 24579, 7329, 509, 41678, 535, 730, 3745, 3967, 843, 1615, 959, 1143, 13975, 12930, 518, 810, 730, 528, 24753, 521, 3136, 1377, 779, 29780, 12548, 50804, 777, 521, 2100, 3852, 1543, 523, 1030, 779, 12377, 535, 918, 54069, 523, 3617, 3962, 535, 9413, 598, 19718, 843, 52709, 1126, 1635, 10412, 6650, 535, 730, 1632, 3967, 3567, 1149, 1514, 6650, 523, 509, 526, 523, 56975, 810, 967, 24339, 708, 1098, 12377, 41975, 3967, 810, 24753, 810, 14954, 791, 779, 11917, 15671, 875, 779, 1517, 803, 779, 5362, 523, 941, 12778, 6835, 730, 1664, 3967, 811, 874, 2251, 875, 768, 3177, 8324, 811, 874, 967, 777, 931, 523, 3877, 730, 3745, 3967, 2251, 521, 779, 8324, 953, 13096, 967, 777, 931, 523, 509, 527, 523, 3617, 3962, 803, 779, 2659, 8900, 3177, 8324, 708, 1098, 5112, 803, 779, 8324, 4908, 884, 730, 1532, 4690, 730, 1718, 530, 1166, 12043, 906, 768, 3347, 3084, 521, 7885, 968, 9213, 833, 810, 2341, 791, 968, 28682, 523, 2479, 1166, 1295, 14332, 35517, 896, 1166, 1014, 4001, 884, 779, 26168, 523, 509, 528, 523, 21640, 1090, 2312, 708, 548, 54069, 17422, 2851, 779, 1423, 523, 941, 10039, 9657, 779, 6844, 7405, 521, 1144, 9388, 730, 527, 521, 9300, 875, 779, 11917, 12562, 3549, 521, 810, 35130, 1547, 27297, 45546, 3256, 779, 8008, 3509, 521, 1895, 916, 1209, 1399, 11583, 20470, 523, 941, 7982, 875, 768, 55097, 22212, 884, 8995, 27784, 6806, 521, 810, 779, 17677, 1178, 2166, 11860, 811, 6715, 1230, 768, 11013, 8153, 797, 3354, 523, 49427, 5313, 14118, 988, 730, 18833, 811, 730, 22989, 16671, 523, 1405, 14954, 791, 779, 15671, 810, 779, 20470, 1316, 5514, 779, 38666, 811, 1973, 3622, 2759, 3217, 768, 1423, 523, 509, 529, 523, 9312, 802, 9233, 1090, 2312, 708, 1098, 1230, 9932, 3185, 802, 9233, 521, 2345, 586, 521, 6892, 779, 12324, 875, 779, 1423, 523, 53813, 953, 730, 531, 521, 21957, 535, 22398, 23023, 730, 527, 521, 26547, 521, 779, 6844, 7405, 730, 527, 521, 9300, 810, 41634, 730, 526, 521, 30542, 523, 2315, 2124, 11761, 953, 730, 529, 521, 25803, 521, 8787, 38666, 17236, 843, 526, 521, 21405, 1109, 12235, 843, 23617, 1109, 11917, 33265, 843, 7537, 518, 810, 7494, 875, 779, 8008, 3509, 5158, 6688, 843, 526, 521, 26668, 1191, 941, 1423, 8557, 916, 768, 41938, 803, 730, 526, 521, 32647, 810, 768, 7493, 8386, 803, 730, 534, 521, 19222, 523, 941, 12324, 1320, 1381, 20425, 968, 902, 7054, 1924, 5725, 521, 1316, 1833, 1295, 4662, 523, 32826, 12043, 779, 12324, 968, 768, 1914, 803, 6108, 521, 916, 1295, 20402, 2723, 523, 509, 530, 523, 32964, 988, 779, 3822, 3892, 708, 26691, 2514, 843, 549, 37986, 3271, 779, 2514, 21440, 730, 5411, 8267, 35304, 884, 779, 1375, 977, 5329, 779, 2254, 810, 50398, 1547, 803, 779, 2759, 5854, 20842, 884, 779, 5253, 523, 1311, 1499, 1672, 811, 8153, 779, 8008, 3509, 875, 21275, 4476, 797, 779, 6844, 810, 856, 5400, 875, 1547, 1209, 20470, 819, 30193, 2514, 843, 545, 6310, 793, 3271, 29504, 779, 6844, 7405, 521, 779, 2514, 9657, 768, 30780, 4915, 797, 4843, 896, 17322, 1919, 523, 835, 3316, 4915, 797, 779, 11917, 856, 13573, 875, 4471, 1288, 768, 35563, 1011, 874, 20560, 963, 768, 16336, 7264, 819, 35798, 4977, 810, 5600, 2514, 843, 549, 831, 833, 3271, 10325, 988, 779, 22212, 7982, 521, 779, 2514, 4468, 2204, 15073, 11513, 15489, 810, 768, 26781, 15761, 884, 13878, 9690, 811, 779, 17677, 536, 1547, 803, 779, 15489, 1052, 1381, 50398, 523, 509, 531, 523, 7329, 12562, 3549, 708, 559, 4182, 4468, 875, 779, 11917, 15671, 896, 779, 12562, 12940, 1178, 1381, 29878, 963, 730, 1133, 521, 6019, 523, 3877, 779, 730, 527, 521, 9300, 988, 779, 6844, 7405, 521, 779, 3549, 2166, 12973, 963, 730, 1632, 521, 14671, 523, 941, 15671, 1178, 6944, 811, 1547, 56302, 8810, 592, 875, 779, 2280, 810, 45777, 1149, 13041, 968, 779, 1764, 803, 3354, 523, 6162, 4584, 2155, 2076, 4526, 1929, 30541, 1232, 1000, 797, 779, 12562, 3494, 2013, 779, 6844, 521, 810, 902, 8535, 8923, 1188, 3999, 779, 12562, 2076, 1523, 1336, 10309, 523, 32826, 5751, 3984, 779, 10039, 1499, 3644, 1523, 3290, 8372, 536, 28682, 1845, 896, 779, 15671, 1499, 2743, 1892, 811, 768, 3177, 8324, 2076, 54526, 1523, 803, 779, 10039, 1090, 4636, 523, 3732, 5484, 9303, 811, 49904, 768, 32673, 5043, 875, 779, 3549, 797, 4963, 521, 810, 2993, 3967, 12547, 3501, 811, 1801, 523, 509, 532, 523, 53776, 803, 12900, 708, 548, 54069, 11860, 811, 2472, 2030, 906, 12377, 810, 953, 840, 5508, 3800, 946, 920, 4346, 523, 9413, 598, 19718, 953, 840, 5508, 3800, 906, 27246, 946, 920, 4346, 523, 2345, 586, 13682, 2208, 906, 3185, 802, 9233, 1720, 4250, 1779, 523, 8726, 3967, 1166, 33256, 875, 779, 2375, 535, 10593, 3102, 810, 15642, 523, 4162, 3290, 15450, 988, 2059, 16397, 521, 768, 7155, 48506, 953, 4908, 523, 10593, 3102, 5068, 730, 2235, 20402, 810, 15642, 730, 1443, 521, 916, 730, 527, 20069, 9244, 12452, 523, 10593, 3102, 953, 13096, 11510, 906, 3185, 802, 9233, 523, 941, 8324, 14954, 791, 2345, 586, 6248, 876, 875, 1210, 1779, 803, 3714, 523, 509, 533, 523, 41453, 730, 526, 535, 39391, 1465, 14122, 884, 779, 7314, 5253, 708, 2364, 4346, 968, 1030, 774, 521, 2341, 791, 968, 15821, 793, 535, 997, 7464, 779, 10039, 47882, 1287, 811, 730, 526, 521, 1086, 988, 1352, 25511, 884, 2759, 39391, 1465, 14122, 963, 779, 9028, 803, 779, 7314, 5253, 521, 3910, 811, 17677, 15001, 2927, 1030, 774, 10201, 896, 28285, 5419, 938, 6654, 979, 3213, 811, 779, 8746, 1327, 521, 1144, 30809, 779, 5799, 810, 12785, 779, 3789, 8034, 875, 10415, 824, 5569, 523, 3732, 5484, 5751, 3984, 779, 1465, 14122, 1499, 8524, 38883, 536, 1030, 774, 25723, 896, 1186, 1499, 2472, 797, 3160, 803, 779, 15565, 1090, 11404, 523, 1019, 2176, 535, 730, 2999, 875, 521, 730, 534, 2723, 521, 730, 532, 41754, 806, 1114, 523, 941, 9557, 953, 7818, 523, 509, 534, 523, 41453, 730, 527, 535, 8900, 22398, 22819, 708, 2364, 4346, 968, 779, 15671, 535, 997, 7464, 779, 8900, 22819, 27870, 988, 730, 1374, 811, 730, 1386, 1094, 8461, 988, 730, 526, 4471, 730, 1718, 531, 2927, 941, 3185, 802, 9233, 59339, 1845, 896, 1033, 1423, 1090, 41938, 4201, 902, 3761, 27759, 875, 2166, 523, 17857, 3967, 1845, 896, 768, 3585, 22819, 1499, 3151, 1612, 1361, 16671, 521, 4155, 797, 779, 1361, 1434, 1872, 884, 13878, 9690, 523, 1019, 2176, 535, 730, 1133, 875, 521, 730, 2117, 2723, 521, 730, 528, 41754, 806, 1114, 523, 941, 9557, 953, 17602, 521, 810, 779, 22819, 32660, 963, 730, 1374, 523, 509, 1155, 523, 41453, 730, 528, 535, 3681, 9041, 708, 2364, 4346, 968, 38992, 1068, 521, 2341, 791, 968, 43842, 535, 997, 7464, 779, 10039, 12881, 768, 3681, 9041, 884, 779, 52352, 13646, 5598, 779, 11917, 2927, 32826, 7951, 779, 4588, 797, 11618, 1203, 5751, 1316, 1499, 2144, 1720, 779, 9041, 810, 3984, 779, 11917, 15671, 40917, 779, 13646, 523, 1311, 953, 11860, 811, 2021, 1287, 768, 3822, 2514, 803, 38992, 1068, 521, 43842, 810, 15642, 811, 2312, 811, 779, 3280, 3177, 8324, 523, 3253, 11386, 953, 4281, 523, 509, 1632, 523, 10972, 1312, 3436, 708, 13792, 518, 7223, 4729, 884, 8995, 27784, 535, 2993, 3967, 4468, 11242, 10978, 1714, 3537, 779, 5630, 963, 2254, 3217, 523, 941, 27246, 1188, 5773, 811, 779, 17677, 14065, 875, 779, 3354, 8153, 811, 2957, 6586, 1818, 819, 10388, 518, 54894, 535, 20470, 938, 4602, 811, 5514, 884, 779, 9168, 3392, 803, 779, 7314, 819, 10120, 518, 941, 29780, 12548, 50804, 777, 1845, 896, 779, 17677, 1090, 8825, 14018, 875, 3681, 7092, 21981, 797, 4963, 810, 896, 779, 9041, 2724, 33460, 819, 9285, 518, 1416, 48053, 535, 779, 17677, 1188, 5840, 7980, 23878, 988, 5105, 536, 779, 6111, 3378, 1188, 874, 16022, 797, 779, 3280, 38666, 819, 11634, 518, 8528, 3711, 937, 1888, 535, 9213, 833, 4468, 896, 779, 1629, 3711, 937, 1888, 5329, 779, 15565, 953, 20425, 797, 6060, 810, 896, 1352, 46967, 2716, 1046, 797, 5277, 536, 779, 15671, 1188, 2565, 8297, 1667, 523, 509, 1374, 523, 13887, 803, 779, 3280, 8324, 708, 1098, 3280, 3177, 8324, 1188, 874, 4908, 797, 4526, 536, 779, 5562, 1188, 874, 9139, 797, 779, 6844, 38666, 523, 941, 8324, 8557, 963, 730, 2197, 535, 1155, 523, 18, 4023, 3636, 953, 779, 2258, 875, 779, 39391, 1465, 14122, 9557, 540, 19, 16, 2119, 730, 525, 535, 14873, 968, 1529, 20, 19, 16, 2119, 730, 526, 535, 13305, 20, 19, 16, 2119, 730, 527, 535, 7951, 968, 1529, 20, 19, 16, 2119, 730, 528, 535, 42869, 1069, 20, 21 ], "markers": [ 1429, 1439, 1447, 1457 ], "logits_raw": [ -4.230611324310303, 0.1285986751317978, -1.4657516479492188, -2.841369867324829 ], "probs_raw": [ 0.010092451237142086, 0.7891826033592224, 0.1602359414100647, 0.04048895090818405 ], "probs": [ 0.04854496940970421, 0.6031103730201721, 0.23998364806175232, 0.10836095362901688 ], "argmax_index": 1, "argmax": "1", "top2_margin": 0.3631267249584198, "near_tie": false, "gold": "2", "correct": false, "host_softmax_max_abs_dp": 7.450580596923828e-09, "batch": { "call": 0, "index": 0, "B": 4 } } ], "response": { "answers": { "fee_motion": { "type": "choice", "choice": "deferred", "confidence": 0.6915529370307922, "probabilities": { "approved": 0.15077616274356842, "rejected": 0.0647965595126152, "deferred": 0.6915529370307922, "withdrawn": 0.09287437796592712 } }, "quorate": { "type": "noul", "noul": 0.9495782256126404 }, "treasurer": { "type": "choice", "choice": "odette", "confidence": 0.7560074925422668, "probabilities": { "odette": 0.7560074925422668, "lewis": 0.24399243295192719 } }, "racks_support": { "type": "score", "score": 1.4081605300307274, "confidence": 0.6031103730201721, "probabilities": { "0": 0.04854496940970421, "1": 0.6031103730201721, "2": 0.23998364806175232, "3": 0.10836095362901688 }, "legend": { "0": "opposed by most", "1": "split", "2": "supported by most", "3": "unanimous" } } }, "usage": { "input_tokens": 5815, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "own_a01", "mode": "text", "source": "own_json", "native": true, "prefix": 0, "usage_input_tokens": 618, "questions": [ { "qid": "first", "type": "choice", "K": 4, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 220, "positions": 220, "bucket": 256, "ids": [ 1, 17, 568, 38685, 593, 32909, 6881, 997, 13806, 522, 5768, 1878, 997, 7820, 6881, 997, 5059, 2874, 884, 8913, 730, 527, 996, 2534, 884, 1973, 4426, 1878, 997, 782, 29500, 6881, 997, 4905, 535, 3654, 1878, 997, 3165, 1012, 592, 6881, 997, 526, 3410, 511, 3205, 37080, 593, 32909, 6881, 997, 13806, 522, 9270, 1878, 997, 7820, 6881, 997, 29408, 13040, 963, 779, 2120, 19025, 22879, 1113, 5663, 3273, 1878, 997, 782, 29500, 6881, 997, 4905, 535, 1374, 1878, 997, 3165, 1012, 592, 6881, 997, 2291, 1912, 4204, 511, 3205, 37080, 593, 32909, 6881, 997, 13806, 522, 8884, 1878, 997, 7820, 6881, 997, 7367, 875, 768, 2341, 6562, 1878, 997, 782, 29500, 6881, 997, 4905, 535, 1664, 1878, 997, 3165, 1012, 592, 6881, 997, 526, 1811, 511, 3205, 37080, 593, 32909, 6881, 997, 13806, 522, 10981, 1878, 997, 7820, 6881, 997, 14100, 965, 803, 15658, 8997, 3627, 779, 9397, 5362, 1878, 997, 782, 29500, 6881, 997, 4905, 535, 4536, 1878, 997, 3165, 1012, 592, 6881, 997, 51493, 511, 20352, 18, 28828, 30437, 1753, 874, 28010, 1509, 540, 19, 16, 17905, 522, 5768, 535, 2343, 2874, 996, 2534, 20, 19, 16, 17905, 522, 9270, 535, 5719, 20391, 5060, 13040, 20, 19, 16, 17905, 522, 8884, 535, 8296, 6562, 20, 19, 16, 17905, 522, 10981, 535, 23250, 800, 19944, 20, 21 ], "markers": [ 179, 190, 201, 210 ], "logits_raw": [ 0.3843047022819519, -2.1585264205932617, -0.326029896736145, 0.20786070823669434 ], "probs_raw": [ 0.4152187407016754, 0.03265422582626343, 0.20407159626483917, 0.34805533289909363 ], "probs": [ 0.3778960704803467, 0.061447251588106155, 0.2275114357471466, 0.33314526081085205 ], "argmax_index": 0, "argmax": "HD-101", "top2_margin": 0.04475080966949463, "near_tie": false, "gold": "HD-104", "correct": false, "host_softmax_max_abs_dp": 2.9802322387695312e-08, "batch": { "call": 0, "index": 0, "B": 3 } }, { "qid": "hazard", "type": "noul", "K": 2, "calibrate": true, "temperature_key": "noul:2", "T": 1.6663223505020142, "prefix": 0, "n_ids": 203, "positions": 203, "bucket": 256, "ids": [ 1, 17, 568, 38685, 593, 32909, 6881, 997, 13806, 522, 5768, 1878, 997, 7820, 6881, 997, 5059, 2874, 884, 8913, 730, 527, 996, 2534, 884, 1973, 4426, 1878, 997, 782, 29500, 6881, 997, 4905, 535, 3654, 1878, 997, 3165, 1012, 592, 6881, 997, 526, 3410, 511, 3205, 37080, 593, 32909, 6881, 997, 13806, 522, 9270, 1878, 997, 7820, 6881, 997, 29408, 13040, 963, 779, 2120, 19025, 22879, 1113, 5663, 3273, 1878, 997, 782, 29500, 6881, 997, 4905, 535, 1374, 1878, 997, 3165, 1012, 592, 6881, 997, 2291, 1912, 4204, 511, 3205, 37080, 593, 32909, 6881, 997, 13806, 522, 8884, 1878, 997, 7820, 6881, 997, 7367, 875, 768, 2341, 6562, 1878, 997, 782, 29500, 6881, 997, 4905, 535, 1664, 1878, 997, 3165, 1012, 592, 6881, 997, 526, 1811, 511, 3205, 37080, 593, 32909, 6881, 997, 13806, 522, 10981, 1878, 997, 7820, 6881, 997, 14100, 965, 803, 15658, 8997, 3627, 779, 9397, 5362, 1878, 997, 782, 29500, 6881, 997, 4905, 535, 4536, 1878, 997, 3165, 1012, 592, 6881, 997, 51493, 511, 20352, 18, 29589, 1523, 30437, 8487, 768, 3023, 6155, 17745, 540, 19, 16, 4831, 535, 1295, 521, 779, 8294, 2177, 1014, 3965, 20, 19, 16, 3347, 535, 18672, 521, 779, 8294, 10995, 20, 21 ], "markers": [ 181, 193 ], "logits_raw": [ 2.477010488510132, 6.635200023651123 ], "probs_raw": [ 0.015395121648907661, 0.9846048951148987 ], "probs": [ 0.9238207936286926, 0.07617922872304916 ], "argmax_index": 0, "argmax": "true", "top2_margin": 0.8476415649056435, "near_tie": false, "gold": "true", "correct": true, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 1, "B": 3 } }, { "qid": "load", "type": "score", "K": 2, "calibrate": true, "temperature_key": "score:2", "T": 1.0, "prefix": 0, "n_ids": 195, "positions": 195, "bucket": 256, "ids": [ 1, 17, 568, 38685, 593, 32909, 6881, 997, 13806, 522, 5768, 1878, 997, 7820, 6881, 997, 5059, 2874, 884, 8913, 730, 527, 996, 2534, 884, 1973, 4426, 1878, 997, 782, 29500, 6881, 997, 4905, 535, 3654, 1878, 997, 3165, 1012, 592, 6881, 997, 526, 3410, 511, 3205, 37080, 593, 32909, 6881, 997, 13806, 522, 9270, 1878, 997, 7820, 6881, 997, 29408, 13040, 963, 779, 2120, 19025, 22879, 1113, 5663, 3273, 1878, 997, 782, 29500, 6881, 997, 4905, 535, 1374, 1878, 997, 3165, 1012, 592, 6881, 997, 2291, 1912, 4204, 511, 3205, 37080, 593, 32909, 6881, 997, 13806, 522, 8884, 1878, 997, 7820, 6881, 997, 7367, 875, 768, 2341, 6562, 1878, 997, 782, 29500, 6881, 997, 4905, 535, 1664, 1878, 997, 3165, 1012, 592, 6881, 997, 526, 1811, 511, 3205, 37080, 593, 32909, 6881, 997, 13806, 522, 10981, 1878, 997, 7820, 6881, 997, 14100, 965, 803, 15658, 8997, 3627, 779, 9397, 5362, 1878, 997, 782, 29500, 6881, 997, 4905, 535, 4536, 1878, 997, 3165, 1012, 592, 6881, 997, 51493, 511, 20352, 18, 4940, 779, 27838, 1391, 38596, 540, 19, 16, 2119, 730, 525, 535, 4381, 20, 19, 16, 2119, 730, 526, 535, 1391, 38596, 20, 21 ], "markers": [ 178, 186 ], "logits_raw": [ -2.7285783290863037, 2.443589687347412 ], "probs_raw": [ 0.005640266463160515, 0.9943597912788391 ], "probs": [ 0.005640266463160515, 0.9943597912788391 ], "argmax_index": 1, "argmax": "1", "top2_margin": 0.9887195248156786, "near_tie": false, "gold": "0", "correct": false, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 2, "B": 3 } } ], "response": { "answers": { "first": { "type": "choice", "choice": "HD-101", "confidence": 0.3778960704803467, "probabilities": { "HD-101": 0.3778960704803467, "HD-102": 0.061447251588106155, "HD-103": 0.2275114357471466, "HD-104": 0.33314526081085205 } }, "hazard": { "type": "noul", "noul": 0.9238207936286926 }, "load": { "type": "score", "score": 0.9943597912788391, "confidence": 0.9943597912788391, "probabilities": { "0": 0.005640266463160515, "1": 0.9943597912788391 }, "legend": { "0": "normal", "1": "overloaded" } } }, "usage": { "input_tokens": 618, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "own_j01", "mode": "text", "source": "own_json", "native": true, "prefix": 0, "usage_input_tokens": 762, "questions": [ { "qid": "overdue", "type": "noul", "K": 2, "calibrate": true, "temperature_key": "noul:2", "T": 1.6663223505020142, "prefix": 0, "n_ids": 247, "positions": 247, "bucket": 256, "ids": [ 1, 17, 38685, 769, 47958, 6881, 997, 15807, 564, 522, 1718, 531, 522, 30907, 527, 1878, 997, 1271, 2466, 6881, 997, 547, 1983, 893, 1174, 18119, 16643, 1878, 997, 38765, 770, 6881, 997, 556, 820, 799, 12890, 12519, 1157, 4847, 1878, 997, 1271, 2686, 6881, 997, 1718, 531, 522, 6326, 522, 5843, 1878, 997, 30717, 6881, 997, 1718, 531, 522, 3654, 522, 5843, 1878, 997, 19248, 6881, 997, 885, 38008, 1878, 997, 2481, 53339, 4294, 806, 6881, 730, 527, 521, 997, 10890, 6881, 1607, 38685, 5027, 6881, 997, 2414, 29248, 47635, 835, 527, 1878, 997, 590, 1937, 6881, 730, 2108, 521, 997, 10829, 572, 51363, 6881, 730, 526, 523, 3520, 3205, 37080, 5027, 6881, 997, 19938, 1477, 2643, 10819, 1878, 997, 590, 1937, 6881, 730, 7896, 521, 997, 10829, 572, 51363, 6881, 730, 525, 523, 5370, 3205, 37080, 5027, 6881, 997, 19589, 36603, 1878, 997, 590, 1937, 6881, 730, 526, 521, 997, 10829, 572, 51363, 6881, 730, 1374, 523, 525, 602, 5086, 997, 4093, 35377, 6881, 730, 22277, 523, 525, 521, 997, 14623, 572, 10233, 6881, 730, 525, 523, 527, 521, 997, 35377, 6881, 730, 25633, 523, 527, 521, 997, 27317, 957, 572, 55913, 6881, 997, 18703, 6337, 2638, 730, 1664, 3378, 1878, 997, 593, 53277, 6881, 997, 1718, 531, 522, 3654, 522, 913, 511, 602, 18, 4940, 779, 60827, 1139, 1391, 30717, 540, 19, 16, 4831, 535, 1295, 521, 779, 8294, 2177, 1014, 3965, 20, 19, 16, 3347, 535, 18672, 521, 779, 8294, 10995, 20, 21 ], "markers": [ 225, 237 ], "logits_raw": [ -0.45539525151252747, 1.4913076162338257 ], "probs_raw": [ 0.12491331994533539, 0.8750866651535034 ], "probs": [ 0.762830913066864, 0.237169086933136 ], "argmax_index": 0, "argmax": "true", "top2_margin": 0.525661826133728, "near_tie": false, "gold": "true", "correct": true, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 1, "B": 3 } }, { "qid": "method", "type": "choice", "K": 4, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 245, "positions": 245, "bucket": 256, "ids": [ 1, 17, 38685, 769, 47958, 6881, 997, 15807, 564, 522, 1718, 531, 522, 30907, 527, 1878, 997, 1271, 2466, 6881, 997, 547, 1983, 893, 1174, 18119, 16643, 1878, 997, 38765, 770, 6881, 997, 556, 820, 799, 12890, 12519, 1157, 4847, 1878, 997, 1271, 2686, 6881, 997, 1718, 531, 522, 6326, 522, 5843, 1878, 997, 30717, 6881, 997, 1718, 531, 522, 3654, 522, 5843, 1878, 997, 19248, 6881, 997, 885, 38008, 1878, 997, 2481, 53339, 4294, 806, 6881, 730, 527, 521, 997, 10890, 6881, 1607, 38685, 5027, 6881, 997, 2414, 29248, 47635, 835, 527, 1878, 997, 590, 1937, 6881, 730, 2108, 521, 997, 10829, 572, 51363, 6881, 730, 526, 523, 3520, 3205, 37080, 5027, 6881, 997, 19938, 1477, 2643, 10819, 1878, 997, 590, 1937, 6881, 730, 7896, 521, 997, 10829, 572, 51363, 6881, 730, 525, 523, 5370, 3205, 37080, 5027, 6881, 997, 19589, 36603, 1878, 997, 590, 1937, 6881, 730, 526, 521, 997, 10829, 572, 51363, 6881, 730, 1374, 523, 525, 602, 5086, 997, 4093, 35377, 6881, 730, 22277, 523, 525, 521, 997, 14623, 572, 10233, 6881, 730, 525, 523, 527, 521, 997, 35377, 6881, 730, 25633, 523, 527, 521, 997, 27317, 957, 572, 55913, 6881, 997, 18703, 6337, 2638, 730, 1664, 3378, 1878, 997, 593, 53277, 6881, 997, 1718, 531, 522, 3654, 522, 913, 511, 602, 18, 4023, 856, 779, 13608, 5559, 811, 4025, 540, 19, 16, 5060, 20, 19, 16, 7493, 41953, 1100, 20, 19, 16, 14184, 20, 19, 16, 1636, 1357, 20, 21 ], "markers": [ 226, 230, 236, 240 ], "logits_raw": [ -9.03740406036377, 1.9014228582382202, -6.792163848876953, -6.893122673034668 ], "probs_raw": [ 1.774932025000453e-05, 0.999663233757019, 0.00016760120342951268, 0.0001515065087005496 ], "probs": [ 0.00040230932063423097, 0.9957361221313477, 0.0020003572572022676, 0.001861173310317099 ], "argmax_index": 1, "argmax": "bank_transfer", "top2_margin": 0.9937357648741454, "near_tie": false, "gold": "bank_transfer", "correct": true, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 2, "B": 3 } }, { "qid": "size", "type": "score", "K": 4, "calibrate": true, "temperature_key": "score:3-5", "T": 1.7301132678985596, "prefix": 0, "n_ids": 270, "positions": 270, "bucket": 512, "ids": [ 1, 17, 38685, 769, 47958, 6881, 997, 15807, 564, 522, 1718, 531, 522, 30907, 527, 1878, 997, 1271, 2466, 6881, 997, 547, 1983, 893, 1174, 18119, 16643, 1878, 997, 38765, 770, 6881, 997, 556, 820, 799, 12890, 12519, 1157, 4847, 1878, 997, 1271, 2686, 6881, 997, 1718, 531, 522, 6326, 522, 5843, 1878, 997, 30717, 6881, 997, 1718, 531, 522, 3654, 522, 5843, 1878, 997, 19248, 6881, 997, 885, 38008, 1878, 997, 2481, 53339, 4294, 806, 6881, 730, 527, 521, 997, 10890, 6881, 1607, 38685, 5027, 6881, 997, 2414, 29248, 47635, 835, 527, 1878, 997, 590, 1937, 6881, 730, 2108, 521, 997, 10829, 572, 51363, 6881, 730, 526, 523, 3520, 3205, 37080, 5027, 6881, 997, 19938, 1477, 2643, 10819, 1878, 997, 590, 1937, 6881, 730, 7896, 521, 997, 10829, 572, 51363, 6881, 730, 525, 523, 5370, 3205, 37080, 5027, 6881, 997, 19589, 36603, 1878, 997, 590, 1937, 6881, 730, 526, 521, 997, 10829, 572, 51363, 6881, 730, 1374, 523, 525, 602, 5086, 997, 4093, 35377, 6881, 730, 22277, 523, 525, 521, 997, 14623, 572, 10233, 6881, 730, 525, 523, 527, 521, 997, 35377, 6881, 730, 25633, 523, 527, 521, 997, 27317, 957, 572, 55913, 6881, 997, 18703, 6337, 2638, 730, 1664, 3378, 1878, 997, 593, 53277, 6881, 997, 1718, 531, 522, 3654, 522, 913, 511, 602, 18, 4023, 2611, 856, 779, 60827, 1139, 3426, 540, 19, 16, 2119, 730, 525, 535, 1490, 730, 2096, 20, 19, 16, 2119, 730, 526, 535, 730, 2096, 811, 730, 7896, 20, 19, 16, 2119, 730, 527, 535, 730, 7896, 811, 730, 4291, 20, 19, 16, 2119, 730, 528, 535, 1391, 730, 4291, 20, 21 ], "markers": [ 226, 236, 248, 260 ], "logits_raw": [ -15.045595169067383, -6.663366317749023, -6.0419511795043945, -7.225985527038574 ], "probs_raw": [ 6.670520815532655e-05, 0.2914173901081085, 0.5424911379814148, 0.1660248339176178 ], "probs": [ 0.002488102298229933, 0.316214919090271, 0.4528668224811554, 0.22843007743358612 ], "argmax_index": 2, "argmax": "2", "top2_margin": 0.1366519033908844, "near_tie": false, "gold": "2", "correct": true, "host_softmax_max_abs_dp": 2.3283064365386963e-10, "batch": { "call": 0, "index": 0, "B": 3 } } ], "response": { "answers": { "overdue": { "type": "noul", "noul": 0.762830913066864 }, "method": { "type": "choice", "choice": "bank_transfer", "confidence": 0.9957361221313477, "probabilities": { "card": 0.00040230932063423097, "bank_transfer": 0.9957361221313477, "cash": 0.0020003572572022676, "cheque": 0.001861173310317099 } }, "size": { "type": "score", "score": 1.9072387963533401, "confidence": 0.4528668224811554, "probabilities": { "0": 0.002488102298229933, "1": 0.316214919090271, "2": 0.4528668224811554, "3": 0.22843007743358612 }, "legend": { "0": "under 100", "1": "100 to 250", "2": "250 to 500", "3": "over 500" } } }, "usage": { "input_tokens": 762, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "own_j02", "mode": "text", "source": "own_json", "native": true, "prefix": 0, "usage_input_tokens": 931, "questions": [ { "qid": "next_action", "type": "choice", "K": 10, "calibrate": true, "temperature_key": "choice:6-10", "T": 1.1751071214675903, "prefix": 0, "n_ids": 378, "positions": 378, "bucket": 512, "ids": [ 1, 17, 38685, 3667, 6590, 6881, 997, 42468, 522, 26260, 1532, 1878, 997, 15162, 6881, 997, 557, 1983, 834, 917, 38281, 1878, 997, 57267, 6881, 997, 1718, 531, 522, 6468, 522, 6135, 730, 1133, 535, 2019, 1878, 997, 26207, 6881, 1607, 38685, 38017, 6881, 997, 556, 2508, 522, 526, 552, 548, 1878, 997, 3055, 6881, 997, 46553, 791, 63437, 730, 526, 14895, 1878, 997, 590, 1937, 6881, 730, 527, 521, 997, 19248, 6881, 997, 2354, 9487, 511, 3205, 37080, 38017, 6881, 997, 549, 41555, 522, 14819, 1878, 997, 3055, 6881, 997, 41388, 32849, 17125, 730, 14819, 836, 1878, 997, 590, 1937, 6881, 730, 526, 521, 997, 19248, 6881, 997, 1066, 803, 9051, 511, 3205, 37080, 38017, 6881, 997, 12411, 542, 4894, 17727, 1878, 997, 3055, 6881, 997, 26818, 13192, 521, 730, 1838, 22173, 1878, 997, 590, 1937, 6881, 730, 528, 521, 997, 19248, 6881, 997, 2354, 9487, 511, 602, 5086, 997, 2354, 11468, 6881, 37080, 8618, 6265, 6881, 997, 2482, 1930, 2286, 1145, 1878, 997, 19132, 800, 36619, 6881, 997, 14402, 1991, 791, 1878, 997, 14402, 1991, 791, 40040, 6881, 997, 1718, 531, 522, 6468, 522, 3654, 730, 1632, 535, 2108, 511, 3205, 997, 27317, 957, 6881, 37080, 14164, 6881, 997, 19563, 1878, 997, 576, 2922, 3107, 6881, 730, 2142, 523, 529, 521, 997, 1891, 1356, 791, 6881, 730, 525, 523, 525, 3205, 997, 38765, 770, 5205, 2176, 6881, 997, 15090, 856, 1727, 17125, 540, 859, 953, 15418, 875, 936, 2927, 602, 18, 3493, 1753, 779, 15565, 1112, 3280, 540, 19, 16, 53467, 5377, 45556, 43840, 535, 12620, 1356, 779, 8688, 896, 953, 1014, 3567, 20, 19, 16, 1039, 959, 572, 3667, 535, 26995, 779, 4394, 2565, 2030, 20, 19, 16, 5254, 572, 1608, 5072, 535, 18388, 779, 2286, 1145, 1695, 779, 61026, 856, 20, 19, 16, 23496, 572, 3667, 535, 2577, 5072, 779, 2565, 20, 19, 16, 2785, 53220, 535, 23361, 779, 11746, 4332, 20, 19, 16, 25668, 30024, 535, 27908, 768, 25668, 5214, 20, 19, 16, 8297, 4370, 831, 5278, 535, 1416, 4128, 15073, 11391, 20, 19, 16, 10547, 4370, 808, 36603, 535, 27589, 779, 13608, 779, 61026, 953, 14465, 20, 19, 16, 24744, 15623, 2265, 535, 32655, 779, 2565, 875, 24744, 4202, 20, 19, 16, 1295, 55692, 535, 28657, 3988, 811, 874, 4007, 20, 21 ], "markers": [ 251, 267, 280, 296, 307, 317, 327, 339, 354, 367 ], "logits_raw": [ -1.7629752159118652, -6.255130290985107, 0.7563826441764832, -5.840479373931885, -10.807558059692383, -8.814050674438477, -13.214742660522461, -4.54431676864624, -5.7599897384643555, -2.928925037384033 ], "probs_raw": [ 0.07224540412425995, 0.0008088949834927917, 0.8973326086997986, 0.001224539359100163, 8.527027603122406e-06, 6.259893416427076e-05, 7.6801677550975e-07, 0.004475989378988743, 0.0013271773932501674, 0.022513583302497864 ], "probs": [ 0.09913931041955948, 0.0021679105702787638, 0.8459538817405701, 0.0030852300114929676, 4.503615491557866e-05, 0.0002456512884236872, 5.806538865726907e-06, 0.009296542964875698, 0.0033039606641978025, 0.03675662726163864 ], "argmax_index": 2, "argmax": "track_parcel", "top2_margin": 0.7468145713210106, "near_tie": false, "gold": "refund_missing_item", "correct": false, "host_softmax_max_abs_dp": 9.313225746154785e-10, "batch": { "call": 0, "index": 0, "B": 3 } }, { "qid": "partial", "type": "noul", "K": 2, "calibrate": true, "temperature_key": "noul:2", "T": 1.6663223505020142, "prefix": 0, "n_ids": 274, "positions": 274, "bucket": 512, "ids": [ 1, 17, 38685, 3667, 6590, 6881, 997, 42468, 522, 26260, 1532, 1878, 997, 15162, 6881, 997, 557, 1983, 834, 917, 38281, 1878, 997, 57267, 6881, 997, 1718, 531, 522, 6468, 522, 6135, 730, 1133, 535, 2019, 1878, 997, 26207, 6881, 1607, 38685, 38017, 6881, 997, 556, 2508, 522, 526, 552, 548, 1878, 997, 3055, 6881, 997, 46553, 791, 63437, 730, 526, 14895, 1878, 997, 590, 1937, 6881, 730, 527, 521, 997, 19248, 6881, 997, 2354, 9487, 511, 3205, 37080, 38017, 6881, 997, 549, 41555, 522, 14819, 1878, 997, 3055, 6881, 997, 41388, 32849, 17125, 730, 14819, 836, 1878, 997, 590, 1937, 6881, 730, 526, 521, 997, 19248, 6881, 997, 1066, 803, 9051, 511, 3205, 37080, 38017, 6881, 997, 12411, 542, 4894, 17727, 1878, 997, 3055, 6881, 997, 26818, 13192, 521, 730, 1838, 22173, 1878, 997, 590, 1937, 6881, 730, 528, 521, 997, 19248, 6881, 997, 2354, 9487, 511, 602, 5086, 997, 2354, 11468, 6881, 37080, 8618, 6265, 6881, 997, 2482, 1930, 2286, 1145, 1878, 997, 19132, 800, 36619, 6881, 997, 14402, 1991, 791, 1878, 997, 14402, 1991, 791, 40040, 6881, 997, 1718, 531, 522, 6468, 522, 3654, 730, 1632, 535, 2108, 511, 3205, 997, 27317, 957, 6881, 37080, 14164, 6881, 997, 19563, 1878, 997, 576, 2922, 3107, 6881, 730, 2142, 523, 529, 521, 997, 1891, 1356, 791, 6881, 730, 525, 523, 525, 3205, 997, 38765, 770, 5205, 2176, 6881, 997, 15090, 856, 1727, 17125, 540, 859, 953, 15418, 875, 936, 2927, 602, 18, 28519, 1550, 1192, 803, 779, 2565, 39993, 540, 19, 16, 4831, 535, 1295, 521, 779, 8294, 2177, 1014, 3965, 20, 19, 16, 3347, 535, 18672, 521, 779, 8294, 10995, 20, 21 ], "markers": [ 252, 264 ], "logits_raw": [ 1.2194080352783203, 5.291305065155029 ], "probs_raw": [ 0.0167593602091074, 0.9832406640052795 ], "probs": [ 0.9200953841209412, 0.07990466803312302 ], "argmax_index": 0, "argmax": "true", "top2_margin": 0.8401907160878181, "near_tie": false, "gold": "true", "correct": true, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 2, "B": 3 } }, { "qid": "charged_claim", "type": "noul", "K": 2, "calibrate": true, "temperature_key": "noul:2", "T": 1.6663223505020142, "prefix": 0, "n_ids": 279, "positions": 279, "bucket": 512, "ids": [ 1, 17, 38685, 3667, 6590, 6881, 997, 42468, 522, 26260, 1532, 1878, 997, 15162, 6881, 997, 557, 1983, 834, 917, 38281, 1878, 997, 57267, 6881, 997, 1718, 531, 522, 6468, 522, 6135, 730, 1133, 535, 2019, 1878, 997, 26207, 6881, 1607, 38685, 38017, 6881, 997, 556, 2508, 522, 526, 552, 548, 1878, 997, 3055, 6881, 997, 46553, 791, 63437, 730, 526, 14895, 1878, 997, 590, 1937, 6881, 730, 527, 521, 997, 19248, 6881, 997, 2354, 9487, 511, 3205, 37080, 38017, 6881, 997, 549, 41555, 522, 14819, 1878, 997, 3055, 6881, 997, 41388, 32849, 17125, 730, 14819, 836, 1878, 997, 590, 1937, 6881, 730, 526, 521, 997, 19248, 6881, 997, 1066, 803, 9051, 511, 3205, 37080, 38017, 6881, 997, 12411, 542, 4894, 17727, 1878, 997, 3055, 6881, 997, 26818, 13192, 521, 730, 1838, 22173, 1878, 997, 590, 1937, 6881, 730, 528, 521, 997, 19248, 6881, 997, 2354, 9487, 511, 602, 5086, 997, 2354, 11468, 6881, 37080, 8618, 6265, 6881, 997, 2482, 1930, 2286, 1145, 1878, 997, 19132, 800, 36619, 6881, 997, 14402, 1991, 791, 1878, 997, 14402, 1991, 791, 40040, 6881, 997, 1718, 531, 522, 6468, 522, 3654, 730, 1632, 535, 2108, 511, 3205, 997, 27317, 957, 6881, 37080, 14164, 6881, 997, 19563, 1878, 997, 576, 2922, 3107, 6881, 730, 2142, 523, 529, 521, 997, 1891, 1356, 791, 6881, 730, 525, 523, 525, 3205, 997, 38765, 770, 5205, 2176, 6881, 997, 15090, 856, 1727, 17125, 540, 859, 953, 15418, 875, 936, 2927, 602, 18, 29589, 779, 13608, 2992, 1186, 10274, 875, 3503, 1186, 2172, 1014, 6984, 540, 19, 16, 4831, 535, 1295, 521, 779, 8294, 2177, 1014, 3965, 20, 19, 16, 3347, 535, 18672, 521, 779, 8294, 10995, 20, 21 ], "markers": [ 257, 269 ], "logits_raw": [ 2.150092363357544, -0.286288321018219 ], "probs_raw": [ 0.9195597767829895, 0.0804402306675911 ], "probs": [ 0.18814170360565186, 0.8118582963943481 ], "argmax_index": 1, "argmax": "false", "top2_margin": 0.6237165927886963, "near_tie": false, "gold": "true", "correct": false, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 1, "B": 3 } } ], "response": { "answers": { "next_action": { "type": "choice", "choice": "track_parcel", "confidence": 0.8459538817405701, "probabilities": { "refund_missing_item": 0.09913931041955948, "resend_order": 0.0021679105702787638, "track_parcel": 0.8459538817405701, "cancel_order": 0.0030852300114929676, "change_address": 4.503615491557866e-05, "discount_code": 0.0002456512884236872, "replace_damaged": 5.806538865726907e-06, "confirm_delivery": 0.009296542964875698, "fraud_review": 0.0033039606641978025, "no_action": 0.03675662726163864 } }, "partial": { "type": "noul", "noul": 0.9200953841209412 }, "charged_claim": { "type": "noul", "noul": 0.18814170360565186 } }, "usage": { "input_tokens": 931, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "own_m01", "mode": "text", "source": "own_text", "native": true, "prefix": 0, "usage_input_tokens": 1408, "questions": [ { "qid": "bedrooms", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 164, "positions": 164, "bucket": 256, "ids": [ 1, 17, 12832, 776, 811, 2944, 1141, 2642, 1050, 799, 40315, 9690, 521, 15398, 799, 12890, 708, 15780, 1985, 791, 5652, 9645, 884, 779, 4976, 8913, 803, 768, 4204, 916, 768, 12372, 523, 55657, 730, 526, 521, 5411, 1094, 5805, 521, 28553, 7449, 811, 4009, 6272, 516, 16429, 521, 9447, 3676, 5699, 730, 1374, 4493, 523, 20448, 988, 730, 526, 4526, 523, 879, 3113, 938, 1014, 4557, 6794, 1883, 523, 873, 1808, 4192, 968, 10449, 843, 15344, 9889, 810, 15622, 1550, 521, 1295, 1312, 5165, 1191, 941, 9645, 856, 9104, 1250, 2075, 10325, 988, 768, 56926, 810, 768, 24910, 7073, 523, 879, 1445, 8784, 797, 779, 39932, 536, 1295, 23171, 3494, 523, 13483, 1327, 884, 826, 4306, 24463, 1685, 730, 1155, 535, 1056, 810, 730, 1532, 535, 1056, 523, 22194, 779, 11380, 521, 873, 3085, 521, 811, 2643, 768, 25647, 523, 18, 4023, 1697, 6100, 14488, 2177, 779, 9645, 1052, 540, 19, 16, 1235, 20, 19, 16, 1547, 20, 19, 16, 2204, 20, 21 ], "markers": [ 152, 156, 160 ], "logits_raw": [ -2.658261299133301, -0.02049952745437622, -5.027771472930908 ], "probs_raw": [ 0.06633322685956955, 0.9274629354476929, 0.006203918252140284 ], "probs": [ 0.1287752240896225, 0.8475256562232971, 0.02369912900030613 ], "argmax_index": 1, "argmax": "two", "top2_margin": 0.7187504321336746, "near_tie": false, "gold": "two", "correct": true, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 7, "B": 8 } }, { "qid": "lift", "type": "noul", "K": 2, "calibrate": true, "temperature_key": "noul:2", "T": 1.6663223505020142, "prefix": 0, "n_ids": 170, "positions": 170, "bucket": 256, "ids": [ 1, 17, 12832, 776, 811, 2944, 1141, 2642, 1050, 799, 40315, 9690, 521, 15398, 799, 12890, 708, 15780, 1985, 791, 5652, 9645, 884, 779, 4976, 8913, 803, 768, 4204, 916, 768, 12372, 523, 55657, 730, 526, 521, 5411, 1094, 5805, 521, 28553, 7449, 811, 4009, 6272, 516, 16429, 521, 9447, 3676, 5699, 730, 1374, 4493, 523, 20448, 988, 730, 526, 4526, 523, 879, 3113, 938, 1014, 4557, 6794, 1883, 523, 873, 1808, 4192, 968, 10449, 843, 15344, 9889, 810, 15622, 1550, 521, 1295, 1312, 5165, 1191, 941, 9645, 856, 9104, 1250, 2075, 10325, 988, 768, 56926, 810, 768, 24910, 7073, 523, 879, 1445, 8784, 797, 779, 39932, 536, 1295, 23171, 3494, 523, 13483, 1327, 884, 826, 4306, 24463, 1685, 730, 1155, 535, 1056, 810, 730, 1532, 535, 1056, 523, 22194, 779, 11380, 521, 873, 3085, 521, 811, 2643, 768, 25647, 523, 18, 4940, 1413, 768, 12372, 540, 19, 16, 4831, 535, 1295, 521, 779, 8294, 2177, 1014, 3965, 20, 19, 16, 3347, 535, 18672, 521, 779, 8294, 10995, 20, 21 ], "markers": [ 148, 160 ], "logits_raw": [ 0.049930721521377563, 9.483137130737305 ], "probs_raw": [ 8.001578680705279e-05, 0.9999200105667114 ], "probs": [ 0.9965332746505737, 0.003466650377959013 ], "argmax_index": 0, "argmax": "true", "top2_margin": 0.9930666242726147, "near_tie": false, "gold": "true", "correct": true, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 6, "B": 8 } }, { "qid": "furnished", "type": "score", "K": 3, "calibrate": true, "temperature_key": "score:3-5", "T": 1.7301132678985596, "prefix": 0, "n_ids": 177, "positions": 177, "bucket": 256, "ids": [ 1, 17, 12832, 776, 811, 2944, 1141, 2642, 1050, 799, 40315, 9690, 521, 15398, 799, 12890, 708, 15780, 1985, 791, 5652, 9645, 884, 779, 4976, 8913, 803, 768, 4204, 916, 768, 12372, 523, 55657, 730, 526, 521, 5411, 1094, 5805, 521, 28553, 7449, 811, 4009, 6272, 516, 16429, 521, 9447, 3676, 5699, 730, 1374, 4493, 523, 20448, 988, 730, 526, 4526, 523, 879, 3113, 938, 1014, 4557, 6794, 1883, 523, 873, 1808, 4192, 968, 10449, 843, 15344, 9889, 810, 15622, 1550, 521, 1295, 1312, 5165, 1191, 941, 9645, 856, 9104, 1250, 2075, 10325, 988, 768, 56926, 810, 768, 24910, 7073, 523, 879, 1445, 8784, 797, 779, 39932, 536, 1295, 23171, 3494, 523, 13483, 1327, 884, 826, 4306, 24463, 1685, 730, 1155, 535, 1056, 810, 730, 1532, 535, 1056, 523, 22194, 779, 11380, 521, 873, 3085, 521, 811, 2643, 768, 25647, 523, 18, 4023, 60603, 856, 779, 9645, 540, 19, 16, 2119, 730, 525, 535, 9104, 1250, 2075, 20, 19, 16, 2119, 730, 526, 535, 20618, 60603, 20, 19, 16, 2119, 730, 527, 535, 7408, 60603, 20, 21 ], "markers": [ 149, 159, 168 ], "logits_raw": [ -0.9351052641868591, -1.015489101409912, -0.2348705232143402 ], "probs_raw": [ 0.25400131940841675, 0.23438279330730438, 0.5116159319877625 ], "probs": [ 0.2895609736442566, 0.27641525864601135, 0.43402379751205444 ], "argmax_index": 2, "argmax": "2", "top2_margin": 0.14446282386779785, "near_tie": false, "gold": "0", "correct": false, "host_softmax_max_abs_dp": 2.9802322387695312e-08, "batch": { "call": 0, "index": 2, "B": 8 } }, { "qid": "rabbit", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 179, "positions": 179, "bucket": 256, "ids": [ 1, 17, 12832, 776, 811, 2944, 1141, 2642, 1050, 799, 40315, 9690, 521, 15398, 799, 12890, 708, 15780, 1985, 791, 5652, 9645, 884, 779, 4976, 8913, 803, 768, 4204, 916, 768, 12372, 523, 55657, 730, 526, 521, 5411, 1094, 5805, 521, 28553, 7449, 811, 4009, 6272, 516, 16429, 521, 9447, 3676, 5699, 730, 1374, 4493, 523, 20448, 988, 730, 526, 4526, 523, 879, 3113, 938, 1014, 4557, 6794, 1883, 523, 873, 1808, 4192, 968, 10449, 843, 15344, 9889, 810, 15622, 1550, 521, 1295, 1312, 5165, 1191, 941, 9645, 856, 9104, 1250, 2075, 10325, 988, 768, 56926, 810, 768, 24910, 7073, 523, 879, 1445, 8784, 797, 779, 39932, 536, 1295, 23171, 3494, 523, 13483, 1327, 884, 826, 4306, 24463, 1685, 730, 1155, 535, 1056, 810, 730, 1532, 535, 1056, 523, 22194, 779, 11380, 521, 873, 3085, 521, 811, 2643, 768, 25647, 523, 18, 45313, 768, 44416, 3089, 768, 36802, 540, 19, 16, 18672, 535, 2595, 19226, 20, 19, 16, 968, 572, 890, 22748, 535, 11005, 916, 779, 2496, 46130, 1090, 10449, 20, 19, 16, 1295, 535, 4524, 6767, 20, 21 ], "markers": [ 150, 157, 172 ], "logits_raw": [ -1.355165719985962, 1.5525476932525635, 0.4870193600654602 ], "probs_raw": [ 0.03902411088347435, 0.7147216200828552, 0.24625429511070251 ], "probs": [ 0.07868150621652603, 0.627973198890686, 0.29334530234336853 ], "argmax_index": 1, "argmax": "by_agreement", "top2_margin": 0.3346278965473175, "near_tie": false, "gold": "no", "correct": false, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 1, "B": 8 } }, { "qid": "parking", "type": "noul", "K": 2, "calibrate": true, "temperature_key": "noul:2", "T": 1.6663223505020142, "prefix": 0, "n_ids": 174, "positions": 174, "bucket": 256, "ids": [ 1, 17, 12832, 776, 811, 2944, 1141, 2642, 1050, 799, 40315, 9690, 521, 15398, 799, 12890, 708, 15780, 1985, 791, 5652, 9645, 884, 779, 4976, 8913, 803, 768, 4204, 916, 768, 12372, 523, 55657, 730, 526, 521, 5411, 1094, 5805, 521, 28553, 7449, 811, 4009, 6272, 516, 16429, 521, 9447, 3676, 5699, 730, 1374, 4493, 523, 20448, 988, 730, 526, 4526, 523, 879, 3113, 938, 1014, 4557, 6794, 1883, 523, 873, 1808, 4192, 968, 10449, 843, 15344, 9889, 810, 15622, 1550, 521, 1295, 1312, 5165, 1191, 941, 9645, 856, 9104, 1250, 2075, 10325, 988, 768, 56926, 810, 768, 24910, 7073, 523, 879, 1445, 8784, 797, 779, 39932, 536, 1295, 23171, 3494, 523, 13483, 1327, 884, 826, 4306, 24463, 1685, 730, 1155, 535, 1056, 810, 730, 1532, 535, 1056, 523, 22194, 779, 11380, 521, 873, 3085, 521, 811, 2643, 768, 25647, 523, 18, 29589, 779, 9645, 2743, 916, 768, 23171, 3494, 540, 19, 16, 4831, 535, 1295, 521, 779, 8294, 2177, 1014, 3965, 20, 19, 16, 3347, 535, 18672, 521, 779, 8294, 10995, 20, 21 ], "markers": [ 152, 164 ], "logits_raw": [ 1.846391201019287, -2.831726312637329 ], "probs_raw": [ 0.9907891154289246, 0.009210866875946522 ], "probs": [ 0.056922901421785355, 0.9430771470069885 ], "argmax_index": 1, "argmax": "false", "top2_margin": 0.8861542455852032, "near_tie": false, "gold": "false", "correct": true, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 5, "B": 8 } }, { "qid": "bill", "type": "choice", "K": 5, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 176, "positions": 176, "bucket": 256, "ids": [ 1, 17, 12832, 776, 811, 2944, 1141, 2642, 1050, 799, 40315, 9690, 521, 15398, 799, 12890, 708, 15780, 1985, 791, 5652, 9645, 884, 779, 4976, 8913, 803, 768, 4204, 916, 768, 12372, 523, 55657, 730, 526, 521, 5411, 1094, 5805, 521, 28553, 7449, 811, 4009, 6272, 516, 16429, 521, 9447, 3676, 5699, 730, 1374, 4493, 523, 20448, 988, 730, 526, 4526, 523, 879, 3113, 938, 1014, 4557, 6794, 1883, 523, 873, 1808, 4192, 968, 10449, 843, 15344, 9889, 810, 15622, 1550, 521, 1295, 1312, 5165, 1191, 941, 9645, 856, 9104, 1250, 2075, 10325, 988, 768, 56926, 810, 768, 24910, 7073, 523, 879, 1445, 8784, 797, 779, 39932, 536, 1295, 23171, 3494, 523, 13483, 1327, 884, 826, 4306, 24463, 1685, 730, 1155, 535, 1056, 810, 730, 1532, 535, 1056, 523, 22194, 779, 11380, 521, 873, 3085, 521, 811, 2643, 768, 25647, 523, 18, 28828, 5363, 856, 4557, 797, 779, 16429, 540, 19, 16, 10366, 20, 19, 16, 4916, 20, 19, 16, 1883, 20, 19, 16, 9399, 20, 19, 16, 10735, 535, 3253, 5363, 856, 4557, 20, 21 ], "markers": [ 151, 155, 159, 163, 167 ], "logits_raw": [ -6.876915454864502, -14.46700668334961, -6.447972774505615, -9.973302841186523, 1.3173890113830566 ], "probs_raw": [ 0.0002760255301836878, 1.3951301980341668e-07, 0.00042387392022646964, 1.2479739780246746e-05, 0.9992875456809998 ], "probs": [ 0.0028496168088167906, 1.2591870472533628e-05, 0.0038713153917342424, 0.0003120219334959984, 0.9929545521736145 ], "argmax_index": 4, "argmax": "none", "top2_margin": 0.9890832367818803, "near_tie": false, "gold": "water", "correct": false, "host_softmax_max_abs_dp": 1.1920928955078125e-07, "batch": { "call": 0, "index": 3, "B": 8 } }, { "qid": "weekday_viewing", "type": "noul", "K": 2, "calibrate": true, "temperature_key": "noul:2", "T": 1.6663223505020142, "prefix": 0, "n_ids": 175, "positions": 175, "bucket": 256, "ids": [ 1, 17, 12832, 776, 811, 2944, 1141, 2642, 1050, 799, 40315, 9690, 521, 15398, 799, 12890, 708, 15780, 1985, 791, 5652, 9645, 884, 779, 4976, 8913, 803, 768, 4204, 916, 768, 12372, 523, 55657, 730, 526, 521, 5411, 1094, 5805, 521, 28553, 7449, 811, 4009, 6272, 516, 16429, 521, 9447, 3676, 5699, 730, 1374, 4493, 523, 20448, 988, 730, 526, 4526, 523, 879, 3113, 938, 1014, 4557, 6794, 1883, 523, 873, 1808, 4192, 968, 10449, 843, 15344, 9889, 810, 15622, 1550, 521, 1295, 1312, 5165, 1191, 941, 9645, 856, 9104, 1250, 2075, 10325, 988, 768, 56926, 810, 768, 24910, 7073, 523, 879, 1445, 8784, 797, 779, 39932, 536, 1295, 23171, 3494, 523, 13483, 1327, 884, 826, 4306, 24463, 1685, 730, 1155, 535, 1056, 810, 730, 1532, 535, 1056, 523, 22194, 779, 11380, 521, 873, 3085, 521, 811, 2643, 768, 25647, 523, 18, 9036, 779, 9645, 874, 16521, 884, 768, 3025, 2367, 540, 19, 16, 4831, 535, 1295, 521, 779, 8294, 2177, 1014, 3965, 20, 19, 16, 3347, 535, 18672, 521, 779, 8294, 10995, 20, 21 ], "markers": [ 153, 165 ], "logits_raw": [ 1.393991470336914, 2.843672752380371 ], "probs_raw": [ 0.1900506168603897, 0.8099493384361267 ], "probs": [ 0.7047432661056519, 0.29525667428970337 ], "argmax_index": 0, "argmax": "true", "top2_margin": 0.4094865918159485, "near_tie": false, "gold": "false", "correct": false, "host_softmax_max_abs_dp": 2.9802322387695312e-08, "batch": { "call": 0, "index": 4, "B": 8 } }, { "qid": "term", "type": "score", "K": 4, "calibrate": true, "temperature_key": "score:3-5", "T": 1.7301132678985596, "prefix": 0, "n_ids": 193, "positions": 193, "bucket": 256, "ids": [ 1, 17, 12832, 776, 811, 2944, 1141, 2642, 1050, 799, 40315, 9690, 521, 15398, 799, 12890, 708, 15780, 1985, 791, 5652, 9645, 884, 779, 4976, 8913, 803, 768, 4204, 916, 768, 12372, 523, 55657, 730, 526, 521, 5411, 1094, 5805, 521, 28553, 7449, 811, 4009, 6272, 516, 16429, 521, 9447, 3676, 5699, 730, 1374, 4493, 523, 20448, 988, 730, 526, 4526, 523, 879, 3113, 938, 1014, 4557, 6794, 1883, 523, 873, 1808, 4192, 968, 10449, 843, 15344, 9889, 810, 15622, 1550, 521, 1295, 1312, 5165, 1191, 941, 9645, 856, 9104, 1250, 2075, 10325, 988, 768, 56926, 810, 768, 24910, 7073, 523, 879, 1445, 8784, 797, 779, 39932, 536, 1295, 23171, 3494, 523, 13483, 1327, 884, 826, 4306, 24463, 1685, 730, 1155, 535, 1056, 810, 730, 1532, 535, 1056, 523, 22194, 779, 11380, 521, 873, 3085, 521, 811, 2643, 768, 25647, 523, 18, 4023, 1787, 856, 779, 9447, 3676, 5699, 540, 19, 16, 2119, 730, 525, 535, 1490, 730, 531, 4493, 20, 19, 16, 2119, 730, 526, 535, 730, 531, 4493, 20, 19, 16, 2119, 730, 527, 535, 730, 1374, 4493, 20, 19, 16, 2119, 730, 528, 535, 1391, 730, 1374, 4493, 20, 21 ], "markers": [ 151, 162, 172, 182 ], "logits_raw": [ -3.809436559677124, -12.830514907836914, 5.591951847076416, 0.9847365617752075 ], "probs_raw": [ 8.178635471267626e-05, 9.882718643439148e-09, 0.9900380373001099, 0.00988015253096819 ], "probs": [ 0.004064416512846947, 2.2106480173533782e-05, 0.9309849143028259, 0.06492854654788971 ], "argmax_index": 2, "argmax": "2", "top2_margin": 0.8660563677549362, "near_tie": false, "gold": "2", "correct": true, "host_softmax_max_abs_dp": 1.1920928955078125e-07, "batch": { "call": 0, "index": 0, "B": 8 } } ], "response": { "answers": { "bedrooms": { "type": "choice", "choice": "two", "confidence": 0.8475256562232971, "probabilities": { "one": 0.1287752240896225, "two": 0.8475256562232971, "three": 0.02369912900030613 } }, "lift": { "type": "noul", "noul": 0.9965332746505737 }, "furnished": { "type": "score", "score": 1.1444628536701202, "confidence": 0.43402379751205444, "probabilities": { "0": 0.2895609736442566, "1": 0.27641525864601135, "2": 0.43402379751205444 }, "legend": { "0": "unfurnished", "1": "partly furnished", "2": "fully furnished" } }, "rabbit": { "type": "choice", "choice": "by_agreement", "confidence": 0.627973198890686, "probabilities": { "yes": 0.07868150621652603, "by_agreement": 0.627973198890686, "no": 0.29334530234336853 } }, "parking": { "type": "noul", "noul": 0.056922901421785355 }, "bill": { "type": "choice", "choice": "none", "confidence": 0.9929545521736145, "probabilities": { "electricity": 0.0028496168088167906, "gas": 1.2591870472533628e-05, "water": 0.0038713153917342424, "internet": 0.0003120219334959984, "none": 0.9929545521736145 } }, "weekday_viewing": { "type": "noul", "noul": 0.7047432661056519 }, "term": { "type": "score", "score": 2.0567775747294945, "confidence": 0.9309849143028259, "probabilities": { "0": 0.004064416512846947, "1": 2.2106480173533782e-05, "2": 0.9309849143028259, "3": 0.06492854654788971 }, "legend": { "0": "under 6 months", "1": "6 months", "2": "12 months", "3": "over 12 months" } } }, "usage": { "input_tokens": 1408, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "own_t01", "mode": "text", "source": "own_text", "native": true, "prefix": 0, "usage_input_tokens": 660, "questions": [ { "qid": "team", "type": "choice", "K": 5, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 187, "positions": 187, "bucket": 256, "ids": [ 1, 17, 561, 32909, 2679, 31464, 526, 1141, 8414, 730, 1718, 531, 522, 6135, 522, 1536, 730, 4905, 535, 1374, 708, 6314, 535, 50642, 843, 38765, 770, 2908, 730, 1718, 528, 1015, 41518, 535, 61958, 791, 11126, 875, 4690, 509, 42120, 521, 859, 6217, 884, 779, 9855, 834, 917, 3016, 786, 984, 1512, 19771, 2984, 523, 4411, 5060, 8294, 4427, 1547, 16620, 803, 730, 2752, 523, 2475, 884, 4690, 730, 526, 521, 1203, 859, 1550, 5068, 1235, 6377, 523, 859, 4001, 45942, 1727, 7493, 810, 1186, 5689, 1123, 811, 3452, 1010, 1509, 523, 13831, 53467, 779, 49006, 7786, 523, 859, 6351, 1236, 1672, 811, 36298, 5514, 969, 875, 1547, 6272, 2158, 859, 6217, 35754, 521, 7929, 4690, 730, 898, 523, 20018, 521, 50642, 18, 28828, 3410, 1753, 9936, 1033, 30437, 1509, 540, 19, 16, 5363, 800, 535, 61958, 774, 521, 60827, 1864, 810, 53467, 592, 20, 19, 16, 5514, 969, 535, 16208, 774, 521, 2286, 3670, 810, 11746, 14721, 20, 19, 16, 3863, 535, 11309, 4086, 521, 2984, 3807, 810, 784, 20231, 20, 19, 16, 10017, 535, 5135, 933, 5647, 50022, 20, 19, 16, 1312, 20, 21 ], "markers": [ 133, 147, 161, 174, 183 ], "logits_raw": [ 5.459228515625, -0.8426796197891235, 1.1409907341003418, -3.7750244140625, -2.193328619003296 ], "probs_raw": [ 0.9845149517059326, 0.0018044234020635486, 0.013117029331624508, 9.612519352231175e-05, 0.00046747553278692067 ], "probs": [ 0.9412418603897095, 0.010438592173159122, 0.04305680841207504, 0.0012850860366597772, 0.0039776465855538845 ], "argmax_index": 0, "argmax": "billing", "top2_margin": 0.8981850519776344, "near_tie": false, "gold": "billing", "correct": true, "host_softmax_max_abs_dp": 3.725290298461914e-09, "batch": { "call": 0, "index": 0, "B": 4 } }, { "qid": "several_actions", "type": "noul", "K": 2, "calibrate": true, "temperature_key": "noul:2", "T": 1.6663223505020142, "prefix": 0, "n_ids": 157, "positions": 157, "bucket": 256, "ids": [ 1, 17, 561, 32909, 2679, 31464, 526, 1141, 8414, 730, 1718, 531, 522, 6135, 522, 1536, 730, 4905, 535, 1374, 708, 6314, 535, 50642, 843, 38765, 770, 2908, 730, 1718, 528, 1015, 41518, 535, 61958, 791, 11126, 875, 4690, 509, 42120, 521, 859, 6217, 884, 779, 9855, 834, 917, 3016, 786, 984, 1512, 19771, 2984, 523, 4411, 5060, 8294, 4427, 1547, 16620, 803, 730, 2752, 523, 2475, 884, 4690, 730, 526, 521, 1203, 859, 1550, 5068, 1235, 6377, 523, 859, 4001, 45942, 1727, 7493, 810, 1186, 5689, 1123, 811, 3452, 1010, 1509, 523, 13831, 53467, 779, 49006, 7786, 523, 859, 6351, 1236, 1672, 811, 36298, 5514, 969, 875, 1547, 6272, 2158, 859, 6217, 35754, 521, 7929, 4690, 730, 898, 523, 20018, 521, 50642, 18, 29589, 779, 30437, 3452, 875, 1209, 1399, 1235, 4272, 540, 19, 16, 4831, 535, 1295, 521, 779, 8294, 2177, 1014, 3965, 20, 19, 16, 3347, 535, 18672, 521, 779, 8294, 10995, 20, 21 ], "markers": [ 135, 147 ], "logits_raw": [ -1.5088059902191162, 2.8384313583374023 ], "probs_raw": [ 0.012777145951986313, 0.9872227907180786 ], "probs": [ 0.931430995464325, 0.06856900453567505 ], "argmax_index": 0, "argmax": "true", "top2_margin": 0.8628619909286499, "near_tie": false, "gold": "true", "correct": true, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 2, "B": 4 } }, { "qid": "upset", "type": "score", "K": 4, "calibrate": true, "temperature_key": "score:3-5", "T": 1.7301132678985596, "prefix": 0, "n_ids": 166, "positions": 166, "bucket": 256, "ids": [ 1, 17, 561, 32909, 2679, 31464, 526, 1141, 8414, 730, 1718, 531, 522, 6135, 522, 1536, 730, 4905, 535, 1374, 708, 6314, 535, 50642, 843, 38765, 770, 2908, 730, 1718, 528, 1015, 41518, 535, 61958, 791, 11126, 875, 4690, 509, 42120, 521, 859, 6217, 884, 779, 9855, 834, 917, 3016, 786, 984, 1512, 19771, 2984, 523, 4411, 5060, 8294, 4427, 1547, 16620, 803, 730, 2752, 523, 2475, 884, 4690, 730, 526, 521, 1203, 859, 1550, 5068, 1235, 6377, 523, 859, 4001, 45942, 1727, 7493, 810, 1186, 5689, 1123, 811, 3452, 1010, 1509, 523, 13831, 53467, 779, 49006, 7786, 523, 859, 6351, 1236, 1672, 811, 36298, 5514, 969, 875, 1547, 6272, 2158, 859, 6217, 35754, 521, 7929, 4690, 730, 898, 523, 20018, 521, 50642, 18, 4023, 30133, 856, 779, 13608, 540, 19, 16, 2119, 730, 525, 535, 19243, 20, 19, 16, 2119, 730, 526, 535, 13545, 876, 35740, 791, 20, 19, 16, 2119, 730, 527, 535, 44658, 20, 19, 16, 2119, 730, 528, 535, 23732, 20, 21 ], "markers": [ 131, 139, 150, 158 ], "logits_raw": [ -0.5534851551055908, -2.147355318069458, -0.1437389999628067, -2.5713460445404053 ], "probs_raw": [ 0.35180142521858215, 0.07146420329809189, 0.5299661159515381, 0.046768318861722946 ], "probs": [ 0.3359356224536896, 0.13370925188064575, 0.4257071018218994, 0.10464800149202347 ], "argmax_index": 2, "argmax": "2", "top2_margin": 0.08977147936820984, "near_tie": false, "gold": "1", "correct": false, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 1, "B": 4 } }, { "qid": "refund", "type": "noul", "K": 2, "calibrate": true, "temperature_key": "noul:2", "T": 1.6663223505020142, "prefix": 0, "n_ids": 150, "positions": 150, "bucket": 256, "ids": [ 1, 17, 561, 32909, 2679, 31464, 526, 1141, 8414, 730, 1718, 531, 522, 6135, 522, 1536, 730, 4905, 535, 1374, 708, 6314, 535, 50642, 843, 38765, 770, 2908, 730, 1718, 528, 1015, 41518, 535, 61958, 791, 11126, 875, 4690, 509, 42120, 521, 859, 6217, 884, 779, 9855, 834, 917, 3016, 786, 984, 1512, 19771, 2984, 523, 4411, 5060, 8294, 4427, 1547, 16620, 803, 730, 2752, 523, 2475, 884, 4690, 730, 526, 521, 1203, 859, 1550, 5068, 1235, 6377, 523, 859, 4001, 45942, 1727, 7493, 810, 1186, 5689, 1123, 811, 3452, 1010, 1509, 523, 13831, 53467, 779, 49006, 7786, 523, 859, 6351, 1236, 1672, 811, 36298, 5514, 969, 875, 1547, 6272, 2158, 859, 6217, 35754, 521, 7929, 4690, 730, 898, 523, 20018, 521, 50642, 18, 9906, 1356, 5940, 540, 19, 16, 4831, 535, 3253, 53467, 856, 21968, 20, 19, 16, 3347, 535, 941, 13608, 22922, 21090, 875, 4636, 1892, 20, 21 ], "markers": [ 129, 138 ], "logits_raw": [ -5.6721014976501465, 12.44875717163086 ], "probs_raw": [ 1.3496180173433459e-08, 1.0 ], "probs": [ 0.9999810457229614, 1.8929660654976033e-05 ], "argmax_index": 0, "argmax": "true", "top2_margin": 0.9999621160623064, "near_tie": false, "gold": "true", "correct": true, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 3, "B": 4 } } ], "response": { "answers": { "team": { "type": "choice", "choice": "billing", "confidence": 0.9412418603897095, "probabilities": { "billing": 0.9412418603897095, "deliveries": 0.010438592173159122, "account": 0.04305680841207504, "technical": 0.0012850860366597772, "other": 0.0039776465855538845 } }, "several_actions": { "type": "noul", "noul": 0.931430995464325 }, "upset": { "type": "score", "score": 1.299067460000515, "confidence": 0.4257071018218994, "probabilities": { "0": 0.3359356224536896, "1": 0.13370925188064575, "2": 0.4257071018218994, "3": 0.10464800149202347 }, "legend": { "0": "calm", "1": "mildly annoyed", "2": "frustrated", "3": "angry" } }, "refund": { "type": "noul", "noul": 0.9999810457229614 } }, "usage": { "input_tokens": 660, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "own_t04", "mode": "text", "source": "own_text", "native": true, "prefix": 0, "usage_input_tokens": 753, "questions": [ { "qid": "crash_owner", "type": "choice", "K": 4, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 191, "positions": 191, "bucket": 256, "ids": [ 1, 17, 7491, 1983, 1395, 57394, 1141, 9573, 3410, 2472, 6688, 521, 22153, 708, 41678, 535, 835, 14007, 521, 6623, 521, 1157, 1837, 860, 521, 911, 14851, 708, 542, 14007, 535, 10414, 779, 49550, 39198, 8660, 536, 13668, 884, 2289, 4202, 2076, 61590, 819, 30819, 535, 779, 21064, 884, 33990, 875, 18204, 730, 1374, 7039, 856, 2510, 2714, 536, 950, 1011, 30961, 936, 1550, 884, 1235, 1927, 9161, 523, 18578, 791, 3032, 779, 6625, 5756, 12116, 779, 2341, 9161, 884, 16012, 819, 2435, 1837, 860, 535, 6872, 8788, 875, 730, 529, 523, 527, 43360, 536, 3988, 5743, 988, 911, 14851, 819, 547, 14851, 535, 41669, 13793, 9167, 536, 1188, 8685, 1157, 1837, 860, 779, 19771, 5592, 5196, 5743, 4008, 819, 15888, 6015, 535, 6872, 730, 529, 523, 527, 14967, 988, 22045, 811, 17980, 1929, 803, 779, 33990, 21064, 523, 6623, 40917, 779, 21064, 536, 835, 14007, 1188, 9256, 916, 1889, 27968, 523, 18, 17408, 40917, 779, 33990, 21064, 540, 19, 16, 768, 14007, 535, 835, 14007, 20, 19, 16, 7572, 535, 6623, 20, 19, 16, 979, 47222, 535, 1157, 1837, 860, 20, 19, 16, 3783, 582, 535, 911, 14851, 20, 21 ], "markers": [ 160, 168, 174, 183 ], "logits_raw": [ -1.1562695503234863, -0.11443820595741272, -2.6990933418273926, -4.24807596206665 ], "probs_raw": [ 0.2442837953567505, 0.692398726940155, 0.05222216248512268, 0.011095328256487846 ], "probs": [ 0.2819421887397766, 0.593431293964386, 0.09365379065275192, 0.030972765758633614 ], "argmax_index": 1, "argmax": "ben", "top2_margin": 0.3114891052246094, "near_tie": false, "gold": "ben", "correct": true, "host_softmax_max_abs_dp": 7.450580596923828e-09, "batch": { "call": 0, "index": 1, "B": 4 } }, { "qid": "blocked", "type": "noul", "K": 2, "calibrate": true, "temperature_key": "noul:2", "T": 1.6663223505020142, "prefix": 0, "n_ids": 182, "positions": 182, "bucket": 256, "ids": [ 1, 17, 7491, 1983, 1395, 57394, 1141, 9573, 3410, 2472, 6688, 521, 22153, 708, 41678, 535, 835, 14007, 521, 6623, 521, 1157, 1837, 860, 521, 911, 14851, 708, 542, 14007, 535, 10414, 779, 49550, 39198, 8660, 536, 13668, 884, 2289, 4202, 2076, 61590, 819, 30819, 535, 779, 21064, 884, 33990, 875, 18204, 730, 1374, 7039, 856, 2510, 2714, 536, 950, 1011, 30961, 936, 1550, 884, 1235, 1927, 9161, 523, 18578, 791, 3032, 779, 6625, 5756, 12116, 779, 2341, 9161, 884, 16012, 819, 2435, 1837, 860, 535, 6872, 8788, 875, 730, 529, 523, 527, 43360, 536, 3988, 5743, 988, 911, 14851, 819, 547, 14851, 535, 41669, 13793, 9167, 536, 1188, 8685, 1157, 1837, 860, 779, 19771, 5592, 5196, 5743, 4008, 819, 15888, 6015, 535, 6872, 730, 529, 523, 527, 14967, 988, 22045, 811, 17980, 1929, 803, 779, 33990, 21064, 523, 6623, 40917, 779, 21064, 536, 835, 14007, 1188, 9256, 916, 1889, 27968, 523, 18, 4940, 1523, 3410, 5484, 26781, 540, 19, 16, 4831, 535, 1295, 521, 779, 8294, 2177, 1014, 3965, 20, 19, 16, 3347, 535, 18672, 521, 779, 8294, 10995, 20, 21 ], "markers": [ 160, 172 ], "logits_raw": [ 2.0561749935150146, -0.3539355993270874 ], "probs_raw": [ 0.9175950288772583, 0.0824049636721611 ], "probs": [ 0.19056159257888794, 0.8094383478164673 ], "argmax_index": 1, "argmax": "false", "top2_margin": 0.6188767552375793, "near_tie": false, "gold": "true", "correct": false, "host_softmax_max_abs_dp": 1.4901161193847656e-08, "batch": { "call": 0, "index": 3, "B": 4 } }, { "qid": "release_day", "type": "choice", "K": 4, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 192, "positions": 192, "bucket": 256, "ids": [ 1, 17, 7491, 1983, 1395, 57394, 1141, 9573, 3410, 2472, 6688, 521, 22153, 708, 41678, 535, 835, 14007, 521, 6623, 521, 1157, 1837, 860, 521, 911, 14851, 708, 542, 14007, 535, 10414, 779, 49550, 39198, 8660, 536, 13668, 884, 2289, 4202, 2076, 61590, 819, 30819, 535, 779, 21064, 884, 33990, 875, 18204, 730, 1374, 7039, 856, 2510, 2714, 536, 950, 1011, 30961, 936, 1550, 884, 1235, 1927, 9161, 523, 18578, 791, 3032, 779, 6625, 5756, 12116, 779, 2341, 9161, 884, 16012, 819, 2435, 1837, 860, 535, 6872, 8788, 875, 730, 529, 523, 527, 43360, 536, 3988, 5743, 988, 911, 14851, 819, 547, 14851, 535, 41669, 13793, 9167, 536, 1188, 8685, 1157, 1837, 860, 779, 19771, 5592, 5196, 5743, 4008, 819, 15888, 6015, 535, 6872, 730, 529, 523, 527, 14967, 988, 22045, 811, 17980, 1929, 803, 779, 33990, 21064, 523, 6623, 40917, 779, 21064, 536, 835, 14007, 1188, 9256, 916, 1889, 27968, 523, 18, 4453, 856, 6872, 730, 529, 523, 527, 2166, 13573, 540, 19, 16, 823, 18908, 535, 22045, 20, 19, 16, 2261, 9008, 535, 16012, 20, 19, 16, 10954, 892, 535, 17980, 20, 19, 16, 10654, 535, 4524, 8091, 20, 21 ], "markers": [ 164, 171, 178, 185 ], "logits_raw": [ -0.7694559693336487, -1.497985601425171, 4.049992561340332, -2.367877960205078 ], "probs_raw": [ 0.007962951436638832, 0.0038430634886026382, 0.9865837693214417, 0.0016102311201393604 ], "probs": [ 0.030133897438645363, 0.01790776662528515, 0.9423384070396423, 0.009619934484362602 ], "argmax_index": 2, "argmax": "monday", "top2_margin": 0.912204509600997, "near_tie": false, "gold": "monday", "correct": true, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 0, "B": 4 } }, { "qid": "blockers", "type": "score", "K": 3, "calibrate": true, "temperature_key": "score:3-5", "T": 1.7301132678985596, "prefix": 0, "n_ids": 188, "positions": 188, "bucket": 256, "ids": [ 1, 17, 7491, 1983, 1395, 57394, 1141, 9573, 3410, 2472, 6688, 521, 22153, 708, 41678, 535, 835, 14007, 521, 6623, 521, 1157, 1837, 860, 521, 911, 14851, 708, 542, 14007, 535, 10414, 779, 49550, 39198, 8660, 536, 13668, 884, 2289, 4202, 2076, 61590, 819, 30819, 535, 779, 21064, 884, 33990, 875, 18204, 730, 1374, 7039, 856, 2510, 2714, 536, 950, 1011, 30961, 936, 1550, 884, 1235, 1927, 9161, 523, 18578, 791, 3032, 779, 6625, 5756, 12116, 779, 2341, 9161, 884, 16012, 819, 2435, 1837, 860, 535, 6872, 8788, 875, 730, 529, 523, 527, 43360, 536, 3988, 5743, 988, 911, 14851, 819, 547, 14851, 535, 41669, 13793, 9167, 536, 1188, 8685, 1157, 1837, 860, 779, 19771, 5592, 5196, 5743, 4008, 819, 15888, 6015, 535, 6872, 730, 529, 523, 527, 14967, 988, 22045, 811, 17980, 1929, 803, 779, 33990, 21064, 523, 6623, 40917, 779, 21064, 536, 835, 14007, 1188, 9256, 916, 1889, 27968, 523, 18, 4023, 1697, 1519, 2312, 768, 4409, 770, 540, 19, 16, 2119, 730, 525, 535, 10735, 20, 19, 16, 2119, 730, 526, 535, 1235, 20, 19, 16, 2119, 730, 527, 535, 1547, 933, 1209, 20, 21 ], "markers": [ 162, 170, 178 ], "logits_raw": [ -2.957319974899292, -4.160699367523193, -2.986135482788086 ], "probs_raw": [ 0.44018465280532837, 0.13213378190994263, 0.42768150568008423 ], "probs": [ 0.40285491943359375, 0.20094417035579681, 0.39620086550712585 ], "argmax_index": 0, "argmax": "0", "top2_margin": 0.0066540539264678955, "near_tie": true, "gold": "1", "correct": false, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 2, "B": 4 } } ], "response": { "answers": { "crash_owner": { "type": "choice", "choice": "ben", "confidence": 0.593431293964386, "probabilities": { "aiko": 0.2819421887397766, "ben": 0.593431293964386, "chloe": 0.09365379065275192, "femi": 0.030972765758633614 } }, "blocked": { "type": "noul", "noul": 0.19056159257888794 }, "release_day": { "type": "choice", "choice": "monday", "confidence": 0.9423384070396423, "probabilities": { "thursday": 0.030133897438645363, "friday": 0.01790776662528515, "monday": 0.9423384070396423, "unknown": 0.009619934484362602 } }, "blockers": { "type": "score", "score": 0.9933459013700485, "confidence": 0.40285491943359375, "probabilities": { "0": 0.40285491943359375, "1": 0.20094417035579681, "2": 0.39620086550712585 }, "legend": { "0": "none", "1": "one", "2": "two or more" } } }, "usage": { "input_tokens": 753, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "own_t06", "mode": "text", "source": "own_text", "native": true, "prefix": 0, "usage_input_tokens": 727, "questions": [ { "qid": "first_request", "type": "choice", "K": 10, "calibrate": true, "temperature_key": "choice:6-10", "T": 1.1751071214675903, "prefix": 0, "n_ids": 294, "positions": 294, "bucket": 512, "ids": [ 1, 17, 568, 913, 535, 5843, 570, 54053, 843, 5652, 730, 25454, 3271, 15012, 521, 779, 16542, 797, 1727, 5362, 7404, 1901, 3822, 810, 936, 1090, 34517, 708, 568, 913, 535, 6135, 570, 13660, 26862, 843, 564, 959, 54898, 2176, 13901, 3271, 859, 6217, 29640, 811, 8498, 896, 510, 859, 6016, 8685, 5791, 988, 12003, 2084, 3599, 819, 568, 913, 535, 6326, 570, 54053, 535, 14165, 523, 1236, 1011, 859, 3089, 779, 5362, 3032, 730, 527, 5806, 27968, 540, 1727, 8842, 856, 963, 730, 528, 708, 568, 913, 535, 6468, 570, 13660, 26862, 535, 25957, 64210, 3032, 730, 1536, 535, 1056, 856, 3023, 875, 730, 898, 14076, 521, 6283, 859, 1614, 936, 1784, 568, 913, 535, 6468, 570, 54053, 535, 18672, 8694, 708, 568, 913, 535, 1698, 570, 54053, 535, 2510, 31731, 4010, 810, 936, 1090, 5629, 52125, 708, 568, 913, 535, 2117, 570, 13660, 26862, 535, 4998, 6650, 521, 1494, 59063, 856, 884, 2772, 1868, 523, 1405, 1188, 874, 916, 1010, 797, 730, 1155, 5112, 521, 810, 1023, 1011, 4876, 1010, 811, 5362, 730, 26740, 1288, 1010, 6863, 523, 18, 3493, 2172, 779, 21522, 3452, 875, 1509, 540, 19, 16, 5362, 52799, 535, 9267, 933, 20056, 811, 779, 5362, 20, 19, 16, 5019, 45625, 1066, 535, 41072, 779, 5362, 2919, 1399, 13376, 20, 19, 16, 25583, 1706, 535, 38332, 25583, 1706, 933, 56043, 20, 19, 16, 12003, 535, 39781, 797, 779, 5362, 856, 11694, 20, 19, 16, 5363, 800, 535, 835, 3952, 1337, 779, 5363, 20, 19, 16, 2643, 800, 572, 6386, 535, 15015, 779, 14721, 933, 5362, 2432, 20, 19, 16, 5807, 43840, 535, 39781, 953, 5807, 20, 19, 16, 10925, 535, 835, 10925, 38203, 20, 19, 16, 5600, 535, 18525, 916, 768, 47502, 933, 8842, 20, 19, 16, 1312, 20, 21 ], "markers": [ 190, 202, 215, 226, 237, 248, 262, 271, 279, 290 ], "logits_raw": [ -3.301970958709717, -0.8134485483169556, -5.719726085662842, 1.0345563888549805, -6.834534168243408, -2.957244634628296, -8.278870582580566, -11.795267105102539, -2.055696725845337, -1.509829044342041 ], "probs_raw": [ 0.009950069710612297, 0.11983334273099899, 0.000886764726601541, 0.7605994939804077, 0.00029083952540531754, 0.014045554213225842, 6.860979192424566e-05, 2.0381423837534385e-06, 0.03460000455379486, 0.05972325801849365 ], "probs": [ 0.01712716370820999, 0.14236073195934296, 0.0021884406451135874, 0.6860777735710144, 0.000847472867462784, 0.022966191172599792, 0.0002479304384905845, 1.2437837540346663e-05, 0.04946316033601761, 0.0787087231874466 ], "argmax_index": 3, "argmax": "maintenance", "top2_margin": 0.5437170416116714, "near_tie": false, "gold": "maintenance", "correct": true, "host_softmax_max_abs_dp": 7.450580596923828e-09, "batch": { "call": 0, "index": 0, "B": 3 } }, { "qid": "fixed", "type": "noul", "K": 2, "calibrate": true, "temperature_key": "noul:2", "T": 1.6663223505020142, "prefix": 0, "n_ids": 217, "positions": 217, "bucket": 256, "ids": [ 1, 17, 568, 913, 535, 5843, 570, 54053, 843, 5652, 730, 25454, 3271, 15012, 521, 779, 16542, 797, 1727, 5362, 7404, 1901, 3822, 810, 936, 1090, 34517, 708, 568, 913, 535, 6135, 570, 13660, 26862, 843, 564, 959, 54898, 2176, 13901, 3271, 859, 6217, 29640, 811, 8498, 896, 510, 859, 6016, 8685, 5791, 988, 12003, 2084, 3599, 819, 568, 913, 535, 6326, 570, 54053, 535, 14165, 523, 1236, 1011, 859, 3089, 779, 5362, 3032, 730, 527, 5806, 27968, 540, 1727, 8842, 856, 963, 730, 528, 708, 568, 913, 535, 6468, 570, 13660, 26862, 535, 25957, 64210, 3032, 730, 1536, 535, 1056, 856, 3023, 875, 730, 898, 14076, 521, 6283, 859, 1614, 936, 1784, 568, 913, 535, 6468, 570, 54053, 535, 18672, 8694, 708, 568, 913, 535, 1698, 570, 54053, 535, 2510, 31731, 4010, 810, 936, 1090, 5629, 52125, 708, 568, 913, 535, 2117, 570, 13660, 26862, 535, 4998, 6650, 521, 1494, 59063, 856, 884, 2772, 1868, 523, 1405, 1188, 874, 916, 1010, 797, 730, 1155, 5112, 521, 810, 1023, 1011, 4876, 1010, 811, 5362, 730, 26740, 1288, 1010, 6863, 523, 18, 24860, 779, 16542, 3220, 1381, 9167, 968, 779, 1764, 803, 779, 22537, 540, 19, 16, 4831, 535, 1295, 521, 779, 8294, 2177, 1014, 3965, 20, 19, 16, 3347, 535, 18672, 521, 779, 8294, 10995, 20, 21 ], "markers": [ 195, 207 ], "logits_raw": [ 2.7867722511291504, -0.9277684688568115 ], "probs_raw": [ 0.9762129783630371, 0.02378701977431774 ], "probs": [ 0.09716012328863144, 0.9028398990631104 ], "argmax_index": 1, "argmax": "false", "top2_margin": 0.8056797757744789, "near_tie": false, "gold": "false", "correct": true, "host_softmax_max_abs_dp": 7.450580596923828e-09, "batch": { "call": 0, "index": 1, "B": 3 } }, { "qid": "patience", "type": "score", "K": 3, "calibrate": true, "temperature_key": "score:3-5", "T": 1.7301132678985596, "prefix": 0, "n_ids": 216, "positions": 216, "bucket": 256, "ids": [ 1, 17, 568, 913, 535, 5843, 570, 54053, 843, 5652, 730, 25454, 3271, 15012, 521, 779, 16542, 797, 1727, 5362, 7404, 1901, 3822, 810, 936, 1090, 34517, 708, 568, 913, 535, 6135, 570, 13660, 26862, 843, 564, 959, 54898, 2176, 13901, 3271, 859, 6217, 29640, 811, 8498, 896, 510, 859, 6016, 8685, 5791, 988, 12003, 2084, 3599, 819, 568, 913, 535, 6326, 570, 54053, 535, 14165, 523, 1236, 1011, 859, 3089, 779, 5362, 3032, 730, 527, 5806, 27968, 540, 1727, 8842, 856, 963, 730, 528, 708, 568, 913, 535, 6468, 570, 13660, 26862, 535, 25957, 64210, 3032, 730, 1536, 535, 1056, 856, 3023, 875, 730, 898, 14076, 521, 6283, 859, 1614, 936, 1784, 568, 913, 535, 6468, 570, 54053, 535, 18672, 8694, 708, 568, 913, 535, 1698, 570, 54053, 535, 2510, 31731, 4010, 810, 936, 1090, 5629, 52125, 708, 568, 913, 535, 2117, 570, 13660, 26862, 535, 4998, 6650, 521, 1494, 59063, 856, 884, 2772, 1868, 523, 1405, 1188, 874, 916, 1010, 797, 730, 1155, 5112, 521, 810, 1023, 1011, 4876, 1010, 811, 5362, 730, 26740, 1288, 1010, 6863, 523, 18, 4023, 5087, 856, 779, 21522, 968, 779, 1764, 540, 19, 16, 2119, 730, 525, 535, 5087, 20, 19, 16, 2119, 730, 526, 535, 2269, 51908, 20, 19, 16, 2119, 730, 527, 535, 23732, 20, 21 ], "markers": [ 191, 199, 208 ], "logits_raw": [ -0.20327231287956238, -2.439265489578247, -3.508890390396118 ], "probs_raw": [ 0.8744603395462036, 0.09346750378608704, 0.03207217901945114 ], "probs": [ 0.7029393911361694, 0.19303540885448456, 0.10402525961399078 ], "argmax_index": 0, "argmax": "0", "top2_margin": 0.5099039822816849, "near_tie": false, "gold": "1", "correct": false, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 2, "B": 3 } } ], "response": { "answers": { "first_request": { "type": "choice", "choice": "maintenance", "confidence": 0.6860777735710144, "probabilities": { "room_service": 0.01712716370820999, "late_checkout": 0.14236073195934296, "towels": 0.0021884406451135874, "maintenance": 0.6860777735710144, "billing": 0.000847472867462784, "booking_change": 0.022966191172599792, "lost_item": 0.0002479304384905845, "noise": 1.2437837540346663e-05, "transport": 0.04946316033601761, "other": 0.0787087231874466 } }, "fixed": { "type": "noul", "noul": 0.09716012328863144 }, "patience": { "type": "score", "score": 0.4010859280824661, "confidence": 0.7029393911361694, "probabilities": { "0": 0.7029393911361694, "1": 0.19303540885448456, "2": 0.10402525961399078 }, "legend": { "0": "patient", "1": "impatient", "2": "angry" } } }, "usage": { "input_tokens": 727, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "own_t07", "mode": "text", "source": "own_text", "native": true, "prefix": 0, "usage_input_tokens": 199, "questions": [ { "qid": "slot", "type": "choice", "K": 4, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 199, "positions": 199, "bucket": 256, "ids": [ 1, 17, 556, 820, 799, 12890, 33496, 1141, 2692, 793, 18284, 19094, 521, 16012, 730, 1443, 535, 3167, 708, 13694, 5052, 521, 1033, 856, 54017, 523, 859, 1052, 768, 16084, 20328, 884, 22092, 963, 730, 534, 535, 1664, 916, 3583, 2412, 786, 523, 39781, 4010, 1287, 963, 1336, 521, 1101, 859, 1595, 811, 4876, 936, 523, 859, 1011, 1112, 22153, 1720, 730, 1386, 535, 1056, 933, 1523, 1394, 884, 22045, 8532, 523, 13831, 1868, 1123, 1892, 811, 10547, 523, 859, 1011, 1901, 3999, 12645, 2013, 779, 2285, 521, 1101, 1720, 730, 1133, 535, 1056, 856, 2238, 523, 28296, 1010, 12588, 22935, 42474, 3280, 3025, 535, 837, 1037, 730, 1632, 535, 1056, 521, 16666, 730, 1386, 535, 1664, 521, 16666, 730, 1417, 535, 1056, 521, 1097, 594, 730, 4528, 535, 1664, 521, 1097, 594, 730, 1374, 535, 1664, 523, 18, 28828, 803, 1518, 3262, 42474, 25113, 779, 59850, 540, 19, 16, 766, 1037, 572, 7984, 525, 535, 22092, 730, 1632, 535, 1056, 20, 19, 16, 28782, 572, 11610, 525, 535, 22153, 730, 1386, 535, 1664, 20, 19, 16, 823, 594, 572, 10293, 525, 535, 22045, 730, 1374, 535, 1664, 20, 19, 16, 10735, 535, 11357, 803, 1377, 25113, 20, 21 ], "markers": [ 149, 163, 176, 190 ], "logits_raw": [ 0.19124549627304077, -1.5787943601608276, 1.2088134288787842, -0.8737671971321106 ], "probs_raw": [ 0.23356255888938904, 0.039781827479600906, 0.6461410522460938, 0.08051449805498123 ], "probs": [ 0.2618938982486725, 0.07396066188812256, 0.5417618751525879, 0.12238364666700363 ], "argmax_index": 2, "argmax": "thu_1230", "top2_margin": 0.2798679769039154, "near_tie": false, "gold": "wed_1530", "correct": false, "host_softmax_max_abs_dp": 5.960464477539063e-08, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "slot": { "type": "choice", "choice": "thu_1230", "confidence": 0.5417618751525879, "probabilities": { "tue_1100": 0.2618938982486725, "wed_1530": 0.07396066188812256, "thu_1230": 0.5417618751525879, "none": 0.12238364666700363 } } }, "usage": { "input_tokens": 199, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "own_t10", "mode": "text", "source": "own_text", "native": true, "prefix": 0, "usage_input_tokens": 490, "questions": [ { "qid": "topic", "type": "choice", "K": 5, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 163, "positions": 163, "bucket": 256, "ids": [ 1, 17, 3097, 779, 17612, 11380, 521, 9855, 834, 917, 7920, 509, 550, 4238, 797, 9645, 730, 534, 810, 859, 1283, 4862, 521, 875, 779, 4976, 1394, 1033, 5805, 521, 1337, 779, 10925, 988, 779, 21780, 7494, 797, 779, 39932, 523, 941, 12719, 938, 1593, 3032, 1720, 44396, 521, 810, 779, 823, 51143, 4834, 1287, 1626, 779, 8913, 803, 1727, 32874, 523, 859, 1052, 5751, 875, 779, 10170, 4934, 811, 874, 6235, 811, 730, 4972, 535, 1056, 522, 2019, 535, 1056, 521, 906, 779, 3622, 7360, 2992, 523, 2182, 6300, 3807, 968, 779, 1764, 803, 3280, 3025, 521, 859, 1188, 9920, 936, 916, 779, 10698, 516, 15671, 523, 509, 566, 2670, 521, 2161, 887, 3309, 843, 40723, 730, 534, 518, 18, 3493, 856, 779, 7627, 1337, 540, 19, 16, 10925, 535, 63144, 20, 19, 16, 38183, 535, 4274, 5569, 20, 19, 16, 16429, 535, 55657, 933, 16620, 20, 19, 16, 23171, 535, 4514, 800, 20, 19, 16, 22906, 535, 873, 1808, 20, 21 ], "markers": [ 128, 134, 141, 149, 156 ], "logits_raw": [ 6.030937671661377, -7.381817817687988, -4.890927314758301, -10.31625747680664, -10.759885787963867 ], "probs_raw": [ 0.9999804496765137, 1.495910510129761e-06, 1.805866304493975e-05, 7.952342429007331e-08, 5.103049005583671e-08 ], "probs": [ 0.9995075464248657, 6.897762068547308e-05, 0.00040875584818422794, 8.47906540002441e-06, 6.176179795147618e-06 ], "argmax_index": 0, "argmax": "noise", "top2_margin": 0.9990987905766815, "near_tie": false, "gold": "noise", "correct": true, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 1, "B": 3 } }, { "qid": "first_time", "type": "noul", "K": 2, "calibrate": true, "temperature_key": "noul:2", "T": 1.6663223505020142, "prefix": 0, "n_ids": 154, "positions": 154, "bucket": 256, "ids": [ 1, 17, 3097, 779, 17612, 11380, 521, 9855, 834, 917, 7920, 509, 550, 4238, 797, 9645, 730, 534, 810, 859, 1283, 4862, 521, 875, 779, 4976, 1394, 1033, 5805, 521, 1337, 779, 10925, 988, 779, 21780, 7494, 797, 779, 39932, 523, 941, 12719, 938, 1593, 3032, 1720, 44396, 521, 810, 779, 823, 51143, 4834, 1287, 1626, 779, 8913, 803, 1727, 32874, 523, 859, 1052, 5751, 875, 779, 10170, 4934, 811, 874, 6235, 811, 730, 4972, 535, 1056, 522, 2019, 535, 1056, 521, 906, 779, 3622, 7360, 2992, 523, 2182, 6300, 3807, 968, 779, 1764, 803, 3280, 3025, 521, 859, 1188, 9920, 936, 916, 779, 10698, 516, 15671, 523, 509, 566, 2670, 521, 2161, 887, 3309, 843, 40723, 730, 534, 518, 18, 4940, 1033, 779, 1509, 1394, 779, 9978, 6537, 2190, 540, 19, 16, 4831, 535, 1295, 521, 779, 8294, 2177, 1014, 3965, 20, 19, 16, 3347, 535, 18672, 521, 779, 8294, 10995, 20, 21 ], "markers": [ 132, 144 ], "logits_raw": [ -0.21387678384780884, 4.600820541381836 ], "probs_raw": [ 0.008044439367949963, 0.991955578327179 ], "probs": [ 0.9473206996917725, 0.05267929285764694 ], "argmax_index": 0, "argmax": "true", "top2_margin": 0.8946414068341255, "near_tie": false, "gold": "false", "correct": false, "host_softmax_max_abs_dp": 3.725290298461914e-09, "batch": { "call": 0, "index": 2, "B": 3 } }, { "qid": "escalation", "type": "score", "K": 4, "calibrate": true, "temperature_key": "score:3-5", "T": 1.7301132678985596, "prefix": 0, "n_ids": 173, "positions": 173, "bucket": 256, "ids": [ 1, 17, 3097, 779, 17612, 11380, 521, 9855, 834, 917, 7920, 509, 550, 4238, 797, 9645, 730, 534, 810, 859, 1283, 4862, 521, 875, 779, 4976, 1394, 1033, 5805, 521, 1337, 779, 10925, 988, 779, 21780, 7494, 797, 779, 39932, 523, 941, 12719, 938, 1593, 3032, 1720, 44396, 521, 810, 779, 823, 51143, 4834, 1287, 1626, 779, 8913, 803, 1727, 32874, 523, 859, 1052, 5751, 875, 779, 10170, 4934, 811, 874, 6235, 811, 730, 4972, 535, 1056, 522, 2019, 535, 1056, 521, 906, 779, 3622, 7360, 2992, 523, 2182, 6300, 3807, 968, 779, 1764, 803, 3280, 3025, 521, 859, 1188, 9920, 936, 916, 779, 10698, 516, 15671, 523, 509, 566, 2670, 521, 2161, 887, 3309, 843, 40723, 730, 534, 518, 18, 4023, 3414, 856, 779, 9978, 9217, 811, 2337, 1033, 540, 19, 16, 2119, 730, 525, 535, 1295, 3759, 3801, 20, 19, 16, 2119, 730, 526, 535, 768, 36685, 20, 19, 16, 2119, 730, 527, 535, 779, 10698, 516, 15671, 20, 19, 16, 2119, 730, 528, 535, 768, 23749, 933, 7326, 20, 21 ], "markers": [ 132, 142, 151, 162 ], "logits_raw": [ -2.038007974624634, -1.6411914825439453, 4.013549327850342, -8.809741973876953 ], "probs_raw": [ 0.0023404823150485754, 0.003480492625385523, 0.9941763877868652, 2.681505065993406e-06 ], "probs": [ 0.02831273339688778, 0.035611558705568314, 0.9355105757713318, 0.000565113325137645 ], "argmax_index": 2, "argmax": "2", "top2_margin": 0.8998990170657635, "near_tie": false, "gold": "2", "correct": true, "host_softmax_max_abs_dp": 5.960464477539063e-08, "batch": { "call": 0, "index": 0, "B": 3 } } ], "response": { "answers": { "topic": { "type": "choice", "choice": "noise", "confidence": 0.9995075464248657, "probabilities": { "noise": 0.9995075464248657, "repairs": 6.897762068547308e-05, "rent": 0.00040875584818422794, "parking": 8.47906540002441e-06, "pets": 6.176179795147618e-06 } }, "first_time": { "type": "noul", "noul": 0.9473206996917725 }, "escalation": { "type": "score", "score": 1.9083280502236448, "confidence": 0.9355105757713318, "probabilities": { "0": 0.02831273339688778, "1": 0.035611558705568314, "2": 0.9355105757713318, "3": 0.000565113325137645 }, "legend": { "0": "no further step", "1": "a reminder", "2": "the residents' committee", "3": "a lawyer or court" } } }, "usage": { "input_tokens": 490, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_005dcf6d2c0f21d64796", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 98, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 98, "positions": 98, "bucket": 256, "ids": [ 1, 17, 1098, 2681, 28117, 1509, 6767, 768, 8153, 777, 811, 7157, 779, 1530, 793, 523, 13629, 779, 5089, 1689, 22922, 46159, 896, 15001, 523, 3253, 2919, 6177, 9647, 523, 18, 14875, 779, 2681, 28117, 1090, 10500, 15001, 6177, 536, 768, 28687, 32491, 1706, 779, 7045, 8825, 523, 3354, 779, 8153, 777, 7157, 779, 1530, 793, 540, 19, 16, 20458, 535, 941, 9519, 7330, 22558, 779, 4272, 20, 19, 16, 33776, 535, 941, 9519, 7330, 14480, 1512, 779, 4272, 20, 19, 16, 28441, 535, 941, 21706, 2203, 2177, 1014, 27169, 3984, 779, 7330, 22558, 779, 4272, 20, 21 ], "markers": [ 58, 69, 81 ], "logits_raw": [ 2.0582985877990723, 1.225986361503601, 2.4099020957946777 ], "probs_raw": [ 0.3500926196575165, 0.15230506658554077, 0.49760234355926514 ], "probs": [ 0.35244399309158325, 0.19448202848434448, 0.4530739486217499 ], "argmax_index": 2, "argmax": "insufficient", "top2_margin": 0.10062995553016663, "near_tie": false, "gold": "prohibited", "correct": false, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "insufficient", "confidence": 0.4530739486217499, "probabilities": { "permitted": 0.35244399309158325, "prohibited": 0.19448202848434448, "insufficient": 0.4530739486217499 } } }, "usage": { "input_tokens": 98, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_00bb4cc043f7d2b288d4", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 75, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 75, "positions": 75, "bucket": 256, "ids": [ 1, 17, 1098, 12153, 4943, 876, 22558, 9365, 779, 11256, 8355, 523, 18, 18693, 902, 3863, 2853, 1550, 1403, 779, 12153, 3575, 6229, 9365, 896, 2853, 523, 3354, 779, 11256, 8355, 874, 7093, 540, 19, 16, 28441, 535, 941, 21706, 2203, 2177, 1014, 27169, 3984, 779, 7330, 22558, 779, 4272, 20, 19, 16, 33776, 535, 941, 9519, 7330, 14480, 1512, 779, 4272, 20, 19, 16, 20458, 535, 941, 9519, 7330, 22558, 779, 4272, 20, 21 ], "markers": [ 35, 52, 64 ], "logits_raw": [ 2.306075096130371, 0.6486483812332153, 4.865217208862305 ], "probs_raw": [ 0.07084484398365021, 0.013505072332918644, 0.9156500697135925 ], "probs": [ 0.1328362673521042, 0.04065634310245514, 0.8265073895454407 ], "argmax_index": 2, "argmax": "permitted", "top2_margin": 0.6936711221933365, "near_tie": false, "gold": "permitted", "correct": true, "host_softmax_max_abs_dp": 1.4901161193847656e-08, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "permitted", "confidence": 0.8265073895454407, "probabilities": { "insufficient": 0.1328362673521042, "prohibited": 0.04065634310245514, "permitted": 0.8265073895454407 } } }, "usage": { "input_tokens": 75, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_02f2b95a5c5a4777a45c", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 82, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 82, "positions": 82, "bucket": 256, "ids": [ 1, 17, 55945, 931, 835, 856, 779, 16495, 57490, 810, 856, 11649, 3222, 811, 9936, 1033, 3869, 523, 46050, 931, 879, 856, 779, 3981, 3321, 57490, 810, 856, 1236, 3222, 811, 9936, 936, 523, 18, 52739, 835, 19148, 835, 856, 3222, 536, 7056, 879, 1550, 1403, 835, 856, 12359, 810, 879, 856, 3222, 523, 19, 16, 879, 535, 46050, 931, 879, 20, 19, 16, 835, 535, 46050, 931, 835, 20, 19, 16, 28441, 535, 42749, 13719, 14350, 779, 21968, 4589, 20, 21 ], "markers": [ 55, 63, 71 ], "logits_raw": [ 2.056809425354004, 1.9211969375610352, 0.8836579918861389 ], "probs_raw": [ 0.4581752121448517, 0.40006986260414124, 0.14175499975681305 ], "probs": [ 0.4273075461387634, 0.38785475492477417, 0.18483762443065643 ], "argmax_index": 0, "argmax": "B", "top2_margin": 0.03945279121398926, "near_tie": false, "gold": "A", "correct": false, "host_softmax_max_abs_dp": 2.9802322387695312e-08, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "B", "confidence": 0.4273075461387634, "probabilities": { "B": 0.4273075461387634, "A": 0.38785475492477417, "insufficient": 0.18483762443065643 } } }, "usage": { "input_tokens": 82, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_03d919fd0e9b3a7cd1e6", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 60, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 60, "positions": 60, "bucket": 256, "ids": [ 1, 17, 1098, 40348, 35283, 768, 5562, 23819, 1203, 6602, 1295, 35101, 5562, 523, 18, 9886, 894, 779, 4673, 535, 779, 35101, 5562, 856, 797, 5105, 523, 19, 16, 28441, 535, 941, 4589, 2177, 1014, 6122, 4021, 20, 19, 16, 7951, 535, 941, 4589, 48308, 779, 4673, 20, 19, 16, 18197, 13735, 535, 941, 4589, 48308, 779, 11977, 20, 21 ], "markers": [ 28, 39, 49 ], "logits_raw": [ 2.0700466632843018, -3.3887336254119873, -0.8867999315261841 ], "probs_raw": [ 0.9467533826828003, 0.004031983204185963, 0.0492146797478199 ], "probs": [ 0.8762499690055847, 0.01774740032851696, 0.10600268840789795 ], "argmax_index": 0, "argmax": "insufficient", "top2_margin": 0.7702472805976868, "near_tie": false, "gold": "insufficient", "correct": true, "host_softmax_max_abs_dp": 1.862645149230957e-09, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "insufficient", "confidence": 0.8762499690055847, "probabilities": { "insufficient": 0.8762499690055847, "supported": 0.01774740032851696, "contradicted": 0.10600268840789795 } } }, "usage": { "input_tokens": 60, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_07ee70cfe4b01330ce89", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 60, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 60, "positions": 60, "bucket": 256, "ids": [ 1, 17, 1098, 3059, 14350, 768, 61026, 32634, 1203, 1295, 9100, 933, 12552, 2203, 523, 18, 9886, 894, 779, 4673, 535, 779, 61026, 1178, 1381, 45943, 523, 19, 16, 7951, 535, 941, 4589, 48308, 779, 4673, 20, 19, 16, 18197, 13735, 535, 941, 4589, 48308, 779, 11977, 20, 19, 16, 28441, 535, 941, 4589, 2177, 1014, 6122, 4021, 20, 21 ], "markers": [ 28, 38, 49 ], "logits_raw": [ -10.376697540283203, -5.948400974273682, 1.1245343685150146 ], "probs_raw": [ 1.0108947208209429e-05, 0.0008470146567560732, 0.9991428852081299 ], "probs": [ 0.0002685684012249112, 0.006351233460009098, 0.9933801889419556 ], "argmax_index": 2, "argmax": "insufficient", "top2_margin": 0.9870289554819465, "near_tie": false, "gold": "insufficient", "correct": true, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "insufficient", "confidence": 0.9933801889419556, "probabilities": { "supported": 0.0002685684012249112, "contradicted": 0.006351233460009098, "insufficient": 0.9933801889419556 } } }, "usage": { "input_tokens": 60, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_0aa1447680e4f25862bb", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 97, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 97, "positions": 97, "bucket": 256, "ids": [ 1, 17, 1098, 26862, 53076, 7944, 779, 7494, 28553, 906, 10274, 523, 941, 10423, 4462, 1856, 7944, 936, 906, 946, 38008, 523, 9758, 7944, 938, 2835, 523, 18, 35516, 7494, 8544, 1403, 779, 28553, 856, 8112, 10274, 536, 797, 19980, 521, 779, 10423, 4462, 1856, 1391, 58340, 779, 26862, 53076, 523, 3354, 936, 874, 5789, 540, 19, 16, 28441, 535, 941, 21706, 2203, 2177, 1014, 27169, 3984, 779, 7330, 22558, 779, 4272, 20, 19, 16, 33776, 535, 941, 9519, 7330, 14480, 1512, 779, 4272, 20, 19, 16, 20458, 535, 941, 9519, 7330, 22558, 779, 4272, 20, 21 ], "markers": [ 57, 74, 86 ], "logits_raw": [ 2.831890106201172, 1.8205256462097168, 2.5837559700012207 ], "probs_raw": [ 0.46642276644706726, 0.16964839398860931, 0.36392882466316223 ], "probs": [ 0.4304542541503906, 0.20901064574718475, 0.3605351746082306 ], "argmax_index": 0, "argmax": "insufficient", "top2_margin": 0.06991907954216003, "near_tie": false, "gold": "prohibited", "correct": false, "host_softmax_max_abs_dp": 2.9802322387695312e-08, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "insufficient", "confidence": 0.4304542541503906, "probabilities": { "insufficient": 0.4304542541503906, "prohibited": 0.20901064574718475, "permitted": 0.3605351746082306 } } }, "usage": { "input_tokens": 97, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_0aa24a6701ac9c56c4a0", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 84, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 84, "positions": 84, "bucket": 256, "ids": [ 1, 17, 55945, 931, 835, 535, 941, 21009, 30403, 32340, 521, 1757, 22922, 46159, 779, 52847, 2076, 779, 3308, 523, 46050, 931, 879, 535, 941, 21009, 30403, 32340, 810, 779, 5156, 3084, 44987, 936, 6728, 797, 5323, 523, 18, 15534, 779, 13719, 916, 6654, 6612, 32340, 15001, 536, 768, 28687, 32491, 1706, 902, 7045, 8825, 523, 19, 16, 28441, 535, 42749, 13719, 14350, 779, 21968, 4589, 20, 19, 16, 879, 535, 46050, 931, 879, 20, 19, 16, 835, 535, 46050, 931, 835, 20, 21 ], "markers": [ 57, 68, 76 ], "logits_raw": [ 2.172837495803833, 2.860612392425537, 1.9425041675567627 ], "probs_raw": [ 0.2643018066883087, 0.52577143907547, 0.20992669463157654 ], "probs": [ 0.28713083267211914, 0.46929967403411865, 0.24356946349143982 ], "argmax_index": 1, "argmax": "B", "top2_margin": 0.1821688413619995, "near_tie": false, "gold": "B", "correct": true, "host_softmax_max_abs_dp": 1.4901161193847656e-08, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "B", "confidence": 0.46929967403411865, "probabilities": { "insufficient": 0.28713083267211914, "B": 0.46929967403411865, "A": 0.24356946349143982 } } }, "usage": { "input_tokens": 84, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_0b43ea8e24e74d621f1b", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 62, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 62, "positions": 62, "bucket": 256, "ids": [ 1, 17, 1098, 23120, 44987, 936, 25539, 779, 16400, 31286, 811, 779, 13608, 1090, 28623, 523, 18, 9886, 894, 779, 4673, 535, 779, 16400, 31286, 1052, 1381, 25539, 523, 19, 16, 28441, 535, 941, 4589, 2177, 1014, 6122, 4021, 20, 19, 16, 7951, 535, 941, 4589, 48308, 779, 4673, 20, 19, 16, 18197, 13735, 535, 941, 4589, 48308, 779, 11977, 20, 21 ], "markers": [ 30, 41, 51 ], "logits_raw": [ -1.1856169700622559, 6.04213809967041, -3.6982507705688477 ], "probs_raw": [ 0.0007255796226672828, 0.9992156028747559, 5.881147808395326e-05 ], "probs": [ 0.005686212796717882, 0.9933690428733826, 0.0009447598131373525 ], "argmax_index": 1, "argmax": "supported", "top2_margin": 0.9876828300766647, "near_tie": false, "gold": "supported", "correct": true, "host_softmax_max_abs_dp": 4.656612873077393e-10, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "supported", "confidence": 0.9933690428733826, "probabilities": { "insufficient": 0.005686212796717882, "supported": 0.9933690428733826, "contradicted": 0.0009447598131373525 } } }, "usage": { "input_tokens": 62, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_0ebd299739302d8b057b", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 88, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 88, "positions": 88, "bucket": 256, "ids": [ 1, 17, 1098, 26862, 3446, 3498, 779, 2391, 931, 1178, 3191, 8784, 536, 779, 51746, 30174, 3498, 936, 856, 797, 8784, 523, 9758, 7944, 938, 2835, 521, 810, 779, 30174, 12199, 19980, 1490, 779, 13061, 523, 18, 9886, 894, 779, 4673, 535, 1490, 779, 21706, 13061, 521, 779, 2391, 931, 856, 6654, 797, 8784, 523, 19, 16, 18197, 13735, 535, 941, 4589, 48308, 779, 11977, 20, 19, 16, 28441, 535, 941, 4589, 2177, 1014, 6122, 4021, 20, 19, 16, 7951, 535, 941, 4589, 48308, 779, 4673, 20, 21 ], "markers": [ 56, 67, 78 ], "logits_raw": [ 2.0599265098571777, 0.8353641033172607, 0.7449114322662354 ], "probs_raw": [ 0.6400587558746338, 0.18810448050498962, 0.1718367636203766 ], "probs": [ 0.5531457662582397, 0.23064285516738892, 0.2162114381790161 ], "argmax_index": 0, "argmax": "contradicted", "top2_margin": 0.32250291109085083, "near_tie": false, "gold": "supported", "correct": false, "host_softmax_max_abs_dp": 1.4901161193847656e-08, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "contradicted", "confidence": 0.5531457662582397, "probabilities": { "contradicted": 0.5531457662582397, "insufficient": 0.23064285516738892, "supported": 0.2162114381790161 } } }, "usage": { "input_tokens": 88, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_1105577da4c8dad4609d", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 92, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 92, "positions": 92, "bucket": 256, "ids": [ 1, 17, 1098, 26862, 3446, 3498, 779, 2391, 931, 856, 797, 8784, 523, 941, 51746, 30174, 3498, 936, 1178, 3191, 8784, 523, 9758, 7944, 938, 2835, 523, 941, 21706, 13061, 3498, 779, 51746, 30174, 12199, 6249, 9448, 19148, 1518, 1547, 5617, 9171, 523, 18, 9886, 894, 779, 4673, 535, 779, 26862, 3446, 7944, 779, 2391, 931, 797, 8784, 523, 19, 16, 18197, 13735, 535, 941, 4589, 48308, 779, 11977, 20, 19, 16, 28441, 535, 941, 4589, 2177, 1014, 6122, 4021, 20, 19, 16, 7951, 535, 941, 4589, 48308, 779, 4673, 20, 21 ], "markers": [ 60, 71, 82 ], "logits_raw": [ 1.914824366569519, 1.321368932723999, -0.17958185076713562 ], "probs_raw": [ 0.5968159437179565, 0.329690158367157, 0.07349391281604767 ], "probs": [ 0.5323480367660522, 0.34840649366378784, 0.11924546957015991 ], "argmax_index": 0, "argmax": "contradicted", "top2_margin": 0.1839415431022644, "near_tie": false, "gold": "supported", "correct": false, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "contradicted", "confidence": 0.5323480367660522, "probabilities": { "contradicted": 0.5323480367660522, "insufficient": 0.34840649366378784, "supported": 0.11924546957015991 } } }, "usage": { "input_tokens": 92, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_19d3d8d0bf773e88be7b", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 64, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 64, "positions": 64, "bucket": 256, "ids": [ 1, 17, 55945, 931, 835, 535, 941, 8825, 953, 14708, 523, 46050, 931, 879, 535, 941, 5771, 953, 5068, 536, 779, 9170, 6728, 1044, 2017, 3019, 523, 18, 52739, 779, 13719, 15603, 902, 14708, 8825, 523, 19, 16, 835, 535, 46050, 931, 835, 20, 19, 16, 879, 535, 46050, 931, 879, 20, 19, 16, 28441, 535, 42749, 13719, 14350, 779, 21968, 4589, 20, 21 ], "markers": [ 37, 45, 53 ], "logits_raw": [ 0.8596923351287842, -1.9263551235198975, -0.09066939353942871 ], "probs_raw": [ 0.6904811263084412, 0.04257814958691597, 0.26694080233573914 ], "probs": [ 0.608325719833374, 0.08314049988985062, 0.3085337281227112 ], "argmax_index": 0, "argmax": "A", "top2_margin": 0.29979199171066284, "near_tie": false, "gold": "A", "correct": true, "host_softmax_max_abs_dp": 2.9802322387695312e-08, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "A", "confidence": 0.608325719833374, "probabilities": { "A": 0.608325719833374, "B": 0.08314049988985062, "insufficient": 0.3085337281227112 } } }, "usage": { "input_tokens": 64, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_19f37d5a97034bd7d04f", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 88, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 88, "positions": 88, "bucket": 256, "ids": [ 1, 17, 1098, 33648, 1090, 37181, 6175, 13546, 779, 30122, 9041, 810, 4943, 876, 13748, 774, 779, 1232, 33388, 920, 2543, 10074, 523, 18, 542, 37181, 6175, 22558, 9544, 8544, 811, 779, 3508, 936, 13546, 523, 3354, 779, 33648, 7157, 779, 1232, 33388, 920, 2543, 10074, 540, 19, 16, 28441, 535, 941, 21706, 2203, 2177, 1014, 27169, 3984, 779, 7330, 22558, 779, 4272, 20, 19, 16, 33776, 535, 941, 9519, 7330, 14480, 1512, 779, 4272, 20, 19, 16, 20458, 535, 941, 9519, 7330, 22558, 779, 4272, 20, 21 ], "markers": [ 48, 65, 77 ], "logits_raw": [ 2.548405170440674, 3.762322187423706, -0.5323378443717957 ], "probs_raw": [ 0.2266252338886261, 0.7629669308662415, 0.010407790541648865 ], "probs": [ 0.28646430373191833, 0.6818165183067322, 0.031719230115413666 ], "argmax_index": 1, "argmax": "prohibited", "top2_margin": 0.39535221457481384, "near_tie": false, "gold": "prohibited", "correct": true, "host_softmax_max_abs_dp": 5.960464477539063e-08, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "prohibited", "confidence": 0.6818165183067322, "probabilities": { "insufficient": 0.28646430373191833, "prohibited": 0.6818165183067322, "permitted": 0.031719230115413666 } } }, "usage": { "input_tokens": 88, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_1a81ead69e93e5d0c86a", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 81, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 81, "positions": 81, "bucket": 256, "ids": [ 1, 17, 1098, 12153, 21968, 902, 3863, 10711, 521, 1203, 779, 7387, 811, 874, 7093, 938, 49033, 523, 18, 18693, 902, 3863, 2853, 1550, 1403, 779, 12153, 3575, 6229, 9365, 896, 2853, 523, 3354, 779, 11256, 8355, 874, 7093, 540, 19, 16, 20458, 535, 941, 9519, 7330, 22558, 779, 4272, 20, 19, 16, 33776, 535, 941, 9519, 7330, 14480, 1512, 779, 4272, 20, 19, 16, 28441, 535, 941, 21706, 2203, 2177, 1014, 27169, 3984, 779, 7330, 22558, 779, 4272, 20, 21 ], "markers": [ 41, 52, 64 ], "logits_raw": [ -5.296720027923584, -2.989729642868042, 2.016939878463745 ], "probs_raw": [ 0.0006615055026486516, 0.006644258741289377, 0.9926942586898804 ], "probs": [ 0.005209682509303093, 0.027071693912148476, 0.9677186608314514 ], "argmax_index": 2, "argmax": "insufficient", "top2_margin": 0.9406469669193029, "near_tie": false, "gold": "insufficient", "correct": true, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "insufficient", "confidence": 0.9677186608314514, "probabilities": { "permitted": 0.005209682509303093, "prohibited": 0.027071693912148476, "insufficient": 0.9677186608314514 } } }, "usage": { "input_tokens": 81, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_1df851a202994a07f123", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 79, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 79, "positions": 79, "bucket": 256, "ids": [ 1, 17, 1098, 11860, 2090, 1321, 6835, 2059, 779, 3956, 8649, 810, 779, 5196, 8170, 523, 941, 11746, 3084, 44987, 779, 3956, 8649, 11423, 810, 22922, 3498, 779, 8170, 953, 1014, 14465, 523, 18, 9886, 894, 779, 4673, 535, 779, 11860, 2090, 1321, 856, 5156, 523, 19, 16, 18197, 13735, 535, 941, 4589, 48308, 779, 11977, 20, 19, 16, 28441, 535, 941, 4589, 2177, 1014, 6122, 4021, 20, 19, 16, 7951, 535, 941, 4589, 48308, 779, 4673, 20, 21 ], "markers": [ 47, 58, 69 ], "logits_raw": [ 0.49528419971466064, 1.0892655849456787, -9.397531509399414 ], "probs_raw": [ 0.35571545362472534, 0.6442665457725525, 1.7976592062041163e-05 ], "probs": [ 0.39535412192344666, 0.6043087244033813, 0.0003372168866917491 ], "argmax_index": 1, "argmax": "insufficient", "top2_margin": 0.2089546024799347, "near_tie": false, "gold": "contradicted", "correct": false, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "insufficient", "confidence": 0.6043087244033813, "probabilities": { "contradicted": 0.39535412192344666, "insufficient": 0.6043087244033813, "supported": 0.0003372168866917491 } } }, "usage": { "input_tokens": 79, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_2175379f4c69e6207626", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 65, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 65, "positions": 65, "bucket": 256, "ids": [ 1, 17, 55945, 931, 835, 535, 6567, 938, 33037, 29067, 7441, 536, 1295, 1927, 953, 2685, 523, 46050, 931, 879, 535, 941, 7850, 33037, 1927, 7410, 523, 18, 52739, 779, 13719, 15603, 902, 6023, 1927, 1725, 523, 19, 16, 28441, 535, 42749, 13719, 14350, 779, 21968, 4589, 20, 19, 16, 879, 535, 46050, 931, 879, 20, 19, 16, 835, 535, 46050, 931, 835, 20, 21 ], "markers": [ 38, 49, 57 ], "logits_raw": [ 0.35558879375457764, -0.2648470997810364, 0.7199549674987793 ], "probs_raw": [ 0.3358735144138336, 0.1806025356054306, 0.4835239052772522 ], "probs": [ 0.34022027254104614, 0.2184140384197235, 0.44136565923690796 ], "argmax_index": 2, "argmax": "A", "top2_margin": 0.10114538669586182, "near_tie": false, "gold": "B", "correct": false, "host_softmax_max_abs_dp": 2.9802322387695312e-08, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "A", "confidence": 0.44136565923690796, "probabilities": { "insufficient": 0.34022027254104614, "B": 0.2184140384197235, "A": 0.44136565923690796 } } }, "usage": { "input_tokens": 65, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_21ca7a8b970e36b8e20f", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 99, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 99, "positions": 99, "bucket": 256, "ids": [ 1, 17, 1098, 16396, 3575, 6229, 779, 5937, 11430, 811, 51132, 41955, 523, 941, 5937, 11430, 17498, 779, 4889, 811, 902, 16701, 521, 1316, 1178, 1295, 1312, 17784, 10562, 523, 18, 1098, 5937, 11430, 1400, 51132, 41955, 1203, 1400, 1014, 40500, 896, 10562, 523, 11005, 30403, 2396, 2053, 1400, 51132, 523, 3354, 779, 16701, 51132, 768, 3346, 540, 19, 16, 33776, 535, 941, 9519, 7330, 14480, 1512, 779, 4272, 20, 19, 16, 28441, 535, 941, 21706, 2203, 2177, 1014, 27169, 3984, 779, 7330, 22558, 779, 4272, 20, 19, 16, 20458, 535, 941, 9519, 7330, 22558, 779, 4272, 20, 21 ], "markers": [ 59, 71, 88 ], "logits_raw": [ 0.8783601522445679, 1.3168089389801025, 3.038339614868164 ], "probs_raw": [ 0.08911655843257904, 0.1381574273109436, 0.7727259397506714 ], "probs": [ 0.14192068576812744, 0.19411849975585938, 0.6639607548713684 ], "argmax_index": 2, "argmax": "permitted", "top2_margin": 0.46984225511550903, "near_tie": false, "gold": "prohibited", "correct": false, "host_softmax_max_abs_dp": 1.4901161193847656e-08, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "permitted", "confidence": 0.6639607548713684, "probabilities": { "prohibited": 0.14192068576812744, "insufficient": 0.19411849975585938, "permitted": 0.6639607548713684 } } }, "usage": { "input_tokens": 99, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_22ec57822db50ec121cf", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 61, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 61, "positions": 61, "bucket": 256, "ids": [ 1, 17, 55945, 931, 835, 535, 859, 64078, 26677, 768, 4025, 17686, 523, 46050, 931, 879, 535, 859, 22879, 26677, 768, 4025, 17686, 523, 18, 52739, 779, 13719, 12469, 768, 4025, 17686, 523, 19, 16, 28441, 535, 42749, 13719, 14350, 779, 21968, 4589, 20, 19, 16, 835, 535, 46050, 931, 835, 20, 19, 16, 879, 535, 46050, 931, 879, 20, 21 ], "markers": [ 34, 45, 53 ], "logits_raw": [ 0.028011880815029144, 0.252865195274353, 1.7705678939819336 ], "probs_raw": [ 0.12556403875350952, 0.157223641872406, 0.7172123193740845 ], "probs": [ 0.17710436880588531, 0.20796306431293488, 0.6149325966835022 ], "argmax_index": 2, "argmax": "B", "top2_margin": 0.4069695323705673, "near_tie": false, "gold": "A", "correct": false, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "B", "confidence": 0.6149325966835022, "probabilities": { "insufficient": 0.17710436880588531, "A": 0.20796306431293488, "B": 0.6149325966835022 } } }, "usage": { "input_tokens": 61, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_25172cd6ddc49ab0ca70", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 70, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 70, "positions": 70, "bucket": 256, "ids": [ 1, 17, 1098, 22773, 22922, 2396, 2240, 3507, 3893, 803, 779, 59534, 523, 18, 31678, 3893, 856, 6767, 1550, 1720, 3507, 25252, 980, 17784, 523, 3354, 936, 6305, 540, 19, 16, 33776, 535, 941, 9519, 7330, 14480, 1512, 779, 4272, 20, 19, 16, 28441, 535, 941, 21706, 2203, 2177, 1014, 27169, 3984, 779, 7330, 22558, 779, 4272, 20, 19, 16, 20458, 535, 941, 9519, 7330, 22558, 779, 4272, 20, 21 ], "markers": [ 30, 42, 59 ], "logits_raw": [ 2.136638641357422, 3.2041687965393066, 0.9671251773834229 ], "probs_raw": [ 0.2370395064353943, 0.6893553733825684, 0.0736050233244896 ], "probs": [ 0.27952614426612854, 0.5992463827133179, 0.1212274357676506 ], "argmax_index": 1, "argmax": "insufficient", "top2_margin": 0.31972023844718933, "near_tie": false, "gold": "permitted", "correct": false, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "insufficient", "confidence": 0.5992463827133179, "probabilities": { "prohibited": 0.27952614426612854, "insufficient": 0.5992463827133179, "permitted": 0.1212274357676506 } } }, "usage": { "input_tokens": 70, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_2608370d01783735c3ea", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 71, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 71, "positions": 71, "bucket": 256, "ids": [ 1, 17, 1098, 40348, 10234, 26956, 779, 7627, 811, 4963, 523, 8874, 3575, 12800, 768, 23819, 27809, 5105, 953, 5405, 810, 4963, 953, 36072, 523, 18, 9886, 894, 779, 4673, 535, 779, 35101, 5562, 856, 797, 4963, 523, 19, 16, 7951, 535, 941, 4589, 48308, 779, 4673, 20, 19, 16, 28441, 535, 941, 4589, 2177, 1014, 6122, 4021, 20, 19, 16, 18197, 13735, 535, 941, 4589, 48308, 779, 11977, 20, 21 ], "markers": [ 39, 49, 60 ], "logits_raw": [ -6.859125137329102, 0.526771068572998, -1.9133896827697754 ], "probs_raw": [ 0.0005699152243323624, 0.9193147420883179, 0.08011537790298462 ], "probs": [ 0.004332481883466244, 0.8473929166793823, 0.14827457070350647 ], "argmax_index": 1, "argmax": "insufficient", "top2_margin": 0.6991183459758759, "near_tie": false, "gold": "contradicted", "correct": false, "host_softmax_max_abs_dp": 4.656612873077393e-10, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "insufficient", "confidence": 0.8473929166793823, "probabilities": { "supported": 0.004332481883466244, "insufficient": 0.8473929166793823, "contradicted": 0.14827457070350647 } } }, "usage": { "input_tokens": 71, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_2708193212d8a4a523c7", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 95, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 95, "positions": 95, "bucket": 256, "ids": [ 1, 17, 1098, 4887, 11917, 810, 2254, 2059, 10547, 1186, 938, 3161, 810, 11500, 875, 779, 8324, 523, 18, 21509, 779, 2254, 1550, 906, 768, 3981, 3321, 1403, 779, 4887, 11917, 856, 58754, 933, 4994, 3972, 1108, 536, 9141, 2643, 779, 4887, 11917, 523, 3354, 779, 2254, 874, 2643, 791, 1490, 1033, 5792, 540, 19, 16, 33776, 535, 941, 9519, 7330, 14480, 1512, 779, 4272, 20, 19, 16, 20458, 535, 941, 9519, 7330, 22558, 779, 4272, 20, 19, 16, 28441, 535, 941, 21706, 2203, 2177, 1014, 27169, 3984, 779, 7330, 22558, 779, 4272, 20, 21 ], "markers": [ 55, 67, 78 ], "logits_raw": [ 0.818483293056488, 3.5373549461364746, 2.113175868988037 ], "probs_raw": [ 0.0504717193543911, 0.765312910079956, 0.18421539664268494 ], "probs": [ 0.09527885168790817, 0.6644770503044128, 0.24024412035942078 ], "argmax_index": 1, "argmax": "permitted", "top2_margin": 0.42423292994499207, "near_tie": false, "gold": "prohibited", "correct": false, "host_softmax_max_abs_dp": 2.9802322387695312e-08, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "permitted", "confidence": 0.6644770503044128, "probabilities": { "prohibited": 0.09527885168790817, "permitted": 0.6644770503044128, "insufficient": 0.24024412035942078 } } }, "usage": { "input_tokens": 95, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_2bf540d515a51755b25c", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 68, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 68, "positions": 68, "bucket": 256, "ids": [ 1, 17, 55945, 931, 835, 535, 13831, 41071, 1124, 779, 11019, 523, 46050, 931, 879, 535, 13085, 1123, 1531, 936, 1499, 2144, 521, 1203, 1112, 1014, 41071, 1124, 936, 4763, 523, 18, 52739, 779, 13719, 3575, 4651, 41071, 3665, 523, 19, 16, 28441, 535, 42749, 13719, 14350, 779, 21968, 4589, 20, 19, 16, 879, 535, 46050, 931, 879, 20, 19, 16, 835, 535, 46050, 931, 835, 20, 21 ], "markers": [ 41, 52, 60 ], "logits_raw": [ 0.8004128932952881, 0.6302740573883057, 0.15096712112426758 ], "probs_raw": [ 0.4226751923561096, 0.35654670000076294, 0.22077812254428864 ], "probs": [ 0.39771389961242676, 0.352198988199234, 0.25008711218833923 ], "argmax_index": 0, "argmax": "insufficient", "top2_margin": 0.04551491141319275, "near_tie": false, "gold": "A", "correct": false, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "insufficient", "confidence": 0.39771389961242676, "probabilities": { "insufficient": 0.39771389961242676, "B": 0.352198988199234, "A": 0.25008711218833923 } } }, "usage": { "input_tokens": 68, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_31eb4533d2c89959ec46", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 90, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 90, "positions": 90, "bucket": 256, "ids": [ 1, 17, 1098, 2681, 28117, 6516, 902, 2960, 15001, 6177, 521, 1203, 13834, 1352, 4140, 4689, 1523, 2919, 9448, 938, 3161, 523, 18, 14875, 779, 2681, 28117, 1090, 10500, 15001, 6177, 536, 768, 28687, 32491, 1706, 779, 7045, 8825, 523, 3354, 779, 8153, 777, 7157, 779, 1530, 793, 540, 19, 16, 28441, 535, 941, 21706, 2203, 2177, 1014, 27169, 3984, 779, 7330, 22558, 779, 4272, 20, 19, 16, 33776, 535, 941, 9519, 7330, 14480, 1512, 779, 4272, 20, 19, 16, 20458, 535, 941, 9519, 7330, 22558, 779, 4272, 20, 21 ], "markers": [ 50, 67, 79 ], "logits_raw": [ 2.566591262817383, 0.23663032054901123, 0.29303932189941406 ], "probs_raw": [ 0.8331629037857056, 0.08106638491153717, 0.08577068150043488 ], "probs": [ 0.7212923765182495, 0.1365465521812439, 0.1421610563993454 ], "argmax_index": 0, "argmax": "insufficient", "top2_margin": 0.5791313201189041, "near_tie": false, "gold": "insufficient", "correct": true, "host_softmax_max_abs_dp": 1.4901161193847656e-08, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "insufficient", "confidence": 0.7212923765182495, "probabilities": { "insufficient": 0.7212923765182495, "prohibited": 0.1365465521812439, "permitted": 0.1421610563993454 } } }, "usage": { "input_tokens": 90, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_3226202ba1ada2570e9f", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 79, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 79, "positions": 79, "bucket": 256, "ids": [ 1, 17, 1098, 3434, 919, 5609, 797, 779, 20610, 1203, 4117, 5329, 936, 523, 18, 15246, 779, 8530, 7330, 521, 768, 2230, 6835, 3822, 797, 779, 20610, 536, 22115, 6955, 856, 28441, 523, 3354, 936, 874, 12800, 540, 19, 16, 33776, 535, 941, 9519, 7330, 14480, 1512, 779, 4272, 20, 19, 16, 20458, 535, 941, 9519, 7330, 22558, 779, 4272, 20, 19, 16, 28441, 535, 941, 21706, 2203, 2177, 1014, 27169, 3984, 779, 7330, 22558, 779, 4272, 20, 21 ], "markers": [ 39, 51, 62 ], "logits_raw": [ 1.931678295135498, 1.6135848760604858, 2.03139328956604 ], "probs_raw": [ 0.3530586063861847, 0.25686243176460266, 0.39007890224456787 ], "probs": [ 0.3483629524707794, 0.2775549590587616, 0.3740820288658142 ], "argmax_index": 2, "argmax": "insufficient", "top2_margin": 0.02571907639503479, "near_tie": false, "gold": "prohibited", "correct": false, "host_softmax_max_abs_dp": 2.9802322387695312e-08, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "insufficient", "confidence": 0.3740820288658142, "probabilities": { "prohibited": 0.3483629524707794, "permitted": 0.2775549590587616, "insufficient": 0.3740820288658142 } } }, "usage": { "input_tokens": 79, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_32f36cd0fbbfb3604586", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 68, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 68, "positions": 68, "bucket": 256, "ids": [ 1, 17, 1098, 11860, 2090, 1321, 6835, 779, 3956, 8649, 810, 5196, 8170, 523, 941, 11746, 3084, 44987, 896, 2059, 11423, 523, 18, 9886, 894, 779, 4673, 535, 779, 11860, 2090, 1321, 856, 5156, 523, 19, 16, 7951, 535, 941, 4589, 48308, 779, 4673, 20, 19, 16, 28441, 535, 941, 4589, 2177, 1014, 6122, 4021, 20, 19, 16, 18197, 13735, 535, 941, 4589, 48308, 779, 11977, 20, 21 ], "markers": [ 36, 46, 57 ], "logits_raw": [ -1.0280404090881348, 0.17128825187683105, -1.9770114421844482 ], "probs_raw": [ 0.21253859996795654, 0.7051793932914734, 0.08228201419115067 ], "probs": [ 0.258857786655426, 0.6097228527069092, 0.1314193308353424 ], "argmax_index": 1, "argmax": "insufficient", "top2_margin": 0.35086506605148315, "near_tie": false, "gold": "supported", "correct": false, "host_softmax_max_abs_dp": 2.9802322387695312e-08, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "insufficient", "confidence": 0.6097228527069092, "probabilities": { "supported": 0.258857786655426, "insufficient": 0.6097228527069092, "contradicted": 0.1314193308353424 } } }, "usage": { "input_tokens": 68, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_33d6e5da58f0fee2490d", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 76, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 76, "positions": 76, "bucket": 256, "ids": [ 1, 17, 1098, 8170, 7850, 11820, 37414, 9007, 4611, 1203, 1014, 1509, 9764, 833, 4611, 523, 18, 1098, 1509, 9764, 833, 59534, 6835, 1550, 1509, 9764, 833, 4611, 523, 3354, 1033, 8170, 6984, 936, 540, 19, 16, 33776, 535, 941, 9519, 7330, 14480, 1512, 779, 4272, 20, 19, 16, 28441, 535, 941, 21706, 2203, 2177, 1014, 27169, 3984, 779, 7330, 22558, 779, 4272, 20, 19, 16, 20458, 535, 941, 9519, 7330, 22558, 779, 4272, 20, 21 ], "markers": [ 36, 48, 65 ], "logits_raw": [ -2.6971306800842285, 2.78682541847229, -2.3960108757019043 ], "probs_raw": [ 0.004112706519663334, 0.9903295040130615, 0.0055577936582267284 ], "probs": [ 0.019044214859604836, 0.9573412537574768, 0.023614587262272835 ], "argmax_index": 1, "argmax": "insufficient", "top2_margin": 0.933726666495204, "near_tie": false, "gold": "prohibited", "correct": false, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "insufficient", "confidence": 0.9573412537574768, "probabilities": { "prohibited": 0.019044214859604836, "insufficient": 0.9573412537574768, "permitted": 0.023614587262272835 } } }, "usage": { "input_tokens": 76, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_36273dd4777876a26338", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 60, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 60, "positions": 60, "bucket": 256, "ids": [ 1, 17, 1098, 3059, 5813, 779, 45853, 810, 8112, 779, 61026, 1090, 8744, 4623, 523, 18, 9886, 894, 779, 4673, 535, 779, 61026, 1178, 1381, 45943, 523, 19, 16, 18197, 13735, 535, 941, 4589, 48308, 779, 11977, 20, 19, 16, 7951, 535, 941, 4589, 48308, 779, 4673, 20, 19, 16, 28441, 535, 941, 4589, 2177, 1014, 6122, 4021, 20, 21 ], "markers": [ 28, 39, 49 ], "logits_raw": [ -2.4476537704467773, 8.073257446289062, -0.6171743869781494 ], "probs_raw": [ 2.696134652069304e-05, 0.9998049139976501, 0.00016815461276564747 ], "probs": [ 0.000543197151273489, 0.9974485039710999, 0.002008317969739437 ], "argmax_index": 1, "argmax": "supported", "top2_margin": 0.9954401860013604, "near_tie": false, "gold": "supported", "correct": true, "host_softmax_max_abs_dp": 2.3283064365386963e-10, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "supported", "confidence": 0.9974485039710999, "probabilities": { "contradicted": 0.000543197151273489, "supported": 0.9974485039710999, "insufficient": 0.002008317969739437 } } }, "usage": { "input_tokens": 60, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_36c71bb649b7d5eccb1f", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 60, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 60, "positions": 60, "bucket": 256, "ids": [ 1, 17, 1098, 59063, 44987, 19459, 60468, 856, 4175, 810, 19459, 42074, 788, 856, 49550, 523, 18, 9886, 894, 779, 4673, 535, 19459, 60468, 856, 4175, 523, 19, 16, 7951, 535, 941, 4589, 48308, 779, 4673, 20, 19, 16, 28441, 535, 941, 4589, 2177, 1014, 6122, 4021, 20, 19, 16, 18197, 13735, 535, 941, 4589, 48308, 779, 11977, 20, 21 ], "markers": [ 28, 38, 49 ], "logits_raw": [ 7.094854354858398, 0.14647667109966278, -2.7314815521240234 ], "probs_raw": [ 0.9989867806434631, 0.0009592188289389014, 5.395556945586577e-05 ], "probs": [ 0.9921786785125732, 0.0069338493049144745, 0.0008874361519701779 ], "argmax_index": 0, "argmax": "supported", "top2_margin": 0.9852448292076588, "near_tie": false, "gold": "supported", "correct": true, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "supported", "confidence": 0.9921786785125732, "probabilities": { "supported": 0.9921786785125732, "insufficient": 0.0069338493049144745, "contradicted": 0.0008874361519701779 } } }, "usage": { "input_tokens": 60, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_37cebc3792e98049439f", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 88, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 88, "positions": 88, "bucket": 256, "ids": [ 1, 17, 1098, 45609, 7330, 9938, 592, 38183, 811, 779, 797, 41722, 23120, 1403, 936, 1011, 2463, 779, 1336, 536, 1517, 779, 5329, 23120, 1550, 1403, 779, 797, 41722, 23120, 4584, 523, 9758, 29142, 10547, 1186, 1011, 2463, 1033, 12940, 523, 18, 9886, 894, 779, 4673, 535, 779, 5329, 23120, 1011, 2463, 1033, 12940, 523, 19, 16, 28441, 535, 941, 4589, 2177, 1014, 6122, 4021, 20, 19, 16, 7951, 535, 941, 4589, 48308, 779, 4673, 20, 19, 16, 18197, 13735, 535, 941, 4589, 48308, 779, 11977, 20, 21 ], "markers": [ 56, 67, 77 ], "logits_raw": [ 1.0944440364837646, 2.0053186416625977, -0.7762060165405273 ], "probs_raw": [ 0.2746860384941101, 0.6830058097839355, 0.042308107018470764 ], "probs": [ 0.3144998848438263, 0.6028423309326172, 0.08265769481658936 ], "argmax_index": 1, "argmax": "supported", "top2_margin": 0.2883424460887909, "near_tie": false, "gold": "supported", "correct": true, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "supported", "confidence": 0.6028423309326172, "probabilities": { "insufficient": 0.3144998848438263, "supported": 0.6028423309326172, "contradicted": 0.08265769481658936 } } }, "usage": { "input_tokens": 88, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_386e10b1c1ce906473a4", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 98, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 98, "positions": 98, "bucket": 256, "ids": [ 1, 17, 49290, 5019, 11528, 6261, 875, 20155, 1230, 1798, 521, 6794, 896, 10144, 39208, 3469, 1134, 15419, 2563, 523, 46050, 931, 835, 535, 4411, 38120, 953, 5019, 1929, 803, 768, 20155, 1230, 1047, 810, 5260, 779, 4225, 19214, 523, 46050, 931, 879, 535, 29207, 953, 5019, 1929, 803, 768, 20155, 1230, 1047, 1203, 39273, 779, 19214, 523, 18, 15534, 779, 38120, 46102, 1490, 779, 9519, 11287, 810, 32294, 523, 19, 16, 879, 535, 46050, 931, 879, 20, 19, 16, 835, 535, 46050, 931, 835, 20, 19, 16, 28441, 535, 42749, 13719, 14350, 779, 21968, 4589, 20, 21 ], "markers": [ 71, 79, 87 ], "logits_raw": [ 1.2154874801635742, 3.1739749908447266, 1.4671909809112549 ], "probs_raw": [ 0.10666880756616592, 0.7561322450637817, 0.13719898462295532 ], "probs": [ 0.16004574298858643, 0.6483826637268066, 0.19157147407531738 ], "argmax_index": 1, "argmax": "A", "top2_margin": 0.45681118965148926, "near_tie": false, "gold": "A", "correct": true, "host_softmax_max_abs_dp": 1.4901161193847656e-08, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "A", "confidence": 0.6483826637268066, "probabilities": { "B": 0.16004574298858643, "A": 0.6483826637268066, "insufficient": 0.19157147407531738 } } }, "usage": { "input_tokens": 98, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_3b0958cabc339005cf66", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 63, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 63, "positions": 63, "bucket": 256, "ids": [ 1, 17, 542, 6634, 35283, 902, 7267, 2630, 19736, 768, 12037, 933, 3984, 779, 3863, 856, 37640, 523, 18, 9886, 894, 779, 4673, 535, 779, 7267, 12037, 953, 37640, 523, 19, 16, 28441, 535, 941, 4589, 2177, 1014, 6122, 4021, 20, 19, 16, 7951, 535, 941, 4589, 48308, 779, 4673, 20, 19, 16, 18197, 13735, 535, 941, 4589, 48308, 779, 11977, 20, 21 ], "markers": [ 31, 42, 52 ], "logits_raw": [ 0.8897602558135986, -6.432289123535156, -0.8964612483978271 ], "probs_raw": [ 0.8559790253639221, 0.000565636670216918, 0.14345543086528778 ], "probs": [ 0.7785053253173828, 0.004166017286479473, 0.21732863783836365 ], "argmax_index": 0, "argmax": "insufficient", "top2_margin": 0.5611766874790192, "near_tie": false, "gold": "insufficient", "correct": true, "host_softmax_max_abs_dp": 1.4901161193847656e-08, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "insufficient", "confidence": 0.7785053253173828, "probabilities": { "insufficient": 0.7785053253173828, "supported": 0.004166017286479473, "contradicted": 0.21732863783836365 } } }, "usage": { "input_tokens": 63, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_3dd9c3ac98f8ddf7bdd3", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 64, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 64, "positions": 64, "bucket": 256, "ids": [ 1, 17, 1098, 3059, 12930, 12874, 875, 45853, 779, 61026, 810, 22922, 3498, 1295, 45853, 1178, 1381, 5813, 523, 18, 9886, 894, 779, 4673, 535, 779, 4479, 6034, 45853, 12874, 523, 19, 16, 18197, 13735, 535, 941, 4589, 48308, 779, 11977, 20, 19, 16, 7951, 535, 941, 4589, 48308, 779, 4673, 20, 19, 16, 28441, 535, 941, 4589, 2177, 1014, 6122, 4021, 20, 21 ], "markers": [ 32, 43, 53 ], "logits_raw": [ -0.29899075627326965, -2.115078926086426, -0.10533981025218964 ], "probs_raw": [ 0.4208163917064667, 0.06845032423734665, 0.5107333064079285 ], "probs": [ 0.4129457473754883, 0.11284510046243668, 0.47420912981033325 ], "argmax_index": 2, "argmax": "insufficient", "top2_margin": 0.06126338243484497, "near_tie": false, "gold": "supported", "correct": false, "host_softmax_max_abs_dp": 2.9802322387695312e-08, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "insufficient", "confidence": 0.47420912981033325, "probabilities": { "contradicted": 0.4129457473754883, "supported": 0.11284510046243668, "insufficient": 0.47420912981033325 } } }, "usage": { "input_tokens": 64, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_3e75f2b84e8e08e3d3f2", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 60, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 60, "positions": 60, "bucket": 256, "ids": [ 1, 17, 55945, 931, 835, 535, 941, 10046, 6289, 779, 13732, 523, 46050, 931, 879, 535, 941, 5647, 1178, 768, 7043, 20453, 523, 18, 52739, 779, 13719, 12469, 768, 4025, 17686, 523, 19, 16, 879, 535, 46050, 931, 879, 20, 19, 16, 835, 535, 46050, 931, 835, 20, 19, 16, 28441, 535, 42749, 13719, 14350, 779, 21968, 4589, 20, 21 ], "markers": [ 33, 41, 49 ], "logits_raw": [ 0.2606961727142334, -3.224853515625, 0.07640176266431808 ], "probs_raw": [ 0.5369623899459839, 0.016450868919491768, 0.4465867280960083 ], "probs": [ 0.5103158354759216, 0.04231620952486992, 0.44736793637275696 ], "argmax_index": 0, "argmax": "B", "top2_margin": 0.06294789910316467, "near_tie": false, "gold": "insufficient", "correct": false, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "B", "confidence": 0.5103158354759216, "probabilities": { "B": 0.5103158354759216, "A": 0.04231620952486992, "insufficient": 0.44736793637275696 } } }, "usage": { "input_tokens": 60, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_3f783eec5c90812479bc", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 76, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 76, "positions": 76, "bucket": 256, "ids": [ 1, 17, 1098, 3434, 919, 5609, 5329, 779, 20610, 810, 4117, 5662, 936, 523, 18, 542, 12518, 1090, 2230, 6835, 4624, 797, 779, 20610, 810, 1178, 1295, 21701, 27434, 523, 3354, 936, 874, 12800, 540, 19, 16, 20458, 535, 941, 9519, 7330, 22558, 779, 4272, 20, 19, 16, 33776, 535, 941, 9519, 7330, 14480, 1512, 779, 4272, 20, 19, 16, 28441, 535, 941, 21706, 2203, 2177, 1014, 27169, 3984, 779, 7330, 22558, 779, 4272, 20, 21 ], "markers": [ 36, 47, 59 ], "logits_raw": [ 2.9592502117156982, 1.7731610536575317, 2.0034730434417725 ], "probs_raw": [ 0.591741681098938, 0.1807258129119873, 0.2275324910879135 ], "probs": [ 0.5171136260032654, 0.22162669897079468, 0.26125967502593994 ], "argmax_index": 0, "argmax": "permitted", "top2_margin": 0.25585395097732544, "near_tie": false, "gold": "prohibited", "correct": false, "host_softmax_max_abs_dp": 2.9802322387695312e-08, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "permitted", "confidence": 0.5171136260032654, "probabilities": { "permitted": 0.5171136260032654, "prohibited": 0.22162669897079468, "insufficient": 0.26125967502593994 } } }, "usage": { "input_tokens": 76, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_4089c199b8b796ec3ed8", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 94, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 94, "positions": 94, "bucket": 256, "ids": [ 1, 17, 1098, 1547, 40111, 938, 5668, 521, 1203, 13834, 16425, 4689, 14503, 2583, 856, 4617, 523, 18, 21509, 779, 2254, 1550, 906, 768, 3981, 3321, 1403, 779, 4887, 11917, 856, 58754, 933, 4994, 3972, 1108, 536, 9141, 2643, 779, 4887, 11917, 523, 3354, 779, 2254, 874, 2643, 791, 1490, 1033, 5792, 540, 19, 16, 28441, 535, 941, 21706, 2203, 2177, 1014, 27169, 3984, 779, 7330, 22558, 779, 4272, 20, 19, 16, 33776, 535, 941, 9519, 7330, 14480, 1512, 779, 4272, 20, 19, 16, 20458, 535, 941, 9519, 7330, 22558, 779, 4272, 20, 21 ], "markers": [ 54, 71, 83 ], "logits_raw": [ 2.3879141807556152, 0.4206157922744751, 2.479020833969116 ], "probs_raw": [ 0.44738325476646423, 0.06255944818258286, 0.4900572597980499 ], "probs": [ 0.432425856590271, 0.10606959462165833, 0.4615044593811035 ], "argmax_index": 2, "argmax": "permitted", "top2_margin": 0.02907860279083252, "near_tie": false, "gold": "insufficient", "correct": false, "host_softmax_max_abs_dp": 7.450580596923828e-09, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "permitted", "confidence": 0.4615044593811035, "probabilities": { "insufficient": 0.432425856590271, "prohibited": 0.10606959462165833, "permitted": 0.4615044593811035 } } }, "usage": { "input_tokens": 94, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_42384af9ae56beb37ff5", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 91, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 91, "positions": 91, "bucket": 256, "ids": [ 1, 17, 1098, 7330, 22558, 2314, 53859, 968, 779, 7576, 1203, 23369, 2796, 1523, 3759, 53859, 523, 46050, 931, 835, 535, 859, 5068, 3346, 17784, 10562, 5828, 988, 779, 7576, 523, 46050, 931, 879, 535, 859, 5068, 936, 1550, 988, 835, 521, 810, 1052, 1295, 1312, 8825, 523, 18, 15534, 779, 13719, 5977, 13289, 10562, 4834, 1550, 1626, 768, 33776, 2341, 53859, 523, 19, 16, 28441, 535, 42749, 13719, 14350, 779, 21968, 4589, 20, 19, 16, 835, 535, 46050, 931, 835, 20, 19, 16, 879, 535, 46050, 931, 879, 20, 21 ], "markers": [ 64, 75, 83 ], "logits_raw": [ 1.6472855806350708, 2.0896692276000977, 2.033372163772583 ], "probs_raw": [ 0.24828527867794037, 0.3864343762397766, 0.36528030037879944 ], "probs": [ 0.27105966210365295, 0.3717978298664093, 0.3571425676345825 ], "argmax_index": 1, "argmax": "A", "top2_margin": 0.014655262231826782, "near_tie": true, "gold": "B", "correct": false, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "A", "confidence": 0.3717978298664093, "probabilities": { "insufficient": 0.27105966210365295, "A": 0.3717978298664093, "B": 0.3571425676345825 } } }, "usage": { "input_tokens": 91, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_42c54fa2ad45743e24d8", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 81, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 81, "positions": 81, "bucket": 256, "ids": [ 1, 17, 49290, 5019, 11528, 6261, 875, 20155, 1230, 1798, 521, 6794, 896, 10144, 39208, 3469, 1134, 15419, 2563, 523, 46050, 931, 835, 810, 46050, 931, 879, 2312, 5019, 11528, 6261, 1203, 3176, 13834, 1230, 1047, 20306, 4689, 19214, 2203, 523, 18, 15534, 779, 38120, 46102, 1490, 779, 9519, 11287, 810, 32294, 523, 19, 16, 835, 535, 46050, 931, 835, 20, 19, 16, 28441, 535, 42749, 13719, 14350, 779, 21968, 4589, 20, 19, 16, 879, 535, 46050, 931, 879, 20, 21 ], "markers": [ 54, 62, 73 ], "logits_raw": [ 0.7830364108085632, 2.160759449005127, 1.3004415035247803 ], "probs_raw": [ 0.1505223959684372, 0.596950888633728, 0.2525267004966736 ], "probs": [ 0.19520875811576843, 0.5222931504249573, 0.2824980318546295 ], "argmax_index": 1, "argmax": "insufficient", "top2_margin": 0.23979511857032776, "near_tie": false, "gold": "insufficient", "correct": true, "host_softmax_max_abs_dp": 1.4901161193847656e-08, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "insufficient", "confidence": 0.5222931504249573, "probabilities": { "A": 0.19520875811576843, "insufficient": 0.5222931504249573, "B": 0.2824980318546295 } } }, "usage": { "input_tokens": 81, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_4444dad25e438f128673", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 76, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 76, "positions": 76, "bucket": 256, "ids": [ 1, 17, 1098, 3434, 919, 5609, 797, 779, 20610, 1203, 4117, 5329, 936, 523, 18, 542, 12518, 1090, 2230, 6835, 4624, 797, 779, 20610, 810, 1178, 1295, 21701, 27434, 523, 3354, 936, 874, 12800, 540, 19, 16, 33776, 535, 941, 9519, 7330, 14480, 1512, 779, 4272, 20, 19, 16, 20458, 535, 941, 9519, 7330, 22558, 779, 4272, 20, 19, 16, 28441, 535, 941, 21706, 2203, 2177, 1014, 27169, 3984, 779, 7330, 22558, 779, 4272, 20, 21 ], "markers": [ 36, 48, 59 ], "logits_raw": [ 2.421286106109619, 4.4906415939331055, 2.1723551750183105 ], "probs_raw": [ 0.10309968143701553, 0.8165203332901001, 0.08038000017404556 ], "probs": [ 0.16071435809135437, 0.7047528624534607, 0.13453277945518494 ], "argmax_index": 1, "argmax": "permitted", "top2_margin": 0.5440385043621063, "near_tie": false, "gold": "permitted", "correct": true, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "permitted", "confidence": 0.7047528624534607, "probabilities": { "prohibited": 0.16071435809135437, "permitted": 0.7047528624534607, "insufficient": 0.13453277945518494 } } }, "usage": { "input_tokens": 76, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_44574c44e243ee1b2f52", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 71, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 71, "positions": 71, "bucket": 256, "ids": [ 1, 17, 1098, 40348, 10234, 26956, 779, 7627, 811, 4963, 523, 8874, 3575, 12800, 768, 23819, 27809, 5105, 953, 5405, 810, 4963, 953, 36072, 523, 18, 9886, 894, 779, 4673, 535, 779, 35101, 5562, 856, 797, 5105, 523, 19, 16, 28441, 535, 941, 4589, 2177, 1014, 6122, 4021, 20, 19, 16, 18197, 13735, 535, 941, 4589, 48308, 779, 11977, 20, 19, 16, 7951, 535, 941, 4589, 48308, 779, 4673, 20, 21 ], "markers": [ 39, 50, 61 ], "logits_raw": [ -0.28942108154296875, -3.0245723724365234, 2.4573111534118652 ], "probs_raw": [ 0.06003668159246445, 0.003895431524142623, 0.9360678791999817 ], "probs": [ 0.12112487852573395, 0.017167197540402412, 0.8617079257965088 ], "argmax_index": 2, "argmax": "supported", "top2_margin": 0.7405830472707748, "near_tie": false, "gold": "supported", "correct": true, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "supported", "confidence": 0.8617079257965088, "probabilities": { "insufficient": 0.12112487852573395, "contradicted": 0.017167197540402412, "supported": 0.8617079257965088 } } }, "usage": { "input_tokens": 71, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_44bae0c8a1615a096c43", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 74, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 74, "positions": 74, "bucket": 256, "ids": [ 1, 17, 55945, 931, 835, 535, 941, 6634, 8269, 896, 779, 7850, 33037, 1927, 7410, 523, 46050, 931, 879, 535, 941, 6634, 15994, 33037, 29067, 7441, 1203, 3498, 1295, 1927, 953, 2685, 523, 18, 52739, 779, 13719, 8960, 1550, 902, 946, 1513, 58009, 1927, 9100, 523, 19, 16, 879, 535, 46050, 931, 879, 20, 19, 16, 835, 535, 46050, 931, 835, 20, 19, 16, 28441, 535, 42749, 13719, 14350, 779, 21968, 4589, 20, 21 ], "markers": [ 47, 55, 63 ], "logits_raw": [ -1.737524390220642, 0.3417292535305023, 0.8451249003410339 ], "probs_raw": [ 0.04498295858502388, 0.35979601740837097, 0.5952211022377014 ], "probs": [ 0.08515271544456482, 0.3760550320148468, 0.5387921929359436 ], "argmax_index": 2, "argmax": "insufficient", "top2_margin": 0.1627371609210968, "near_tie": false, "gold": "B", "correct": false, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "insufficient", "confidence": 0.5387921929359436, "probabilities": { "B": 0.08515271544456482, "A": 0.3760550320148468, "insufficient": 0.5387921929359436 } } }, "usage": { "input_tokens": 74, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_4533dd91267b4d815690", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 93, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 93, "positions": 93, "bucket": 256, "ids": [ 1, 17, 1098, 8414, 5799, 22654, 1178, 768, 12803, 13789, 810, 1236, 6034, 5231, 41095, 523, 9758, 11445, 938, 11649, 523, 18, 41914, 1587, 22879, 8414, 5799, 523, 45681, 936, 875, 768, 12803, 13789, 521, 6794, 896, 5231, 41095, 3469, 4201, 936, 797, 808, 16979, 523, 3354, 1033, 1789, 874, 11086, 540, 19, 16, 20458, 535, 941, 9519, 7330, 22558, 779, 4272, 20, 19, 16, 28441, 535, 941, 21706, 2203, 2177, 1014, 27169, 3984, 779, 7330, 22558, 779, 4272, 20, 19, 16, 33776, 535, 941, 9519, 7330, 14480, 1512, 779, 4272, 20, 21 ], "markers": [ 53, 64, 81 ], "logits_raw": [ 3.4331963062286377, 2.286992311477661, 1.4278408288955688 ], "probs_raw": [ 0.6884902119636536, 0.218830406665802, 0.09267935901880264 ], "probs": [ 0.5953508019447327, 0.2625322937965393, 0.14211688935756683 ], "argmax_index": 0, "argmax": "permitted", "top2_margin": 0.33281850814819336, "near_tie": false, "gold": "prohibited", "correct": false, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "permitted", "confidence": 0.5953508019447327, "probabilities": { "permitted": 0.5953508019447327, "insufficient": 0.2625322937965393, "prohibited": 0.14211688935756683 } } }, "usage": { "input_tokens": 93, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_46b7029b9a704138b77a", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 81, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 81, "positions": 81, "bucket": 256, "ids": [ 1, 17, 3312, 17980, 779, 12153, 30403, 33300, 1450, 52518, 523, 2357, 22092, 1689, 22922, 46159, 896, 52847, 523, 941, 3084, 3498, 17688, 1486, 23496, 7045, 15001, 521, 810, 1295, 2919, 15001, 953, 2997, 523, 18, 9886, 894, 779, 4673, 535, 779, 12153, 6516, 15001, 884, 17980, 523, 19, 16, 28441, 535, 941, 4589, 2177, 1014, 6122, 4021, 20, 19, 16, 18197, 13735, 535, 941, 4589, 48308, 779, 11977, 20, 19, 16, 7951, 535, 941, 4589, 48308, 779, 4673, 20, 21 ], "markers": [ 49, 60, 71 ], "logits_raw": [ 0.612383246421814, 0.810704231262207, 2.3589706420898438 ], "probs_raw": [ 0.12571735680103302, 0.1532939225435257, 0.7209887504577637 ], "probs": [ 0.17748259007930756, 0.2044944167137146, 0.6180230379104614 ], "argmax_index": 2, "argmax": "supported", "top2_margin": 0.4135286211967468, "near_tie": false, "gold": "supported", "correct": true, "host_softmax_max_abs_dp": 1.4901161193847656e-08, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "supported", "confidence": 0.6180230379104614, "probabilities": { "insufficient": 0.17748259007930756, "contradicted": 0.2044944167137146, "supported": 0.6180230379104614 } } }, "usage": { "input_tokens": 81, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_46d6a0731bd288c370f0", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 67, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 67, "positions": 67, "bucket": 256, "ids": [ 1, 17, 30729, 856, 5668, 797, 31804, 1337, 18577, 521, 1203, 1295, 10562, 25274, 933, 53859, 4450, 938, 21706, 523, 18, 9886, 894, 779, 4673, 535, 8376, 1178, 10562, 811, 51132, 16993, 18577, 523, 19, 16, 28441, 535, 941, 4589, 2177, 1014, 6122, 4021, 20, 19, 16, 18197, 13735, 535, 941, 4589, 48308, 779, 11977, 20, 19, 16, 7951, 535, 941, 4589, 48308, 779, 4673, 20, 21 ], "markers": [ 35, 46, 57 ], "logits_raw": [ 2.5726563930511475, -0.306399405002594, -4.652129173278809 ], "probs_raw": [ 0.9461489915847778, 0.053162023425102234, 0.0006890883669257164 ], "probs": [ 0.882128119468689, 0.1128116250038147, 0.005060174502432346 ], "argmax_index": 0, "argmax": "insufficient", "top2_margin": 0.7693164944648743, "near_tie": false, "gold": "insufficient", "correct": true, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "insufficient", "confidence": 0.882128119468689, "probabilities": { "insufficient": 0.882128119468689, "contradicted": 0.1128116250038147, "supported": 0.005060174502432346 } } }, "usage": { "input_tokens": 67, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_4aa83af95bb69e90bc2d", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 97, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 97, "positions": 97, "bucket": 256, "ids": [ 1, 17, 1098, 16396, 3575, 6229, 779, 5937, 11430, 811, 51132, 41955, 523, 941, 5937, 11430, 17498, 779, 4889, 811, 902, 16701, 521, 1316, 1178, 1295, 1312, 17784, 10562, 523, 18, 1098, 5937, 11430, 1400, 40500, 3346, 17784, 811, 902, 16701, 521, 810, 5296, 1101, 25274, 779, 16701, 17784, 10562, 523, 3354, 1033, 16701, 51132, 540, 19, 16, 20458, 535, 941, 9519, 7330, 22558, 779, 4272, 20, 19, 16, 33776, 535, 941, 9519, 7330, 14480, 1512, 779, 4272, 20, 19, 16, 28441, 535, 941, 21706, 2203, 2177, 1014, 27169, 3984, 779, 7330, 22558, 779, 4272, 20, 21 ], "markers": [ 57, 68, 80 ], "logits_raw": [ 2.168987989425659, 0.28088173270225525, 1.7221136093139648 ], "probs_raw": [ 0.5583527684211731, 0.08451125770807266, 0.35713595151901245 ], "probs": [ 0.5034531950950623, 0.1306791752576828, 0.36586761474609375 ], "argmax_index": 0, "argmax": "permitted", "top2_margin": 0.1375855803489685, "near_tie": false, "gold": "permitted", "correct": true, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "permitted", "confidence": 0.5034531950950623, "probabilities": { "permitted": 0.5034531950950623, "prohibited": 0.1306791752576828, "insufficient": 0.36586761474609375 } } }, "usage": { "input_tokens": 97, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_4aee37a6a155747766aa", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 95, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 95, "positions": 95, "bucket": 256, "ids": [ 1, 17, 1098, 26862, 53076, 7944, 779, 28553, 946, 38008, 521, 2158, 779, 10423, 4462, 1856, 7944, 936, 10274, 523, 9758, 7944, 938, 2835, 523, 18, 35516, 7494, 8544, 1403, 779, 28553, 856, 8112, 10274, 536, 797, 19980, 521, 779, 10423, 4462, 1856, 1391, 58340, 779, 26862, 53076, 523, 3354, 936, 874, 5789, 540, 19, 16, 20458, 535, 941, 9519, 7330, 22558, 779, 4272, 20, 19, 16, 33776, 535, 941, 9519, 7330, 14480, 1512, 779, 4272, 20, 19, 16, 28441, 535, 941, 21706, 2203, 2177, 1014, 27169, 3984, 779, 7330, 22558, 779, 4272, 20, 21 ], "markers": [ 55, 66, 78 ], "logits_raw": [ 1.37359619140625, 1.4480979442596436, 2.7218360900878906 ], "probs_raw": [ 0.16869127750396729, 0.18173906207084656, 0.6495696902275085 ], "probs": [ 0.21392832696437836, 0.22562186419963837, 0.5604497790336609 ], "argmax_index": 2, "argmax": "insufficient", "top2_margin": 0.3348279148340225, "near_tie": false, "gold": "permitted", "correct": false, "host_softmax_max_abs_dp": 1.4901161193847656e-08, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "insufficient", "confidence": 0.5604497790336609, "probabilities": { "permitted": 0.21392832696437836, "prohibited": 0.22562186419963837, "insufficient": 0.5604497790336609 } } }, "usage": { "input_tokens": 95, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_4b3d794c004566610181", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 71, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 71, "positions": 71, "bucket": 256, "ids": [ 1, 17, 1098, 6765, 4943, 876, 22558, 18109, 16711, 803, 779, 2984, 523, 18, 560, 965, 16711, 1550, 916, 779, 6765, 1090, 11599, 15001, 523, 3354, 16711, 874, 4997, 540, 19, 16, 33776, 535, 941, 9519, 7330, 14480, 1512, 779, 4272, 20, 19, 16, 28441, 535, 941, 21706, 2203, 2177, 1014, 27169, 3984, 779, 7330, 22558, 779, 4272, 20, 19, 16, 20458, 535, 941, 9519, 7330, 22558, 779, 4272, 20, 21 ], "markers": [ 31, 43, 60 ], "logits_raw": [ -0.2797635793685913, 1.3045634031295776, 4.427176475524902 ], "probs_raw": [ 0.008577145636081696, 0.04182225093245506, 0.9496005773544312 ], "probs": [ 0.030341804027557373, 0.09409179538488388, 0.8755663633346558 ], "argmax_index": 2, "argmax": "permitted", "top2_margin": 0.7814745679497719, "near_tie": false, "gold": "permitted", "correct": true, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "permitted", "confidence": 0.8755663633346558, "probabilities": { "prohibited": 0.030341804027557373, "insufficient": 0.09409179538488388, "permitted": 0.8755663633346558 } } }, "usage": { "input_tokens": 71, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_4d7a003342db5b0a17e3", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 79, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 79, "positions": 79, "bucket": 256, "ids": [ 1, 17, 55945, 931, 835, 535, 941, 42271, 44987, 43152, 803, 768, 8825, 5771, 1203, 22922, 3498, 1295, 9170, 6177, 1178, 1381, 2047, 523, 46050, 931, 879, 535, 941, 42271, 44987, 896, 779, 8825, 953, 14708, 523, 18, 52739, 779, 13719, 15603, 43152, 916, 779, 9170, 2510, 1044, 2017, 3019, 523, 19, 16, 28441, 535, 42749, 13719, 14350, 779, 21968, 4589, 20, 19, 16, 835, 535, 46050, 931, 835, 20, 19, 16, 879, 535, 46050, 931, 879, 20, 21 ], "markers": [ 52, 63, 71 ], "logits_raw": [ -1.0214812755584717, -1.9063987731933594, -0.2571205794811249 ], "probs_raw": [ 0.2808697521686554, 0.11592849344015121, 0.6032017469406128 ], "probs": [ 0.30695444345474243, 0.16313357651233673, 0.529911994934082 ], "argmax_index": 2, "argmax": "B", "top2_margin": 0.2229575514793396, "near_tie": false, "gold": "A", "correct": false, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "B", "confidence": 0.529911994934082, "probabilities": { "insufficient": 0.30695444345474243, "A": 0.16313357651233673, "B": 0.529911994934082 } } }, "usage": { "input_tokens": 79, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_4da6aedc1139224eeff2", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 104, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 104, "positions": 104, "bucket": 256, "ids": [ 1, 17, 49290, 5019, 11528, 6261, 875, 20155, 1230, 1798, 521, 6794, 896, 10144, 39208, 3469, 1134, 15419, 2563, 523, 46050, 931, 835, 535, 4411, 38120, 953, 5019, 1929, 803, 768, 20155, 1230, 1047, 810, 5260, 779, 4225, 19214, 523, 46050, 931, 879, 535, 29207, 953, 5019, 1929, 803, 768, 20155, 1230, 1047, 1203, 39273, 779, 19214, 523, 18, 15534, 779, 38120, 1134, 15419, 2441, 10285, 968, 779, 19214, 32294, 9618, 779, 1230, 1047, 11287, 523, 19, 16, 28441, 535, 42749, 13719, 14350, 779, 21968, 4589, 20, 19, 16, 879, 535, 46050, 931, 879, 20, 19, 16, 835, 535, 46050, 931, 835, 20, 21 ], "markers": [ 77, 88, 96 ], "logits_raw": [ 1.825056552886963, 0.6948303580284119, 2.1762826442718506 ], "probs_raw": [ 0.3644622266292572, 0.117706798017025, 0.517831027507782 ], "probs": [ 0.3661385178565979, 0.1633097380399704, 0.4705516993999481 ], "argmax_index": 2, "argmax": "A", "top2_margin": 0.10441318154335022, "near_tie": false, "gold": "B", "correct": false, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "A", "confidence": 0.4705516993999481, "probabilities": { "insufficient": 0.3661385178565979, "B": 0.1633097380399704, "A": 0.4705516993999481 } } }, "usage": { "input_tokens": 104, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_51ceab3ce622c7c759f1", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 78, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 78, "positions": 78, "bucket": 256, "ids": [ 1, 17, 10720, 768, 25473, 813, 1914, 521, 768, 3338, 7061, 1165, 779, 12037, 803, 768, 33390, 7267, 523, 941, 3893, 6634, 4478, 1033, 856, 37640, 810, 1295, 2151, 7267, 953, 7935, 523, 18, 9886, 894, 779, 4673, 535, 779, 7267, 12037, 953, 37640, 523, 19, 16, 18197, 13735, 535, 941, 4589, 48308, 779, 11977, 20, 19, 16, 28441, 535, 941, 4589, 2177, 1014, 6122, 4021, 20, 19, 16, 7951, 535, 941, 4589, 48308, 779, 4673, 20, 21 ], "markers": [ 46, 57, 68 ], "logits_raw": [ -0.9429095387458801, 0.016189217567443848, 6.5538153648376465 ], "probs_raw": [ 0.0005537894903682172, 0.0014450271846726537, 0.9980011582374573 ], "probs": [ 0.004657903220504522, 0.009241331368684769, 0.9861007332801819 ], "argmax_index": 2, "argmax": "supported", "top2_margin": 0.9768594019114971, "near_tie": false, "gold": "supported", "correct": true, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "supported", "confidence": 0.9861007332801819, "probabilities": { "contradicted": 0.004657903220504522, "insufficient": 0.009241331368684769, "supported": 0.9861007332801819 } } }, "usage": { "input_tokens": 78, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_533d4423d311de82b27d", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 82, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 82, "positions": 82, "bucket": 256, "ids": [ 1, 17, 55945, 931, 835, 535, 941, 13608, 21090, 1531, 902, 41071, 3665, 1499, 2144, 1203, 22922, 45818, 3785, 936, 41071, 1975, 4763, 523, 46050, 931, 879, 535, 941, 13608, 13765, 592, 779, 15565, 811, 41071, 1124, 779, 11019, 523, 18, 52739, 779, 13719, 14065, 1550, 875, 2203, 2158, 916, 4289, 800, 52847, 523, 19, 16, 28441, 535, 42749, 13719, 14350, 779, 21968, 4589, 20, 19, 16, 835, 535, 46050, 931, 835, 20, 19, 16, 879, 535, 46050, 931, 879, 20, 21 ], "markers": [ 55, 66, 74 ], "logits_raw": [ 1.1153349876403809, 0.07922947406768799, -2.398470163345337 ], "probs_raw": [ 0.7222212553024292, 0.25626859068870544, 0.021510181948542595 ], "probs": [ 0.6417179703712463, 0.3061330020427704, 0.052149027585983276 ], "argmax_index": 0, "argmax": "insufficient", "top2_margin": 0.33558496832847595, "near_tie": false, "gold": "A", "correct": false, "host_softmax_max_abs_dp": 2.9802322387695312e-08, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "insufficient", "confidence": 0.6417179703712463, "probabilities": { "insufficient": 0.6417179703712463, "A": 0.3061330020427704, "B": 0.052149027585983276 } } }, "usage": { "input_tokens": 82, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_56c419b12cd37cbe790c", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 74, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 74, "positions": 74, "bucket": 256, "ids": [ 1, 17, 1098, 21522, 16461, 521, 997, 1098, 21008, 3498, 25583, 1706, 938, 21706, 521, 1203, 10735, 1166, 4617, 523, 859, 25202, 896, 8294, 2927, 18, 9886, 894, 779, 4673, 535, 779, 21522, 30261, 774, 779, 21008, 1090, 50624, 8294, 523, 19, 16, 18197, 13735, 535, 941, 4589, 48308, 779, 11977, 20, 19, 16, 28441, 535, 941, 4589, 2177, 1014, 6122, 4021, 20, 19, 16, 7951, 535, 941, 4589, 48308, 779, 4673, 20, 21 ], "markers": [ 42, 53, 64 ], "logits_raw": [ 0.183104008436203, -0.15960577130317688, -3.6311588287353516 ], "probs_raw": [ 0.5774011611938477, 0.4098648726940155, 0.012733984738588333 ], "probs": [ 0.5410023927688599, 0.42352527379989624, 0.03547229990363121 ], "argmax_index": 0, "argmax": "contradicted", "top2_margin": 0.11747711896896362, "near_tie": false, "gold": "contradicted", "correct": true, "host_softmax_max_abs_dp": 2.9802322387695312e-08, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "contradicted", "confidence": 0.5410023927688599, "probabilities": { "contradicted": 0.5410023927688599, "insufficient": 0.42352527379989624, "supported": 0.03547229990363121 } } }, "usage": { "input_tokens": 74, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_579beed184ffc1104aa9", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 72, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 72, "positions": 72, "bucket": 256, "ids": [ 1, 17, 542, 5156, 13443, 6835, 2059, 10621, 523, 46050, 931, 835, 535, 941, 39037, 856, 12930, 1203, 779, 33407, 2312, 856, 26117, 523, 46050, 931, 879, 535, 9758, 779, 39037, 810, 33407, 2312, 938, 12930, 523, 18, 15534, 779, 5156, 13509, 13443, 523, 19, 16, 879, 535, 46050, 931, 879, 20, 19, 16, 835, 535, 46050, 931, 835, 20, 19, 16, 28441, 535, 42749, 13719, 14350, 779, 21968, 4589, 20, 21 ], "markers": [ 45, 53, 61 ], "logits_raw": [ 3.4579966068267822, 1.1308937072753906, 2.0269265174865723 ], "probs_raw": [ 0.7481496334075928, 0.07300297170877457, 0.17884740233421326 ], "probs": [ 0.6453805565834045, 0.12242553383111954, 0.2321939319372177 ], "argmax_index": 0, "argmax": "B", "top2_margin": 0.41318662464618683, "near_tie": false, "gold": "B", "correct": true, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "B", "confidence": 0.6453805565834045, "probabilities": { "B": 0.6453805565834045, "A": 0.12242553383111954, "insufficient": 0.2321939319372177 } } }, "usage": { "input_tokens": 72, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_587ced261c6092cd1325", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 77, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 77, "positions": 77, "bucket": 256, "ids": [ 1, 17, 55945, 931, 835, 535, 941, 13608, 21090, 1531, 902, 41071, 3665, 1499, 2144, 1203, 22922, 45818, 3785, 936, 41071, 1975, 4763, 523, 46050, 931, 879, 535, 941, 13608, 13765, 592, 779, 15565, 811, 41071, 1124, 779, 11019, 523, 18, 52739, 779, 13719, 3575, 4651, 41071, 3665, 523, 19, 16, 28441, 535, 42749, 13719, 14350, 779, 21968, 4589, 20, 19, 16, 879, 535, 46050, 931, 879, 20, 19, 16, 835, 535, 46050, 931, 835, 20, 21 ], "markers": [ 50, 61, 69 ], "logits_raw": [ 1.8004045486450195, -1.860771656036377, -0.186965674161911 ], "probs_raw": [ 0.8600244522094727, 0.022104576230049133, 0.11787096410989761 ], "probs": [ 0.7604895234107971, 0.05562574788928032, 0.18388469517230988 ], "argmax_index": 0, "argmax": "insufficient", "top2_margin": 0.5766048282384872, "near_tie": false, "gold": "B", "correct": false, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "insufficient", "confidence": 0.7604895234107971, "probabilities": { "insufficient": 0.7604895234107971, "B": 0.05562574788928032, "A": 0.18388469517230988 } } }, "usage": { "input_tokens": 77, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_599b3092b6fd18312d23", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 79, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 79, "positions": 79, "bucket": 256, "ids": [ 1, 17, 55945, 931, 835, 535, 29117, 800, 953, 30403, 810, 779, 5156, 3084, 3498, 15001, 6728, 797, 5323, 523, 46050, 931, 879, 535, 29117, 800, 953, 30403, 521, 1757, 22922, 47927, 523, 18, 15534, 779, 13719, 916, 6654, 6612, 32340, 15001, 536, 768, 28687, 32491, 1706, 902, 7045, 8825, 523, 19, 16, 879, 535, 46050, 931, 879, 20, 19, 16, 28441, 535, 42749, 13719, 14350, 779, 21968, 4589, 20, 19, 16, 835, 535, 46050, 931, 835, 20, 21 ], "markers": [ 52, 60, 71 ], "logits_raw": [ 3.349609851837158, 2.186221122741699, 2.6548283100128174 ], "probs_raw": [ 0.5519954562187195, 0.17245754599571228, 0.27554699778556824 ], "probs": [ 0.4891495704650879, 0.21306899189949036, 0.29778143763542175 ], "argmax_index": 0, "argmax": "B", "top2_margin": 0.19136813282966614, "near_tie": false, "gold": "A", "correct": false, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "B", "confidence": 0.4891495704650879, "probabilities": { "B": 0.4891495704650879, "insufficient": 0.21306899189949036, "A": 0.29778143763542175 } } }, "usage": { "input_tokens": 79, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_5a37d1dfcad20810a76a", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 83, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 83, "positions": 83, "bucket": 256, "ids": [ 1, 17, 1098, 4887, 11917, 810, 2254, 2059, 10547, 1186, 938, 3161, 810, 11500, 875, 779, 8324, 523, 18, 15246, 779, 8530, 5792, 521, 4021, 11500, 3161, 35908, 1400, 874, 2643, 791, 2630, 17150, 523, 3354, 779, 2254, 874, 2643, 791, 540, 19, 16, 33776, 535, 941, 9519, 7330, 14480, 1512, 779, 4272, 20, 19, 16, 20458, 535, 941, 9519, 7330, 22558, 779, 4272, 20, 19, 16, 28441, 535, 941, 21706, 2203, 2177, 1014, 27169, 3984, 779, 7330, 22558, 779, 4272, 20, 21 ], "markers": [ 43, 55, 66 ], "logits_raw": [ -0.7450401186943054, 2.489084005355835, 2.3256301879882812 ], "probs_raw": [ 0.02085919864475727, 0.5294926166534424, 0.4496481716632843 ], "probs": [ 0.049891456961631775, 0.5027567744255066, 0.4473517835140228 ], "argmax_index": 1, "argmax": "permitted", "top2_margin": 0.055404990911483765, "near_tie": false, "gold": "permitted", "correct": true, "host_softmax_max_abs_dp": 2.9802322387695312e-08, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "permitted", "confidence": 0.5027567744255066, "probabilities": { "prohibited": 0.049891456961631775, "permitted": 0.5027567744255066, "insufficient": 0.4473517835140228 } } }, "usage": { "input_tokens": 83, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_5d31b4bf2d0fc193f53b", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 86, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 86, "positions": 86, "bucket": 256, "ids": [ 1, 17, 1098, 49104, 7944, 3822, 884, 779, 2090, 1321, 1203, 14350, 1295, 11746, 3084, 875, 4021, 8688, 523, 18, 8475, 32144, 2524, 2090, 1321, 8544, 1403, 2059, 779, 19089, 16687, 810, 37610, 1126, 1052, 1381, 21706, 523, 3354, 2524, 2090, 1321, 874, 12043, 540, 19, 16, 20458, 535, 941, 9519, 7330, 22558, 779, 4272, 20, 19, 16, 28441, 535, 941, 21706, 2203, 2177, 1014, 27169, 3984, 779, 7330, 22558, 779, 4272, 20, 19, 16, 33776, 535, 941, 9519, 7330, 14480, 1512, 779, 4272, 20, 21 ], "markers": [ 46, 57, 74 ], "logits_raw": [ 1.4837925434112549, 2.482841968536377, -2.960299491882324 ], "probs_raw": [ 0.26828014850616455, 0.7285681962966919, 0.0031516929157078266 ], "probs": [ 0.32433220744132996, 0.6621069312095642, 0.013560839928686619 ], "argmax_index": 1, "argmax": "insufficient", "top2_margin": 0.33777472376823425, "near_tie": false, "gold": "insufficient", "correct": true, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "insufficient", "confidence": 0.6621069312095642, "probabilities": { "permitted": 0.32433220744132996, "insufficient": 0.6621069312095642, "prohibited": 0.013560839928686619 } } }, "usage": { "input_tokens": 86, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_6149a17bc154f9c5b4a0", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 86, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 86, "positions": 86, "bucket": 256, "ids": [ 1, 17, 1098, 1789, 5792, 44117, 60706, 43675, 9618, 8414, 25885, 521, 1203, 22922, 13748, 774, 1883, 5896, 1919, 1403, 779, 29114, 856, 60706, 523, 1470, 8414, 29114, 856, 60706, 1929, 1883, 11851, 936, 523, 3253, 1312, 3360, 856, 2251, 523, 18, 9886, 894, 779, 4673, 535, 779, 29114, 1178, 768, 15868, 523, 19, 16, 7951, 535, 941, 4589, 48308, 779, 4673, 20, 19, 16, 28441, 535, 941, 4589, 2177, 1014, 6122, 4021, 20, 19, 16, 18197, 13735, 535, 941, 4589, 48308, 779, 11977, 20, 21 ], "markers": [ 54, 64, 75 ], "logits_raw": [ -0.638434648513794, 1.2378493547439575, 1.305659532546997 ], "probs_raw": [ 0.0688871443271637, 0.4497777223587036, 0.4813351035118103 ], "probs": [ 0.1132504865527153, 0.4326384365558624, 0.4541110396385193 ], "argmax_index": 2, "argmax": "contradicted", "top2_margin": 0.02147260308265686, "near_tie": false, "gold": "supported", "correct": false, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "contradicted", "confidence": 0.4541110396385193, "probabilities": { "supported": 0.1132504865527153, "insufficient": 0.4326384365558624, "contradicted": 0.4541110396385193 } } }, "usage": { "input_tokens": 86, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_61ac20b07a5189231e15", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 85, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 85, "positions": 85, "bucket": 256, "ids": [ 1, 17, 1098, 6765, 22558, 10004, 2782, 768, 2984, 797, 902, 7477, 4611, 62248, 810, 4943, 876, 23369, 2796, 18109, 16711, 523, 18, 32135, 4611, 19494, 856, 6767, 1403, 779, 6765, 22558, 896, 1517, 523, 3354, 779, 2984, 4650, 797, 779, 4611, 62248, 540, 19, 16, 33776, 535, 941, 9519, 7330, 14480, 1512, 779, 4272, 20, 19, 16, 20458, 535, 941, 9519, 7330, 22558, 779, 4272, 20, 19, 16, 28441, 535, 941, 21706, 2203, 2177, 1014, 27169, 3984, 779, 7330, 22558, 779, 4272, 20, 21 ], "markers": [ 45, 57, 68 ], "logits_raw": [ 1.741729736328125, 2.7703092098236084, 1.8944990634918213 ], "probs_raw": [ 0.2015257179737091, 0.5636855363845825, 0.23478880524635315 ], "probs": [ 0.23807977139949799, 0.49638843536376953, 0.2655317783355713 ], "argmax_index": 1, "argmax": "permitted", "top2_margin": 0.23085665702819824, "near_tie": false, "gold": "permitted", "correct": true, "host_softmax_max_abs_dp": 1.4901161193847656e-08, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "permitted", "confidence": 0.49638843536376953, "probabilities": { "prohibited": 0.23807977139949799, "permitted": 0.49638843536376953, "insufficient": 0.2655317783355713 } } }, "usage": { "input_tokens": 85, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_64d63f86e328aff1481c", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 97, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 97, "positions": 97, "bucket": 256, "ids": [ 1, 17, 1098, 45609, 7330, 9938, 592, 38183, 811, 779, 797, 41722, 23120, 1403, 936, 1011, 2463, 779, 1336, 536, 1517, 779, 5329, 23120, 1550, 1403, 779, 797, 41722, 23120, 4584, 523, 941, 797, 41722, 23120, 4584, 2463, 1033, 12940, 536, 779, 5329, 23120, 1011, 523, 18, 9886, 894, 779, 4673, 535, 779, 45609, 7330, 9938, 592, 1033, 12940, 811, 779, 5329, 23120, 523, 19, 16, 18197, 13735, 535, 941, 4589, 48308, 779, 11977, 20, 19, 16, 28441, 535, 941, 4589, 2177, 1014, 6122, 4021, 20, 19, 16, 7951, 535, 941, 4589, 48308, 779, 4673, 20, 21 ], "markers": [ 65, 76, 87 ], "logits_raw": [ 1.448075771331787, 0.11779351532459259, 1.8871092796325684 ], "probs_raw": [ 0.3551628291606903, 0.0939059853553772, 0.5509311556816101 ], "probs": [ 0.36297598481178284, 0.1403396874666214, 0.49668434262275696 ], "argmax_index": 2, "argmax": "supported", "top2_margin": 0.13370835781097412, "near_tie": false, "gold": "supported", "correct": true, "host_softmax_max_abs_dp": 2.9802322387695312e-08, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "supported", "confidence": 0.49668434262275696, "probabilities": { "contradicted": 0.36297598481178284, "insufficient": 0.1403396874666214, "supported": 0.49668434262275696 } } }, "usage": { "input_tokens": 97, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_674ca4e9346216255466", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 66, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 66, "positions": 66, "bucket": 256, "ids": [ 1, 17, 1098, 10423, 57119, 12199, 19980, 521, 1203, 46050, 931, 835, 810, 46050, 931, 879, 3176, 1550, 2826, 6289, 2630, 57119, 933, 22537, 5742, 523, 18, 15534, 779, 13719, 8177, 4262, 811, 779, 19670, 10423, 57119, 523, 19, 16, 28441, 535, 42749, 13719, 14350, 779, 21968, 4589, 20, 19, 16, 879, 535, 46050, 931, 879, 20, 19, 16, 835, 535, 46050, 931, 835, 20, 21 ], "markers": [ 39, 50, 58 ], "logits_raw": [ 3.3179497718811035, 1.9383889436721802, 1.3750910758972168 ], "probs_raw": [ 0.7168546915054321, 0.1804244965314865, 0.10272078961133957 ], "probs": [ 0.6161913871765137, 0.23000140488147736, 0.15380728244781494 ], "argmax_index": 0, "argmax": "insufficient", "top2_margin": 0.3861899822950363, "near_tie": false, "gold": "insufficient", "correct": true, "host_softmax_max_abs_dp": 1.4901161193847656e-08, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "insufficient", "confidence": 0.6161913871765137, "probabilities": { "insufficient": 0.6161913871765137, "B": 0.23000140488147736, "A": 0.15380728244781494 } } }, "usage": { "input_tokens": 66, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_695dd61195ca025c4f4c", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 69, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 69, "positions": 69, "bucket": 256, "ids": [ 1, 17, 55945, 931, 835, 535, 941, 28356, 893, 44987, 896, 779, 52414, 1079, 953, 14465, 523, 46050, 931, 879, 535, 941, 28356, 893, 7637, 811, 5514, 768, 52414, 1079, 27968, 523, 18, 52739, 779, 13719, 13591, 768, 3866, 11746, 523, 19, 16, 879, 535, 46050, 931, 879, 20, 19, 16, 28441, 535, 42749, 13719, 14350, 779, 21968, 4589, 20, 19, 16, 835, 535, 46050, 931, 835, 20, 21 ], "markers": [ 42, 50, 61 ], "logits_raw": [ -0.545011043548584, -1.2503870725631714, -0.6067150831222534 ], "probs_raw": [ 0.41083216667175293, 0.2029193788766861, 0.3862484395503998 ], "probs": [ 0.3904627561569214, 0.2359113097190857, 0.3736259341239929 ], "argmax_index": 0, "argmax": "B", "top2_margin": 0.016836822032928467, "near_tie": true, "gold": "B", "correct": true, "host_softmax_max_abs_dp": 2.9802322387695312e-08, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "B", "confidence": 0.3904627561569214, "probabilities": { "B": 0.3904627561569214, "insufficient": 0.2359113097190857, "A": 0.3736259341239929 } } }, "usage": { "input_tokens": 69, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_6d64a8459f64eac7f4cc", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 94, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 94, "positions": 94, "bucket": 256, "ids": [ 1, 17, 1098, 45609, 7330, 9938, 592, 38183, 811, 779, 797, 41722, 23120, 1403, 936, 1011, 2463, 779, 1336, 536, 1517, 779, 5329, 23120, 1550, 1403, 779, 797, 41722, 23120, 4584, 523, 42749, 23120, 1090, 4978, 811, 2463, 1033, 12940, 856, 21706, 523, 18, 9886, 894, 779, 4673, 535, 779, 45609, 7330, 9938, 592, 1033, 12940, 811, 779, 5329, 23120, 523, 19, 16, 28441, 535, 941, 4589, 2177, 1014, 6122, 4021, 20, 19, 16, 18197, 13735, 535, 941, 4589, 48308, 779, 11977, 20, 19, 16, 7951, 535, 941, 4589, 48308, 779, 4673, 20, 21 ], "markers": [ 62, 73, 84 ], "logits_raw": [ 0.3540732264518738, 1.012641191482544, -0.8345171213150024 ], "probs_raw": [ 0.30895915627479553, 0.5969163775444031, 0.09412453323602676 ], "probs": [ 0.3301945626735687, 0.5285418629646301, 0.14126355946063995 ], "argmax_index": 1, "argmax": "contradicted", "top2_margin": 0.1983473002910614, "near_tie": false, "gold": "insufficient", "correct": false, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "contradicted", "confidence": 0.5285418629646301, "probabilities": { "insufficient": 0.3301945626735687, "contradicted": 0.5285418629646301, "supported": 0.14126355946063995 } } }, "usage": { "input_tokens": 94, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_6db17cc22bd656d78558", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 87, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 87, "positions": 87, "bucket": 256, "ids": [ 1, 17, 1098, 3084, 35283, 779, 26862, 53076, 810, 4462, 1856, 1203, 14350, 13834, 28553, 5742, 523, 18, 35516, 7494, 8544, 1403, 779, 28553, 856, 8112, 10274, 536, 797, 19980, 521, 779, 10423, 4462, 1856, 1391, 58340, 779, 26862, 53076, 523, 3354, 936, 874, 5789, 540, 19, 16, 33776, 535, 941, 9519, 7330, 14480, 1512, 779, 4272, 20, 19, 16, 20458, 535, 941, 9519, 7330, 22558, 779, 4272, 20, 19, 16, 28441, 535, 941, 21706, 2203, 2177, 1014, 27169, 3984, 779, 7330, 22558, 779, 4272, 20, 21 ], "markers": [ 47, 59, 70 ], "logits_raw": [ 1.8385608196258545, 2.026137590408325, 2.54236102104187 ], "probs_raw": [ 0.23653285205364227, 0.285334974527359, 0.4781322181224823 ], "probs": [ 0.2633902430534363, 0.3011564314365387, 0.4354533553123474 ], "argmax_index": 2, "argmax": "insufficient", "top2_margin": 0.13429692387580872, "near_tie": false, "gold": "insufficient", "correct": true, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "insufficient", "confidence": 0.4354533553123474, "probabilities": { "prohibited": 0.2633902430534363, "permitted": 0.3011564314365387, "insufficient": 0.4354533553123474 } } }, "usage": { "input_tokens": 87, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_712f6ffb763eef7c9a12", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 71, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 71, "positions": 71, "bucket": 256, "ids": [ 1, 17, 55945, 931, 835, 535, 941, 10046, 31494, 768, 18958, 875, 768, 4025, 17686, 1550, 811, 17575, 2723, 936, 523, 46050, 931, 879, 535, 941, 10046, 30261, 774, 26677, 768, 4025, 17686, 523, 18, 52739, 779, 13719, 12469, 768, 4025, 17686, 523, 19, 16, 879, 535, 46050, 931, 879, 20, 19, 16, 835, 535, 46050, 931, 835, 20, 19, 16, 28441, 535, 42749, 13719, 14350, 779, 21968, 4589, 20, 21 ], "markers": [ 44, 52, 60 ], "logits_raw": [ -0.45648908615112305, 0.1176786869764328, 0.13668951392173767 ], "probs_raw": [ 0.21808426082134247, 0.38724181056022644, 0.39467400312423706 ], "probs": [ 0.24785038828849792, 0.37352126836776733, 0.37862834334373474 ], "argmax_index": 2, "argmax": "insufficient", "top2_margin": 0.005107074975967407, "near_tie": true, "gold": "B", "correct": false, "host_softmax_max_abs_dp": 1.4901161193847656e-08, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "insufficient", "confidence": 0.37862834334373474, "probabilities": { "B": 0.24785038828849792, "A": 0.37352126836776733, "insufficient": 0.37862834334373474 } } }, "usage": { "input_tokens": 71, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_75b470586834df8c568d", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 67, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 67, "positions": 67, "bucket": 256, "ids": [ 1, 17, 1098, 21522, 52075, 779, 21008, 1203, 4852, 1295, 7082, 1337, 25583, 1706, 933, 779, 5776, 523, 18, 9886, 894, 779, 4673, 535, 779, 21522, 30261, 774, 779, 21008, 1090, 50624, 8294, 523, 19, 16, 7951, 535, 941, 4589, 48308, 779, 4673, 20, 19, 16, 18197, 13735, 535, 941, 4589, 48308, 779, 11977, 20, 19, 16, 28441, 535, 941, 4589, 2177, 1014, 6122, 4021, 20, 21 ], "markers": [ 35, 45, 56 ], "logits_raw": [ -4.853909492492676, 0.4020563066005707, 0.866544246673584 ], "probs_raw": [ 0.0020090413745492697, 0.3851463794708252, 0.6128445863723755 ], "probs": [ 0.009686571545898914, 0.4137553572654724, 0.5765581130981445 ], "argmax_index": 2, "argmax": "insufficient", "top2_margin": 0.16280275583267212, "near_tie": false, "gold": "insufficient", "correct": true, "host_softmax_max_abs_dp": 5.960464477539063e-08, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "insufficient", "confidence": 0.5765581130981445, "probabilities": { "supported": 0.009686571545898914, "contradicted": 0.4137553572654724, "insufficient": 0.5765581130981445 } } }, "usage": { "input_tokens": 67, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_76c7a79b1ff2363a60c5", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 62, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 62, "positions": 62, "bucket": 256, "ids": [ 1, 17, 1098, 59063, 3498, 19459, 42074, 788, 856, 4175, 523, 50432, 60468, 856, 22922, 49550, 523, 18, 9886, 894, 779, 4673, 535, 19459, 42074, 788, 856, 4175, 523, 19, 16, 7951, 535, 941, 4589, 48308, 779, 4673, 20, 19, 16, 28441, 535, 941, 4589, 2177, 1014, 6122, 4021, 20, 19, 16, 18197, 13735, 535, 941, 4589, 48308, 779, 11977, 20, 21 ], "markers": [ 30, 40, 51 ], "logits_raw": [ 3.787590980529785, 1.6854655742645264, -0.33135542273521423 ], "probs_raw": [ 0.878381073474884, 0.10733503848314285, 0.014283912256360054 ], "probs": [ 0.7839997410774231, 0.17464955151081085, 0.04135073721408844 ], "argmax_index": 0, "argmax": "supported", "top2_margin": 0.6093501895666122, "near_tie": false, "gold": "supported", "correct": true, "host_softmax_max_abs_dp": 1.4901161193847656e-08, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "supported", "confidence": 0.7839997410774231, "probabilities": { "supported": 0.7839997410774231, "insufficient": 0.17464955151081085, "contradicted": 0.04135073721408844 } } }, "usage": { "input_tokens": 62, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_774f7f9245f83c796d44", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 70, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 70, "positions": 70, "bucket": 256, "ids": [ 1, 17, 1098, 7576, 1090, 53859, 7330, 856, 21706, 521, 1203, 46050, 931, 835, 810, 46050, 931, 879, 3176, 1550, 1149, 6289, 810, 1295, 10562, 37078, 523, 18, 15534, 779, 13719, 12467, 6612, 3346, 17784, 10562, 1490, 779, 9519, 53859, 7330, 523, 19, 16, 879, 535, 46050, 931, 879, 20, 19, 16, 835, 535, 46050, 931, 835, 20, 19, 16, 28441, 535, 42749, 13719, 14350, 779, 21968, 4589, 20, 21 ], "markers": [ 43, 51, 59 ], "logits_raw": [ 2.467068910598755, 1.1629825830459595, 1.8013712167739868 ], "probs_raw": [ 0.5601188540458679, 0.1520276963710785, 0.2878533899784088 ], "probs": [ 0.4961564838886261, 0.1954556256532669, 0.3083878755569458 ], "argmax_index": 0, "argmax": "B", "top2_margin": 0.1877686083316803, "near_tie": false, "gold": "insufficient", "correct": false, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "B", "confidence": 0.4961564838886261, "probabilities": { "B": 0.4961564838886261, "A": 0.1954556256532669, "insufficient": 0.3083878755569458 } } }, "usage": { "input_tokens": 70, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_796d5c0da6eaadae6a3f", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 78, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 78, "positions": 78, "bucket": 256, "ids": [ 1, 17, 55945, 931, 835, 535, 941, 21009, 30403, 32340, 521, 1757, 22922, 46159, 779, 52847, 2076, 779, 3308, 523, 46050, 931, 879, 535, 941, 21009, 30403, 32340, 810, 779, 5156, 3084, 44987, 936, 6728, 797, 5323, 523, 18, 15534, 779, 13719, 62439, 902, 52847, 896, 953, 2919, 47927, 523, 19, 16, 879, 535, 46050, 931, 879, 20, 19, 16, 835, 535, 46050, 931, 835, 20, 19, 16, 28441, 535, 42749, 13719, 14350, 779, 21968, 4589, 20, 21 ], "markers": [ 51, 59, 67 ], "logits_raw": [ 3.8280439376831055, 1.9527827501296997, 1.7718702554702759 ], "probs_raw": [ 0.7804832458496094, 0.11965972185134888, 0.09985702484846115 ], "probs": [ 0.6701697111129761, 0.17555658519268036, 0.15427377820014954 ], "argmax_index": 0, "argmax": "B", "top2_margin": 0.4946131259202957, "near_tie": false, "gold": "A", "correct": false, "host_softmax_max_abs_dp": 1.4901161193847656e-08, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "B", "confidence": 0.6701697111129761, "probabilities": { "B": 0.6701697111129761, "A": 0.17555658519268036, "insufficient": 0.15427377820014954 } } }, "usage": { "input_tokens": 78, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_7ada97b6f99f4788e5c7", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 74, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 74, "positions": 74, "bucket": 256, "ids": [ 1, 17, 55945, 931, 835, 535, 941, 42271, 44987, 43152, 803, 768, 8825, 5771, 1203, 22922, 3498, 1295, 9170, 6177, 1178, 1381, 2047, 523, 46050, 931, 879, 535, 941, 42271, 44987, 896, 779, 8825, 953, 14708, 523, 18, 52739, 779, 13719, 15603, 902, 14708, 8825, 523, 19, 16, 879, 535, 46050, 931, 879, 20, 19, 16, 835, 535, 46050, 931, 835, 20, 19, 16, 28441, 535, 42749, 13719, 14350, 779, 21968, 4589, 20, 21 ], "markers": [ 47, 55, 63 ], "logits_raw": [ 0.2991751730442047, -2.6162571907043457, -0.412447452545166 ], "probs_raw": [ 0.6472375988960266, 0.03506771847605705, 0.31769469380378723 ], "probs": [ 0.579340934753418, 0.07218901813030243, 0.3484700620174408 ], "argmax_index": 0, "argmax": "B", "top2_margin": 0.23087087273597717, "near_tie": false, "gold": "B", "correct": true, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "B", "confidence": 0.579340934753418, "probabilities": { "B": 0.579340934753418, "A": 0.07218901813030243, "insufficient": 0.3484700620174408 } } }, "usage": { "input_tokens": 74, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_7ed75d854d99b3a81916", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 76, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 76, "positions": 76, "bucket": 256, "ids": [ 1, 17, 55945, 931, 835, 856, 16495, 810, 46050, 931, 879, 856, 779, 3981, 3321, 521, 1203, 13834, 13719, 1090, 4978, 811, 9936, 779, 3869, 856, 21706, 523, 18, 52739, 835, 19148, 835, 856, 3222, 536, 7056, 879, 1550, 1403, 835, 856, 12359, 810, 879, 856, 3222, 523, 19, 16, 28441, 535, 42749, 13719, 14350, 779, 21968, 4589, 20, 19, 16, 879, 535, 46050, 931, 879, 20, 19, 16, 835, 535, 46050, 931, 835, 20, 21 ], "markers": [ 49, 60, 68 ], "logits_raw": [ 2.1733789443969727, 1.3448199033737183, 1.0819072723388672 ], "probs_raw": [ 0.5642066597938538, 0.2463766634464264, 0.18941661715507507 ], "probs": [ 0.49705639481544495, 0.2750169038772583, 0.22792671620845795 ], "argmax_index": 0, "argmax": "insufficient", "top2_margin": 0.22203949093818665, "near_tie": false, "gold": "insufficient", "correct": true, "host_softmax_max_abs_dp": 1.4901161193847656e-08, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "insufficient", "confidence": 0.49705639481544495, "probabilities": { "insufficient": 0.49705639481544495, "B": 0.2750169038772583, "A": 0.22792671620845795 } } }, "usage": { "input_tokens": 76, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_823bf736b92e9d8fa66d", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 75, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 75, "positions": 75, "bucket": 256, "ids": [ 1, 17, 1098, 8170, 15791, 875, 4611, 521, 1203, 18724, 2203, 856, 1014, 21706, 523, 18, 1098, 1509, 9764, 833, 59534, 6835, 1550, 1509, 9764, 833, 4611, 523, 3354, 1033, 8170, 6984, 936, 540, 19, 16, 33776, 535, 941, 9519, 7330, 14480, 1512, 779, 4272, 20, 19, 16, 28441, 535, 941, 21706, 2203, 2177, 1014, 27169, 3984, 779, 7330, 22558, 779, 4272, 20, 19, 16, 20458, 535, 941, 9519, 7330, 22558, 779, 4272, 20, 21 ], "markers": [ 35, 47, 64 ], "logits_raw": [ -1.5380306243896484, 2.647414207458496, -0.08526135236024857 ], "probs_raw": [ 0.014084970578551292, 0.9257027506828308, 0.06021236255764961 ], "probs": [ 0.042185135185718536, 0.8387300372123718, 0.11908479779958725 ], "argmax_index": 1, "argmax": "insufficient", "top2_margin": 0.7196452394127846, "near_tie": false, "gold": "insufficient", "correct": true, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "insufficient", "confidence": 0.8387300372123718, "probabilities": { "prohibited": 0.042185135185718536, "insufficient": 0.8387300372123718, "permitted": 0.11908479779958725 } } }, "usage": { "input_tokens": 75, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_82c2cd97c160ad27951b", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 62, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 62, "positions": 62, "bucket": 256, "ids": [ 1, 17, 55945, 931, 835, 535, 941, 8449, 16057, 768, 60621, 523, 46050, 931, 879, 535, 941, 12153, 6602, 768, 2375, 3955, 523, 18, 52739, 779, 13719, 14065, 811, 12940, 779, 7605, 8449, 523, 19, 16, 28441, 535, 42749, 13719, 14350, 779, 21968, 4589, 20, 19, 16, 879, 535, 46050, 931, 879, 20, 19, 16, 835, 535, 46050, 931, 835, 20, 21 ], "markers": [ 35, 46, 54 ], "logits_raw": [ 0.8707722425460815, -2.1442296504974365, 0.39768466353416443 ], "probs_raw": [ 0.5980427265167236, 0.029331453144550323, 0.37262576818466187 ], "probs": [ 0.5466614961624146, 0.06344036757946014, 0.3898981213569641 ], "argmax_index": 0, "argmax": "insufficient", "top2_margin": 0.15676337480545044, "near_tie": false, "gold": "insufficient", "correct": true, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "insufficient", "confidence": 0.5466614961624146, "probabilities": { "insufficient": 0.5466614961624146, "B": 0.06344036757946014, "A": 0.3898981213569641 } } }, "usage": { "input_tokens": 62, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_834f3b3e10cc61b33d25", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 61, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 61, "positions": 61, "bucket": 256, "ids": [ 1, 17, 1098, 59063, 3498, 19459, 42074, 788, 856, 4175, 523, 50432, 60468, 856, 22922, 49550, 523, 18, 9886, 894, 779, 4673, 535, 19459, 60468, 856, 4175, 523, 19, 16, 7951, 535, 941, 4589, 48308, 779, 4673, 20, 19, 16, 28441, 535, 941, 4589, 2177, 1014, 6122, 4021, 20, 19, 16, 18197, 13735, 535, 941, 4589, 48308, 779, 11977, 20, 21 ], "markers": [ 29, 39, 50 ], "logits_raw": [ -2.509235382080078, 1.4758033752441406, 1.9639065265655518 ], "probs_raw": [ 0.00702153705060482, 0.37766996026039124, 0.6153085231781006 ], "probs": [ 0.023447468876838684, 0.4040045142173767, 0.572547972202301 ], "argmax_index": 2, "argmax": "contradicted", "top2_margin": 0.16854345798492432, "near_tie": false, "gold": "contradicted", "correct": true, "host_softmax_max_abs_dp": 1.862645149230957e-09, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "contradicted", "confidence": 0.572547972202301, "probabilities": { "supported": 0.023447468876838684, "insufficient": 0.4040045142173767, "contradicted": 0.572547972202301 } } }, "usage": { "input_tokens": 61, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_843185706adff0c60f8b", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 69, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 69, "positions": 69, "bucket": 256, "ids": [ 1, 17, 1098, 10449, 6835, 3956, 8649, 810, 768, 5196, 8170, 521, 1203, 779, 11746, 3084, 6602, 1295, 2203, 1337, 4021, 8688, 523, 18, 9886, 894, 779, 4673, 535, 779, 11860, 2090, 1321, 856, 5156, 523, 19, 16, 18197, 13735, 535, 941, 4589, 48308, 779, 11977, 20, 19, 16, 7951, 535, 941, 4589, 48308, 779, 4673, 20, 19, 16, 28441, 535, 941, 4589, 2177, 1014, 6122, 4021, 20, 21 ], "markers": [ 37, 48, 58 ], "logits_raw": [ 0.3888823986053467, -8.769230842590332, 1.5510286092758179 ], "probs_raw": [ 0.2382715493440628, 2.5104656742769293e-05, 0.761703372001648 ], "probs": [ 0.30347737669944763, 0.00043750001350417733, 0.6960850954055786 ], "argmax_index": 2, "argmax": "insufficient", "top2_margin": 0.392607718706131, "near_tie": false, "gold": "insufficient", "correct": true, "host_softmax_max_abs_dp": 2.9103830456733704e-11, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "insufficient", "confidence": 0.6960850954055786, "probabilities": { "contradicted": 0.30347737669944763, "supported": 0.00043750001350417733, "insufficient": 0.6960850954055786 } } }, "usage": { "input_tokens": 69, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_86a15618ba4cb90a307e", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 64, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 64, "positions": 64, "bucket": 256, "ids": [ 1, 17, 1098, 3059, 12930, 12874, 875, 45853, 779, 61026, 810, 22922, 3498, 1295, 45853, 1178, 1381, 5813, 523, 18, 9886, 894, 779, 4673, 535, 779, 61026, 1178, 1381, 45943, 523, 19, 16, 18197, 13735, 535, 941, 4589, 48308, 779, 11977, 20, 19, 16, 28441, 535, 941, 4589, 2177, 1014, 6122, 4021, 20, 19, 16, 7951, 535, 941, 4589, 48308, 779, 4673, 20, 21 ], "markers": [ 32, 43, 54 ], "logits_raw": [ 1.7928204536437988, 1.682106852531433, -5.491433143615723 ], "probs_raw": [ 0.5274591445922852, 0.4721788465976715, 0.0003619742637965828 ], "probs": [ 0.5182804465293884, 0.4788702130317688, 0.002849374897778034 ], "argmax_index": 0, "argmax": "contradicted", "top2_margin": 0.03941023349761963, "near_tie": false, "gold": "contradicted", "correct": true, "host_softmax_max_abs_dp": 2.3283064365386963e-10, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "contradicted", "confidence": 0.5182804465293884, "probabilities": { "contradicted": 0.5182804465293884, "insufficient": 0.4788702130317688, "supported": 0.002849374897778034 } } }, "usage": { "input_tokens": 64, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_8a4c3de28be95c105ca1", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 78, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 78, "positions": 78, "bucket": 256, "ids": [ 1, 17, 1098, 22773, 2396, 2240, 3525, 768, 6643, 59534, 1203, 4943, 876, 916, 4289, 592, 17784, 875, 3507, 3893, 523, 18, 31678, 3893, 856, 6767, 1550, 1720, 3507, 25252, 980, 17784, 523, 3354, 936, 6305, 540, 19, 16, 20458, 535, 941, 9519, 7330, 22558, 779, 4272, 20, 19, 16, 28441, 535, 941, 21706, 2203, 2177, 1014, 27169, 3984, 779, 7330, 22558, 779, 4272, 20, 19, 16, 33776, 535, 941, 9519, 7330, 14480, 1512, 779, 4272, 20, 21 ], "markers": [ 38, 49, 66 ], "logits_raw": [ -1.75178062915802, 2.2357535362243652, 0.6280560493469238 ], "probs_raw": [ 0.015214934945106506, 0.8204160332679749, 0.16436903178691864 ], "probs": [ 0.04213201254606247, 0.7272379398345947, 0.230630025267601 ], "argmax_index": 1, "argmax": "insufficient", "top2_margin": 0.4966079145669937, "near_tie": false, "gold": "prohibited", "correct": false, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "insufficient", "confidence": 0.7272379398345947, "probabilities": { "permitted": 0.04213201254606247, "insufficient": 0.7272379398345947, "prohibited": 0.230630025267601 } } }, "usage": { "input_tokens": 78, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_8c992e17e4fb54b044f5", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 68, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 68, "positions": 68, "bucket": 256, "ids": [ 1, 17, 55945, 931, 835, 535, 941, 28356, 893, 44987, 896, 779, 52414, 1079, 953, 14465, 523, 46050, 931, 879, 535, 941, 28356, 893, 7637, 811, 5514, 768, 52414, 1079, 27968, 523, 18, 52739, 779, 13719, 15603, 7850, 11746, 523, 19, 16, 879, 535, 46050, 931, 879, 20, 19, 16, 835, 535, 46050, 931, 835, 20, 19, 16, 28441, 535, 42749, 13719, 14350, 779, 21968, 4589, 20, 21 ], "markers": [ 41, 49, 57 ], "logits_raw": [ 1.3221626281738281, 0.9677733778953552, -2.26184344291687 ], "probs_raw": [ 0.5782465934753418, 0.40569886565208435, 0.016054581850767136 ], "probs": [ 0.5394794940948486, 0.41882413625717163, 0.041696369647979736 ], "argmax_index": 0, "argmax": "B", "top2_margin": 0.120655357837677, "near_tie": false, "gold": "A", "correct": false, "host_softmax_max_abs_dp": 3.725290298461914e-09, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "B", "confidence": 0.5394794940948486, "probabilities": { "B": 0.5394794940948486, "A": 0.41882413625717163, "insufficient": 0.041696369647979736 } } }, "usage": { "input_tokens": 68, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_8e8c3804a3c15ebb31e3", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 73, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 73, "positions": 73, "bucket": 256, "ids": [ 1, 17, 1098, 21522, 16461, 521, 997, 1098, 21008, 3498, 25583, 1706, 938, 21706, 523, 859, 11667, 535, 7448, 25583, 1706, 1166, 4617, 2013, 1727, 5776, 2927, 18, 9886, 894, 779, 4673, 535, 779, 21522, 38419, 779, 50624, 8294, 523, 19, 16, 28441, 535, 941, 4589, 2177, 1014, 6122, 4021, 20, 19, 16, 7951, 535, 941, 4589, 48308, 779, 4673, 20, 19, 16, 18197, 13735, 535, 941, 4589, 48308, 779, 11977, 20, 21 ], "markers": [ 41, 52, 62 ], "logits_raw": [ -0.09096686542034149, -3.5231728553771973, 1.6575649976730347 ], "probs_raw": [ 0.14752578735351562, 0.0047673797234892845, 0.8477067947387695 ], "probs": [ 0.21866793930530548, 0.018836576491594315, 0.7624955177307129 ], "argmax_index": 2, "argmax": "contradicted", "top2_margin": 0.5438275784254074, "near_tie": false, "gold": "contradicted", "correct": true, "host_softmax_max_abs_dp": 1.862645149230957e-09, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "contradicted", "confidence": 0.7624955177307129, "probabilities": { "insufficient": 0.21866793930530548, "supported": 0.018836576491594315, "contradicted": 0.7624955177307129 } } }, "usage": { "input_tokens": 73, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_90494c76078ffc3c717a", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 90, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 90, "positions": 90, "bucket": 256, "ids": [ 1, 17, 1098, 8414, 22654, 1178, 768, 12803, 13789, 810, 779, 28332, 44987, 1295, 41095, 953, 5231, 523, 18, 41914, 1587, 22879, 8414, 5799, 523, 45681, 936, 875, 768, 12803, 13789, 521, 6794, 896, 5231, 41095, 3469, 4201, 936, 797, 808, 16979, 523, 3354, 1033, 1789, 874, 11086, 540, 19, 16, 20458, 535, 941, 9519, 7330, 22558, 779, 4272, 20, 19, 16, 28441, 535, 941, 21706, 2203, 2177, 1014, 27169, 3984, 779, 7330, 22558, 779, 4272, 20, 19, 16, 33776, 535, 941, 9519, 7330, 14480, 1512, 779, 4272, 20, 21 ], "markers": [ 50, 61, 78 ], "logits_raw": [ 3.778726100921631, 1.7921221256256104, 0.7049117684364319 ], "probs_raw": [ 0.8450194001197815, 0.11590322852134705, 0.039077434688806534 ], "probs": [ 0.7389857172966003, 0.17878293991088867, 0.08223139494657516 ], "argmax_index": 0, "argmax": "permitted", "top2_margin": 0.5602027773857117, "near_tie": false, "gold": "permitted", "correct": true, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "permitted", "confidence": 0.7389857172966003, "probabilities": { "permitted": 0.7389857172966003, "insufficient": 0.17878293991088867, "prohibited": 0.08223139494657516 } } }, "usage": { "input_tokens": 90, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_92ec2ffa3dd46ada1335", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 91, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 91, "positions": 91, "bucket": 256, "ids": [ 1, 17, 1098, 1789, 5792, 44117, 60706, 43675, 9618, 8414, 25885, 521, 1203, 22922, 13748, 774, 1883, 5896, 1919, 1403, 779, 29114, 856, 60706, 523, 1470, 8414, 29114, 856, 60706, 1929, 1883, 11851, 936, 523, 3253, 1312, 3360, 856, 2251, 523, 18, 9886, 894, 779, 4673, 535, 1033, 29114, 4357, 10812, 875, 1789, 1490, 779, 9519, 5792, 523, 19, 16, 18197, 13735, 535, 941, 4589, 48308, 779, 11977, 20, 19, 16, 28441, 535, 941, 4589, 2177, 1014, 6122, 4021, 20, 19, 16, 7951, 535, 941, 4589, 48308, 779, 4673, 20, 21 ], "markers": [ 59, 70, 81 ], "logits_raw": [ 1.858945608139038, 1.736254096031189, 0.3094581365585327 ], "probs_raw": [ 0.4768959879875183, 0.4218319058418274, 0.1012720987200737 ], "probs": [ 0.4451003968715668, 0.40775102376937866, 0.14714854955673218 ], "argmax_index": 0, "argmax": "contradicted", "top2_margin": 0.03734937310218811, "near_tie": false, "gold": "contradicted", "correct": true, "host_softmax_max_abs_dp": 1.4901161193847656e-08, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "contradicted", "confidence": 0.4451003968715668, "probabilities": { "contradicted": 0.4451003968715668, "insufficient": 0.40775102376937866, "supported": 0.14714854955673218 } } }, "usage": { "input_tokens": 91, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_942ec65dac0bfb9874f9", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 62, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 62, "positions": 62, "bucket": 256, "ids": [ 1, 17, 1098, 2525, 7867, 8269, 768, 28623, 22820, 1203, 1295, 2565, 933, 27742, 5742, 523, 18, 9886, 894, 779, 4673, 535, 779, 16400, 31286, 1052, 1381, 25539, 523, 19, 16, 28441, 535, 941, 4589, 2177, 1014, 6122, 4021, 20, 19, 16, 7951, 535, 941, 4589, 48308, 779, 4673, 20, 19, 16, 18197, 13735, 535, 941, 4589, 48308, 779, 11977, 20, 21 ], "markers": [ 30, 41, 51 ], "logits_raw": [ 3.149710178375244, -5.195099830627441, -0.9152504205703735 ], "probs_raw": [ 0.9828963875770569, 0.00023356221208814532, 0.016870107501745224 ], "probs": [ 0.945720911026001, 0.0024374090135097504, 0.051841605454683304 ], "argmax_index": 0, "argmax": "insufficient", "top2_margin": 0.8938793055713177, "near_tie": false, "gold": "insufficient", "correct": true, "host_softmax_max_abs_dp": 3.725290298461914e-09, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "insufficient", "confidence": 0.945720911026001, "probabilities": { "insufficient": 0.945720911026001, "supported": 0.0024374090135097504, "contradicted": 0.051841605454683304 } } }, "usage": { "input_tokens": 62, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_9473603b44af63beab79", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 76, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 76, "positions": 76, "bucket": 256, "ids": [ 1, 17, 1098, 8170, 7850, 1509, 9764, 833, 4611, 1203, 1014, 11820, 37414, 9007, 4611, 523, 18, 1098, 1509, 9764, 833, 59534, 6835, 1550, 1509, 9764, 833, 4611, 523, 3354, 1033, 8170, 6984, 936, 540, 19, 16, 20458, 535, 941, 9519, 7330, 22558, 779, 4272, 20, 19, 16, 28441, 535, 941, 21706, 2203, 2177, 1014, 27169, 3984, 779, 7330, 22558, 779, 4272, 20, 19, 16, 33776, 535, 941, 9519, 7330, 14480, 1512, 779, 4272, 20, 21 ], "markers": [ 36, 47, 64 ], "logits_raw": [ 0.09622183442115784, 2.969804525375366, -1.9030358791351318 ], "probs_raw": [ 0.053090523928403854, 0.9397191405296326, 0.007190356031060219 ], "probs": [ 0.11075764894485474, 0.8626878261566162, 0.026554502546787262 ], "argmax_index": 1, "argmax": "insufficient", "top2_margin": 0.7519301772117615, "near_tie": false, "gold": "permitted", "correct": false, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "insufficient", "confidence": 0.8626878261566162, "probabilities": { "permitted": 0.11075764894485474, "insufficient": 0.8626878261566162, "prohibited": 0.026554502546787262 } } }, "usage": { "input_tokens": 76, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_960a4bcd27363e6db4b9", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 79, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 79, "positions": 79, "bucket": 256, "ids": [ 1, 17, 55945, 931, 835, 535, 835, 15001, 1534, 856, 8815, 1203, 1352, 4450, 810, 2919, 3807, 938, 10144, 523, 46050, 931, 879, 535, 941, 3308, 2154, 856, 21706, 2630, 15001, 3129, 523, 18, 15534, 779, 13719, 916, 6654, 6612, 32340, 15001, 536, 768, 28687, 32491, 1706, 902, 7045, 8825, 523, 19, 16, 879, 535, 46050, 931, 879, 20, 19, 16, 835, 535, 46050, 931, 835, 20, 19, 16, 28441, 535, 42749, 13719, 14350, 779, 21968, 4589, 20, 21 ], "markers": [ 52, 60, 68 ], "logits_raw": [ 2.380281925201416, 3.187776803970337, 1.8507490158081055 ], "probs_raw": [ 0.2610172927379608, 0.585274875164032, 0.15370787680149078 ], "probs": [ 0.2885650098323822, 0.5137538909912109, 0.19768109917640686 ], "argmax_index": 1, "argmax": "A", "top2_margin": 0.22518888115882874, "near_tie": false, "gold": "insufficient", "correct": false, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "A", "confidence": 0.5137538909912109, "probabilities": { "B": 0.2885650098323822, "A": 0.5137538909912109, "insufficient": 0.19768109917640686 } } }, "usage": { "input_tokens": 79, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_97a35499bd8717f3e786", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 92, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 92, "positions": 92, "bucket": 256, "ids": [ 1, 17, 1098, 2681, 28117, 1509, 17436, 1530, 793, 9544, 521, 1757, 22922, 15011, 936, 2076, 779, 5089, 523, 3253, 2919, 6177, 9647, 523, 18, 14875, 779, 2681, 28117, 1090, 10500, 15001, 6177, 536, 768, 28687, 32491, 1706, 779, 7045, 8825, 523, 3354, 779, 8153, 777, 7157, 779, 1530, 793, 540, 19, 16, 33776, 535, 941, 9519, 7330, 14480, 1512, 779, 4272, 20, 19, 16, 28441, 535, 941, 21706, 2203, 2177, 1014, 27169, 3984, 779, 7330, 22558, 779, 4272, 20, 19, 16, 20458, 535, 941, 9519, 7330, 22558, 779, 4272, 20, 21 ], "markers": [ 52, 64, 81 ], "logits_raw": [ 1.8441698551177979, 2.0381903648376465, 1.5088130235671997 ], "probs_raw": [ 0.34138965606689453, 0.41448840498924255, 0.24412190914154053 ], "probs": [ 0.34064170718193054, 0.39128151535987854, 0.26807668805122375 ], "argmax_index": 1, "argmax": "insufficient", "top2_margin": 0.050639808177948, "near_tie": false, "gold": "permitted", "correct": false, "host_softmax_max_abs_dp": 2.9802322387695312e-08, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "insufficient", "confidence": 0.39128151535987854, "probabilities": { "prohibited": 0.34064170718193054, "insufficient": 0.39128151535987854, "permitted": 0.26807668805122375 } } }, "usage": { "input_tokens": 92, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_97d2ed2591fc7638806c", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 86, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 86, "positions": 86, "bucket": 256, "ids": [ 1, 17, 1098, 1789, 5792, 44117, 60706, 43675, 9618, 8414, 25885, 1203, 13748, 774, 1883, 5896, 1919, 1403, 60706, 523, 941, 29114, 856, 60706, 521, 1203, 779, 3360, 810, 6771, 803, 1883, 5896, 938, 49033, 523, 18, 9886, 894, 779, 4673, 535, 1033, 29114, 4357, 10812, 875, 1789, 1490, 779, 9519, 5792, 523, 19, 16, 28441, 535, 941, 4589, 2177, 1014, 6122, 4021, 20, 19, 16, 7951, 535, 941, 4589, 48308, 779, 4673, 20, 19, 16, 18197, 13735, 535, 941, 4589, 48308, 779, 11977, 20, 21 ], "markers": [ 54, 65, 75 ], "logits_raw": [ 1.4517431259155273, -2.3006997108459473, 0.18750600516796112 ], "probs_raw": [ 0.765746533870697, 0.017964694648981094, 0.21628868579864502 ], "probs": [ 0.6784985661506653, 0.04649622365832329, 0.2750052809715271 ], "argmax_index": 0, "argmax": "insufficient", "top2_margin": 0.4034932851791382, "near_tie": false, "gold": "insufficient", "correct": true, "host_softmax_max_abs_dp": 2.9802322387695312e-08, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "insufficient", "confidence": 0.6784985661506653, "probabilities": { "insufficient": 0.6784985661506653, "supported": 0.04649622365832329, "contradicted": 0.2750052809715271 } } }, "usage": { "input_tokens": 86, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_980e76145c925747220c", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 98, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 98, "positions": 98, "bucket": 256, "ids": [ 1, 17, 1098, 26862, 53076, 7944, 779, 7494, 28553, 906, 10274, 523, 941, 10423, 4462, 1856, 7944, 936, 906, 946, 38008, 523, 9758, 7944, 938, 2835, 523, 18, 35516, 7494, 8544, 1403, 779, 28553, 856, 8112, 10274, 536, 1490, 1033, 8530, 13061, 521, 779, 26862, 53076, 1391, 58340, 779, 4462, 1856, 523, 3354, 936, 874, 5789, 540, 19, 16, 28441, 535, 941, 21706, 2203, 2177, 1014, 27169, 3984, 779, 7330, 22558, 779, 4272, 20, 19, 16, 33776, 535, 941, 9519, 7330, 14480, 1512, 779, 4272, 20, 19, 16, 20458, 535, 941, 9519, 7330, 22558, 779, 4272, 20, 21 ], "markers": [ 58, 75, 87 ], "logits_raw": [ 2.910322427749634, 2.614384651184082, 2.6769022941589355 ], "probs_raw": [ 0.3943755030632019, 0.2933498024940491, 0.3122747242450714 ], "probs": [ 0.37652453780174255, 0.30477800965309143, 0.3186974823474884 ], "argmax_index": 0, "argmax": "insufficient", "top2_margin": 0.05782705545425415, "near_tie": false, "gold": "permitted", "correct": false, "host_softmax_max_abs_dp": 2.9802322387695312e-08, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "insufficient", "confidence": 0.37652453780174255, "probabilities": { "insufficient": 0.37652453780174255, "prohibited": 0.30477800965309143, "permitted": 0.3186974823474884 } } }, "usage": { "input_tokens": 98, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_9ac5cc82c4126c91174a", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 89, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 89, "positions": 89, "bucket": 256, "ids": [ 1, 17, 1098, 8414, 22654, 856, 57077, 521, 1203, 3984, 41095, 953, 5231, 856, 1014, 2538, 523, 18, 41914, 1587, 22879, 8414, 5799, 523, 45681, 936, 875, 768, 12803, 13789, 521, 6794, 896, 5231, 41095, 3469, 4201, 936, 797, 808, 16979, 523, 3354, 1033, 1789, 874, 11086, 540, 19, 16, 20458, 535, 941, 9519, 7330, 22558, 779, 4272, 20, 19, 16, 28441, 535, 941, 21706, 2203, 2177, 1014, 27169, 3984, 779, 7330, 22558, 779, 4272, 20, 19, 16, 33776, 535, 941, 9519, 7330, 14480, 1512, 779, 4272, 20, 21 ], "markers": [ 49, 60, 77 ], "logits_raw": [ 1.1729261875152588, 2.151684284210205, 0.8149329423904419 ], "probs_raw": [ 0.22934584319591522, 0.6103235483169556, 0.16033059358596802 ], "probs": [ 0.2641015648841858, 0.5313907265663147, 0.2045077234506607 ], "argmax_index": 1, "argmax": "insufficient", "top2_margin": 0.2672891616821289, "near_tie": false, "gold": "insufficient", "correct": true, "host_softmax_max_abs_dp": 2.9802322387695312e-08, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "insufficient", "confidence": 0.5313907265663147, "probabilities": { "permitted": 0.2641015648841858, "insufficient": 0.5313907265663147, "prohibited": 0.2045077234506607 } } }, "usage": { "input_tokens": 89, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_a0128e551d095f3449f8", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 68, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 68, "positions": 68, "bucket": 256, "ids": [ 1, 17, 1098, 7576, 6602, 8376, 2314, 10562, 811, 51132, 16993, 18577, 536, 779, 3084, 44987, 1033, 8825, 856, 6612, 523, 18, 9886, 894, 779, 4673, 535, 8376, 1178, 10562, 811, 51132, 16993, 18577, 523, 19, 16, 7951, 535, 941, 4589, 48308, 779, 4673, 20, 19, 16, 28441, 535, 941, 4589, 2177, 1014, 6122, 4021, 20, 19, 16, 18197, 13735, 535, 941, 4589, 48308, 779, 11977, 20, 21 ], "markers": [ 36, 46, 57 ], "logits_raw": [ 6.080697059631348, 0.6055939793586731, -2.5855391025543213 ], "probs_raw": [ 0.9956568479537964, 0.004171600099653006, 0.0001715581165626645 ], "probs": [ 0.9784088134765625, 0.01958679035305977, 0.0020043274853378534 ], "argmax_index": 0, "argmax": "supported", "top2_margin": 0.9588220231235027, "near_tie": false, "gold": "supported", "correct": true, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "supported", "confidence": 0.9784088134765625, "probabilities": { "supported": 0.9784088134765625, "insufficient": 0.01958679035305977, "contradicted": 0.0020043274853378534 } } }, "usage": { "input_tokens": 68, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_a02fd582d2881cc462b2", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 85, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 85, "positions": 85, "bucket": 256, "ids": [ 1, 17, 1098, 33648, 1090, 37181, 6175, 13546, 779, 30122, 9041, 810, 4943, 876, 13748, 774, 779, 1232, 33388, 920, 2543, 10074, 523, 18, 542, 37181, 6175, 22558, 9544, 8544, 811, 779, 3508, 936, 13546, 523, 3354, 779, 33648, 7157, 779, 30122, 9041, 540, 19, 16, 33776, 535, 941, 9519, 7330, 14480, 1512, 779, 4272, 20, 19, 16, 28441, 535, 941, 21706, 2203, 2177, 1014, 27169, 3984, 779, 7330, 22558, 779, 4272, 20, 19, 16, 20458, 535, 941, 9519, 7330, 22558, 779, 4272, 20, 21 ], "markers": [ 45, 57, 74 ], "logits_raw": [ 2.111523151397705, 1.9371681213378906, 3.128459930419922 ], "probs_raw": [ 0.2171688675880432, 0.18242156505584717, 0.6004095673561096 ], "probs": [ 0.25312647223472595, 0.22348424792289734, 0.5233892798423767 ], "argmax_index": 2, "argmax": "permitted", "top2_margin": 0.27026280760765076, "near_tie": false, "gold": "permitted", "correct": true, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "permitted", "confidence": 0.5233892798423767, "probabilities": { "prohibited": 0.25312647223472595, "insufficient": 0.22348424792289734, "permitted": 0.5233892798423767 } } }, "usage": { "input_tokens": 85, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_a0a39ace7e5337d038a7", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 83, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 83, "positions": 83, "bucket": 256, "ids": [ 1, 17, 1098, 3863, 12153, 22558, 9365, 779, 6332, 2154, 1203, 4943, 876, 23369, 2796, 9365, 779, 11256, 8355, 523, 18, 18693, 902, 3863, 2853, 1550, 1403, 779, 12153, 3575, 6229, 9365, 896, 2853, 523, 3354, 779, 11256, 8355, 874, 7093, 540, 19, 16, 33776, 535, 941, 9519, 7330, 14480, 1512, 779, 4272, 20, 19, 16, 20458, 535, 941, 9519, 7330, 22558, 779, 4272, 20, 19, 16, 28441, 535, 941, 21706, 2203, 2177, 1014, 27169, 3984, 779, 7330, 22558, 779, 4272, 20, 21 ], "markers": [ 43, 55, 66 ], "logits_raw": [ 4.502979278564453, 2.016558885574341, 1.5348275899887085 ], "probs_raw": [ 0.881363570690155, 0.07333587110042572, 0.045300524681806564 ], "probs": [ 0.7756208181381226, 0.13130463659763336, 0.09307458251714706 ], "argmax_index": 0, "argmax": "prohibited", "top2_margin": 0.6443161815404892, "near_tie": false, "gold": "prohibited", "correct": true, "host_softmax_max_abs_dp": 1.4901161193847656e-08, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "prohibited", "confidence": 0.7756208181381226, "probabilities": { "prohibited": 0.7756208181381226, "permitted": 0.13130463659763336, "insufficient": 0.09307458251714706 } } }, "usage": { "input_tokens": 83, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_a22dab8ba073ff6cdb8d", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 94, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 94, "positions": 94, "bucket": 256, "ids": [ 1, 17, 49290, 5019, 11528, 6261, 875, 20155, 1230, 1798, 521, 6794, 896, 10144, 39208, 3469, 1134, 15419, 2563, 523, 46050, 931, 835, 535, 4411, 5019, 1230, 1047, 8896, 38120, 39273, 779, 19214, 523, 46050, 931, 879, 535, 4411, 5019, 1230, 1047, 8896, 38120, 1178, 779, 19214, 523, 9758, 1230, 1798, 938, 20155, 523, 18, 15534, 779, 38120, 46102, 1490, 779, 9519, 11287, 810, 32294, 523, 19, 16, 28441, 535, 42749, 13719, 14350, 779, 21968, 4589, 20, 19, 16, 879, 535, 46050, 931, 879, 20, 19, 16, 835, 535, 46050, 931, 835, 20, 21 ], "markers": [ 67, 78, 86 ], "logits_raw": [ 2.378375768661499, 0.7194362282752991, 1.3831477165222168 ], "probs_raw": [ 0.6410338878631592, 0.1220148503780365, 0.23695123195648193 ], "probs": [ 0.5565075874328613, 0.1701427549123764, 0.27334970235824585 ], "argmax_index": 0, "argmax": "insufficient", "top2_margin": 0.2831578850746155, "near_tie": false, "gold": "B", "correct": false, "host_softmax_max_abs_dp": 1.4901161193847656e-08, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "insufficient", "confidence": 0.5565075874328613, "probabilities": { "insufficient": 0.5565075874328613, "B": 0.1701427549123764, "A": 0.27334970235824585 } } }, "usage": { "input_tokens": 94, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_a3f18f3a63d45345942b", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 70, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 70, "positions": 70, "bucket": 256, "ids": [ 1, 17, 1098, 2525, 7867, 13055, 16400, 31286, 523, 941, 23120, 44987, 1186, 1052, 1014, 4763, 1381, 25539, 811, 779, 13608, 1090, 28623, 523, 18, 9886, 894, 779, 4673, 535, 779, 16400, 31286, 1052, 1381, 25539, 523, 19, 16, 7951, 535, 941, 4589, 48308, 779, 4673, 20, 19, 16, 28441, 535, 941, 4589, 2177, 1014, 6122, 4021, 20, 19, 16, 18197, 13735, 535, 941, 4589, 48308, 779, 11977, 20, 21 ], "markers": [ 38, 48, 59 ], "logits_raw": [ -10.35418701171875, 2.370882034301758, 1.7727797031402588 ], "probs_raw": [ 1.919910801007063e-06, 0.6452207565307617, 0.3547772765159607 ], "probs": [ 6.825674063293263e-05, 0.605174720287323, 0.394756942987442 ], "argmax_index": 1, "argmax": "insufficient", "top2_margin": 0.21041777729988098, "near_tie": false, "gold": "contradicted", "correct": false, "host_softmax_max_abs_dp": 2.9802322387695312e-08, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "insufficient", "confidence": 0.605174720287323, "probabilities": { "supported": 6.825674063293263e-05, "insufficient": 0.605174720287323, "contradicted": 0.394756942987442 } } }, "usage": { "input_tokens": 70, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_a9230edfc6ec9e524af0", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 75, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 75, "positions": 75, "bucket": 256, "ids": [ 1, 17, 542, 5156, 13509, 13443, 6835, 2059, 768, 39037, 810, 768, 33407, 2312, 523, 46050, 931, 835, 535, 9758, 10621, 938, 12930, 523, 46050, 931, 879, 535, 11005, 779, 39037, 856, 12930, 536, 779, 33407, 2312, 856, 26117, 523, 18, 15534, 779, 5156, 13509, 13443, 523, 19, 16, 879, 535, 46050, 931, 879, 20, 19, 16, 28441, 535, 42749, 13719, 14350, 779, 21968, 4589, 20, 19, 16, 835, 535, 46050, 931, 835, 20, 21 ], "markers": [ 48, 56, 67 ], "logits_raw": [ -0.8022832274436951, 0.6754550933837891, 0.012994378805160522 ], "probs_raw": [ 0.1308416724205017, 0.5734819173812866, 0.29567641019821167 ], "probs": [ 0.17655335366725922, 0.5073632001876831, 0.31608352065086365 ], "argmax_index": 1, "argmax": "insufficient", "top2_margin": 0.19127967953681946, "near_tie": false, "gold": "A", "correct": false, "host_softmax_max_abs_dp": 2.9802322387695312e-08, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "insufficient", "confidence": 0.5073632001876831, "probabilities": { "B": 0.17655335366725922, "insufficient": 0.5073632001876831, "A": 0.31608352065086365 } } }, "usage": { "input_tokens": 75, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_acacd8b46acd25cb954b", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 94, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 94, "positions": 94, "bucket": 256, "ids": [ 1, 17, 1098, 4887, 11917, 856, 11649, 58754, 521, 2158, 779, 2254, 856, 3161, 810, 11500, 523, 18, 21509, 779, 2254, 1550, 906, 768, 3981, 3321, 1403, 779, 4887, 11917, 856, 58754, 933, 4994, 3972, 1108, 536, 9141, 2643, 779, 4887, 11917, 523, 3354, 779, 2254, 874, 2643, 791, 1490, 1033, 5792, 540, 19, 16, 28441, 535, 941, 21706, 2203, 2177, 1014, 27169, 3984, 779, 7330, 22558, 779, 4272, 20, 19, 16, 20458, 535, 941, 9519, 7330, 22558, 779, 4272, 20, 19, 16, 33776, 535, 941, 9519, 7330, 14480, 1512, 779, 4272, 20, 21 ], "markers": [ 54, 71, 82 ], "logits_raw": [ 1.6015324592590332, 4.51754903793335, 1.0473811626434326 ], "probs_raw": [ 0.04989488422870636, 0.9214375615119934, 0.028667587786912918 ], "probs": [ 0.10307374596595764, 0.8275465965270996, 0.06937965005636215 ], "argmax_index": 1, "argmax": "permitted", "top2_margin": 0.724472850561142, "near_tie": false, "gold": "permitted", "correct": true, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "permitted", "confidence": 0.8275465965270996, "probabilities": { "insufficient": 0.10307374596595764, "permitted": 0.8275465965270996, "prohibited": 0.06937965005636215 } } }, "usage": { "input_tokens": 94, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_b45c63fb92729af27f05", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 83, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 83, "positions": 83, "bucket": 256, "ids": [ 1, 17, 3312, 17980, 779, 12153, 30403, 33300, 1450, 52518, 523, 2357, 22092, 1689, 22922, 46159, 896, 52847, 523, 941, 3084, 3498, 17688, 1486, 23496, 7045, 15001, 521, 810, 1295, 2919, 15001, 953, 2997, 523, 18, 9886, 894, 779, 4673, 535, 15001, 811, 19382, 779, 52518, 856, 6654, 6612, 523, 19, 16, 18197, 13735, 535, 941, 4589, 48308, 779, 11977, 20, 19, 16, 7951, 535, 941, 4589, 48308, 779, 4673, 20, 19, 16, 28441, 535, 941, 4589, 2177, 1014, 6122, 4021, 20, 21 ], "markers": [ 51, 62, 72 ], "logits_raw": [ 3.0344765186309814, -6.221322059631348, 1.7618658542633057 ], "probs_raw": [ 0.7811310291290283, 7.464173540938646e-05, 0.2187943458557129 ], "probs": [ 0.7121281027793884, 0.0009574248106218874, 0.2869144380092621 ], "argmax_index": 0, "argmax": "contradicted", "top2_margin": 0.42521366477012634, "near_tie": false, "gold": "contradicted", "correct": true, "host_softmax_max_abs_dp": 5.820766091346741e-11, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "contradicted", "confidence": 0.7121281027793884, "probabilities": { "contradicted": 0.7121281027793884, "supported": 0.0009574248106218874, "insufficient": 0.2869144380092621 } } }, "usage": { "input_tokens": 83, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_b49ecda75086d8c68d2a", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 61, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 61, "positions": 61, "bucket": 256, "ids": [ 1, 17, 55945, 931, 835, 535, 941, 28356, 893, 6602, 902, 4332, 523, 46050, 931, 879, 535, 941, 52414, 1079, 6034, 809, 814, 969, 523, 18, 52739, 779, 13719, 15603, 7850, 11746, 523, 19, 16, 879, 535, 46050, 931, 879, 20, 19, 16, 835, 535, 46050, 931, 835, 20, 19, 16, 28441, 535, 42749, 13719, 14350, 779, 21968, 4589, 20, 21 ], "markers": [ 34, 42, 50 ], "logits_raw": [ 1.6683297157287598, -2.664180040359497, -0.32061368227005005 ], "probs_raw": [ 0.8695845007896423, 0.01142159104347229, 0.11899389326572418 ], "probs": [ 0.7771173715591431, 0.035188402980566025, 0.1876942366361618 ], "argmax_index": 0, "argmax": "B", "top2_margin": 0.5894231349229813, "near_tie": false, "gold": "insufficient", "correct": false, "host_softmax_max_abs_dp": 1.4901161193847656e-08, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "B", "confidence": 0.7771173715591431, "probabilities": { "B": 0.7771173715591431, "A": 0.035188402980566025, "insufficient": 0.1876942366361618 } } }, "usage": { "input_tokens": 61, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_b531dd603bf080dab820", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 84, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 84, "positions": 84, "bucket": 256, "ids": [ 1, 17, 1098, 49104, 21706, 779, 19089, 16687, 1203, 22922, 2172, 1014, 7018, 779, 37610, 1126, 523, 18, 8475, 32144, 2524, 2090, 1321, 8544, 1403, 2059, 779, 19089, 16687, 810, 37610, 1126, 1052, 1381, 21706, 523, 3354, 2524, 2090, 1321, 874, 12043, 540, 19, 16, 33776, 535, 941, 9519, 7330, 14480, 1512, 779, 4272, 20, 19, 16, 20458, 535, 941, 9519, 7330, 22558, 779, 4272, 20, 19, 16, 28441, 535, 941, 21706, 2203, 2177, 1014, 27169, 3984, 779, 7330, 22558, 779, 4272, 20, 21 ], "markers": [ 44, 56, 67 ], "logits_raw": [ -0.07308841496706009, 2.8655457496643066, 3.741370916366577 ], "probs_raw": [ 0.015327489003539085, 0.2895366847515106, 0.695135772228241 ], "probs": [ 0.04096178337931633, 0.33422577381134033, 0.6248124837875366 ], "argmax_index": 2, "argmax": "insufficient", "top2_margin": 0.2905867099761963, "near_tie": false, "gold": "prohibited", "correct": false, "host_softmax_max_abs_dp": 2.9802322387695312e-08, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "insufficient", "confidence": 0.6248124837875366, "probabilities": { "prohibited": 0.04096178337931633, "permitted": 0.33422577381134033, "insufficient": 0.6248124837875366 } } }, "usage": { "input_tokens": 84, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_b5e6788d95b8395d2888", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 78, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 78, "positions": 78, "bucket": 256, "ids": [ 1, 17, 1098, 11860, 2090, 1321, 6835, 2059, 779, 3956, 8649, 810, 779, 5196, 8170, 523, 941, 11746, 3084, 44987, 779, 3956, 8649, 11423, 810, 22922, 3498, 779, 8170, 953, 1014, 14465, 523, 18, 9886, 894, 779, 4673, 535, 779, 3956, 8649, 1166, 14465, 523, 19, 16, 28441, 535, 941, 4589, 2177, 1014, 6122, 4021, 20, 19, 16, 18197, 13735, 535, 941, 4589, 48308, 779, 11977, 20, 19, 16, 7951, 535, 941, 4589, 48308, 779, 4673, 20, 21 ], "markers": [ 46, 57, 68 ], "logits_raw": [ 0.7756969332695007, 0.1709093302488327, -0.5731619596481323 ], "probs_raw": [ 0.5537936687469482, 0.30247679352760315, 0.1437295377254486 ], "probs": [ 0.49243271350860596, 0.3196847140789032, 0.18788257241249084 ], "argmax_index": 0, "argmax": "insufficient", "top2_margin": 0.17274799942970276, "near_tie": false, "gold": "supported", "correct": false, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "insufficient", "confidence": 0.49243271350860596, "probabilities": { "insufficient": 0.49243271350860596, "contradicted": 0.3196847140789032, "supported": 0.18788257241249084 } } }, "usage": { "input_tokens": 78, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_b7c4e62d07986286fce0", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 75, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 75, "positions": 75, "bucket": 256, "ids": [ 1, 17, 2797, 6765, 1090, 15001, 7627, 9647, 521, 1203, 779, 5582, 936, 6007, 938, 49033, 523, 18, 560, 965, 16711, 1550, 916, 779, 6765, 1090, 11599, 15001, 523, 3354, 16711, 874, 4997, 540, 19, 16, 28441, 535, 941, 21706, 2203, 2177, 1014, 27169, 3984, 779, 7330, 22558, 779, 4272, 20, 19, 16, 20458, 535, 941, 9519, 7330, 22558, 779, 4272, 20, 19, 16, 33776, 535, 941, 9519, 7330, 14480, 1512, 779, 4272, 20, 21 ], "markers": [ 35, 52, 63 ], "logits_raw": [ 2.9559473991394043, -0.06785887479782104, 0.48010173439979553 ], "probs_raw": [ 0.8828402757644653, 0.04292000085115433, 0.07423967868089676 ], "probs": [ 0.7776677012443542, 0.08968294411897659, 0.13264939188957214 ], "argmax_index": 0, "argmax": "insufficient", "top2_margin": 0.6450183093547821, "near_tie": false, "gold": "insufficient", "correct": true, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "insufficient", "confidence": 0.7776677012443542, "probabilities": { "insufficient": 0.7776677012443542, "permitted": 0.08968294411897659, "prohibited": 0.13264939188957214 } } }, "usage": { "input_tokens": 75, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_b8091be929b1e489fcc4", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 78, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 78, "positions": 78, "bucket": 256, "ids": [ 1, 17, 10720, 768, 25473, 813, 1914, 521, 768, 3338, 7061, 1165, 779, 12037, 803, 768, 33390, 7267, 523, 941, 3893, 6634, 4478, 1033, 856, 37640, 810, 1295, 2151, 7267, 953, 7935, 523, 18, 9886, 894, 779, 4673, 535, 768, 2151, 7267, 953, 7935, 523, 19, 16, 7951, 535, 941, 4589, 48308, 779, 4673, 20, 19, 16, 18197, 13735, 535, 941, 4589, 48308, 779, 11977, 20, 19, 16, 28441, 535, 941, 4589, 2177, 1014, 6122, 4021, 20, 21 ], "markers": [ 46, 56, 67 ], "logits_raw": [ -9.860912322998047, 4.097040176391602, 1.2136549949645996 ], "probs_raw": [ 8.212900866055861e-07, 0.9470182657241821, 0.05298099294304848 ], "probs": [ 4.146222636336461e-05, 0.8868883848190308, 0.11307018250226974 ], "argmax_index": 1, "argmax": "contradicted", "top2_margin": 0.773818202316761, "near_tie": false, "gold": "contradicted", "correct": true, "host_softmax_max_abs_dp": 7.450580596923828e-09, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "contradicted", "confidence": 0.8868883848190308, "probabilities": { "supported": 4.146222636336461e-05, "contradicted": 0.8868883848190308, "insufficient": 0.11307018250226974 } } }, "usage": { "input_tokens": 78, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_b863998d4be2721a5dfd", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 92, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 92, "positions": 92, "bucket": 256, "ids": [ 1, 17, 1098, 45609, 7330, 9938, 592, 38183, 811, 779, 797, 41722, 23120, 1403, 936, 1011, 2463, 779, 1336, 536, 1517, 779, 5329, 23120, 1550, 1403, 779, 797, 41722, 23120, 4584, 523, 9758, 29142, 10547, 1186, 1011, 2463, 1033, 12940, 523, 18, 9886, 894, 779, 4673, 535, 779, 45609, 7330, 9938, 592, 1033, 12940, 811, 779, 5329, 23120, 523, 19, 16, 7951, 535, 941, 4589, 48308, 779, 4673, 20, 19, 16, 28441, 535, 941, 4589, 2177, 1014, 6122, 4021, 20, 19, 16, 18197, 13735, 535, 941, 4589, 48308, 779, 11977, 20, 21 ], "markers": [ 60, 70, 81 ], "logits_raw": [ -0.9867814779281616, 0.012546658515930176, -0.5340939164161682 ], "probs_raw": [ 0.18907205760478973, 0.5136059522628784, 0.29732200503349304 ], "probs": [ 0.22605858743190765, 0.4615784287452698, 0.31236302852630615 ], "argmax_index": 1, "argmax": "insufficient", "top2_margin": 0.14921540021896362, "near_tie": false, "gold": "contradicted", "correct": false, "host_softmax_max_abs_dp": 1.4901161193847656e-08, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "insufficient", "confidence": 0.4615784287452698, "probabilities": { "supported": 0.22605858743190765, "insufficient": 0.4615784287452698, "contradicted": 0.31236302852630615 } } }, "usage": { "input_tokens": 92, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_b8c5b9b285cbdddc9f8d", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 78, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 78, "positions": 78, "bucket": 256, "ids": [ 1, 17, 55945, 931, 835, 856, 779, 16495, 57490, 1203, 856, 11649, 12359, 811, 9936, 1033, 3869, 523, 46050, 931, 879, 856, 779, 3981, 3321, 810, 856, 11649, 3222, 523, 18, 52739, 835, 19148, 835, 856, 3222, 536, 7056, 879, 1550, 1403, 835, 856, 12359, 810, 879, 856, 3222, 523, 19, 16, 835, 535, 46050, 931, 835, 20, 19, 16, 28441, 535, 42749, 13719, 14350, 779, 21968, 4589, 20, 19, 16, 879, 535, 46050, 931, 879, 20, 21 ], "markers": [ 51, 59, 70 ], "logits_raw": [ 0.475014865398407, 1.241917371749878, 2.9685797691345215 ], "probs_raw": [ 0.0655418112874031, 0.1411171853542328, 0.7933409810066223 ], "probs": [ 0.11538352072238922, 0.19955480098724365, 0.6850616335868835 ], "argmax_index": 2, "argmax": "B", "top2_margin": 0.4855068325996399, "near_tie": false, "gold": "B", "correct": true, "host_softmax_max_abs_dp": 7.450580596923828e-09, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "B", "confidence": 0.6850616335868835, "probabilities": { "A": 0.11538352072238922, "insufficient": 0.19955480098724365, "B": 0.6850616335868835 } } }, "usage": { "input_tokens": 78, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_b8e2750307df9761e716", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 81, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 81, "positions": 81, "bucket": 256, "ids": [ 1, 17, 1098, 49104, 21706, 2059, 779, 19089, 16687, 810, 779, 37610, 1126, 523, 18, 8475, 32144, 2524, 2090, 1321, 8544, 1403, 2059, 779, 19089, 16687, 810, 37610, 1126, 1052, 1381, 21706, 523, 3354, 2524, 2090, 1321, 874, 12043, 540, 19, 16, 20458, 535, 941, 9519, 7330, 22558, 779, 4272, 20, 19, 16, 33776, 535, 941, 9519, 7330, 14480, 1512, 779, 4272, 20, 19, 16, 28441, 535, 941, 21706, 2203, 2177, 1014, 27169, 3984, 779, 7330, 22558, 779, 4272, 20, 21 ], "markers": [ 41, 52, 64 ], "logits_raw": [ 6.782066822052002, -1.3062357902526855, 3.482503652572632 ], "probs_raw": [ 0.9641282558441162, 0.0002960940182674676, 0.03557564318180084 ], "probs": [ 0.9109132885932922, 0.0028198061045259237, 0.08626683801412582 ], "argmax_index": 0, "argmax": "permitted", "top2_margin": 0.8246464505791664, "near_tie": false, "gold": "permitted", "correct": true, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "permitted", "confidence": 0.9109132885932922, "probabilities": { "permitted": 0.9109132885932922, "prohibited": 0.0028198061045259237, "insufficient": 0.08626683801412582 } } }, "usage": { "input_tokens": 81, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_be2dd67bc733c1213bef", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 87, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 87, "positions": 87, "bucket": 256, "ids": [ 1, 17, 1098, 49104, 21706, 779, 19089, 16687, 1203, 22922, 2172, 1014, 7018, 779, 37610, 1126, 523, 18, 8475, 32144, 779, 14144, 58931, 1403, 779, 16687, 1178, 1381, 21706, 536, 779, 37610, 1126, 856, 1014, 4225, 875, 1033, 58931, 523, 3354, 1033, 58931, 874, 12043, 540, 19, 16, 33776, 535, 941, 9519, 7330, 14480, 1512, 779, 4272, 20, 19, 16, 20458, 535, 941, 9519, 7330, 22558, 779, 4272, 20, 19, 16, 28441, 535, 941, 21706, 2203, 2177, 1014, 27169, 3984, 779, 7330, 22558, 779, 4272, 20, 21 ], "markers": [ 47, 59, 70 ], "logits_raw": [ 0.8967389464378357, 1.5261001586914062, 3.427116632461548 ], "probs_raw": [ 0.06478925049304962, 0.12157130241394043, 0.8136394619941711 ], "probs": [ 0.11543203890323639, 0.18095681071281433, 0.7036111950874329 ], "argmax_index": 2, "argmax": "insufficient", "top2_margin": 0.5226543843746185, "near_tie": false, "gold": "permitted", "correct": false, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "insufficient", "confidence": 0.7036111950874329, "probabilities": { "prohibited": 0.11543203890323639, "permitted": 0.18095681071281433, "insufficient": 0.7036111950874329 } } }, "usage": { "input_tokens": 87, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_be866e1dbec5d42d9423", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 88, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 88, "positions": 88, "bucket": 256, "ids": [ 1, 17, 1098, 1789, 5792, 44117, 60706, 43675, 9618, 8414, 25885, 521, 1203, 13748, 774, 1883, 5896, 1919, 1403, 60706, 523, 1470, 8414, 29114, 1178, 768, 12803, 45836, 15868, 521, 916, 1295, 1883, 5896, 933, 1312, 32294, 523, 18, 9886, 894, 779, 4673, 535, 1033, 29114, 4357, 10812, 875, 1789, 1490, 779, 9519, 5792, 523, 19, 16, 18197, 13735, 535, 941, 4589, 48308, 779, 11977, 20, 19, 16, 28441, 535, 941, 4589, 2177, 1014, 6122, 4021, 20, 19, 16, 7951, 535, 941, 4589, 48308, 779, 4673, 20, 21 ], "markers": [ 56, 67, 78 ], "logits_raw": [ 1.425513744354248, 1.8368489742279053, 0.7128006219863892 ], "probs_raw": [ 0.3334285318851471, 0.5030873417854309, 0.16348405182361603 ], "probs": [ 0.33983734250068665, 0.455911785364151, 0.20425090193748474 ], "argmax_index": 1, "argmax": "insufficient", "top2_margin": 0.11607444286346436, "near_tie": false, "gold": "supported", "correct": false, "host_softmax_max_abs_dp": 2.9802322387695312e-08, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "insufficient", "confidence": 0.455911785364151, "probabilities": { "contradicted": 0.33983734250068665, "insufficient": 0.455911785364151, "supported": 0.20425090193748474 } } }, "usage": { "input_tokens": 88, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_c100e84585149a04be49", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 81, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 81, "positions": 81, "bucket": 256, "ids": [ 1, 17, 542, 5156, 13509, 13443, 6835, 2059, 768, 39037, 810, 768, 33407, 2312, 523, 46050, 931, 835, 535, 9758, 10621, 938, 12930, 523, 46050, 931, 879, 535, 11005, 779, 39037, 856, 12930, 536, 779, 33407, 2312, 856, 26117, 523, 18, 15534, 779, 13443, 8960, 779, 39037, 1203, 10144, 779, 33407, 2312, 523, 19, 16, 879, 535, 46050, 931, 879, 20, 19, 16, 835, 535, 46050, 931, 835, 20, 19, 16, 28441, 535, 42749, 13719, 14350, 779, 21968, 4589, 20, 21 ], "markers": [ 54, 62, 70 ], "logits_raw": [ -0.07913510501384735, -0.2963172197341919, 0.7918941378593445 ], "probs_raw": [ 0.23842720687389374, 0.19188226759433746, 0.5696905255317688 ], "probs": [ 0.2688654065132141, 0.23022787272930145, 0.5009067058563232 ], "argmax_index": 2, "argmax": "insufficient", "top2_margin": 0.23204129934310913, "near_tie": false, "gold": "B", "correct": false, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "insufficient", "confidence": 0.5009067058563232, "probabilities": { "B": 0.2688654065132141, "A": 0.23022787272930145, "insufficient": 0.5009067058563232 } } }, "usage": { "input_tokens": 81, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_c4e9bd463b6fde5be743", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 60, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 60, "positions": 60, "bucket": 256, "ids": [ 1, 17, 542, 5156, 13443, 6835, 2059, 10621, 523, 46050, 931, 835, 810, 46050, 931, 879, 3619, 1550, 13443, 20317, 810, 1295, 28261, 2203, 523, 18, 15534, 779, 5156, 13509, 13443, 523, 19, 16, 879, 535, 46050, 931, 879, 20, 19, 16, 28441, 535, 42749, 13719, 14350, 779, 21968, 4589, 20, 19, 16, 835, 535, 46050, 931, 835, 20, 21 ], "markers": [ 33, 41, 52 ], "logits_raw": [ 2.186413526535034, 1.998923420906067, 2.937082290649414 ], "probs_raw": [ 0.253327876329422, 0.21001826226711273, 0.5366538763046265 ], "probs": [ 0.27900218963623047, 0.24402931332588196, 0.47696855664253235 ], "argmax_index": 2, "argmax": "A", "top2_margin": 0.19796636700630188, "near_tie": false, "gold": "insufficient", "correct": false, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "A", "confidence": 0.47696855664253235, "probabilities": { "B": 0.27900218963623047, "insufficient": 0.24402931332588196, "A": 0.47696855664253235 } } }, "usage": { "input_tokens": 60, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_c6a411e7f2bbdf377173", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 92, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 92, "positions": 92, "bucket": 256, "ids": [ 1, 17, 1098, 7576, 6516, 903, 1653, 10562, 811, 51132, 16993, 18577, 810, 22922, 23369, 1339, 3759, 53859, 523, 903, 1653, 1757, 5689, 8376, 811, 51132, 1377, 523, 941, 7330, 4478, 902, 18135, 42183, 53859, 25274, 1295, 10562, 536, 8376, 1178, 1295, 1312, 8825, 523, 18, 9886, 894, 779, 4673, 535, 8376, 1178, 10562, 811, 51132, 16993, 18577, 523, 19, 16, 18197, 13735, 535, 941, 4589, 48308, 779, 11977, 20, 19, 16, 28441, 535, 941, 4589, 2177, 1014, 6122, 4021, 20, 19, 16, 7951, 535, 941, 4589, 48308, 779, 4673, 20, 21 ], "markers": [ 60, 71, 82 ], "logits_raw": [ 1.1703366041183472, 1.2382922172546387, -0.3189012110233307 ], "probs_raw": [ 0.4355661869049072, 0.4661942422389984, 0.09823955595493317 ], "probs": [ 0.41755813360214233, 0.438327819108963, 0.14411407709121704 ], "argmax_index": 1, "argmax": "insufficient", "top2_margin": 0.02076968550682068, "near_tie": false, "gold": "contradicted", "correct": false, "host_softmax_max_abs_dp": 1.4901161193847656e-08, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "insufficient", "confidence": 0.438327819108963, "probabilities": { "contradicted": 0.41755813360214233, "insufficient": 0.438327819108963, "supported": 0.14411407709121704 } } }, "usage": { "input_tokens": 92, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_c89b86e3f55123bfa506", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 68, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 68, "positions": 68, "bucket": 256, "ids": [ 1, 17, 1098, 6634, 11854, 768, 2151, 5854, 7267, 7935, 5329, 768, 4795, 810, 4943, 876, 3498, 779, 12037, 953, 1014, 37640, 523, 18, 9886, 894, 779, 4673, 535, 779, 7267, 12037, 953, 37640, 523, 19, 16, 7951, 535, 941, 4589, 48308, 779, 4673, 20, 19, 16, 18197, 13735, 535, 941, 4589, 48308, 779, 11977, 20, 19, 16, 28441, 535, 941, 4589, 2177, 1014, 6122, 4021, 20, 21 ], "markers": [ 36, 46, 57 ], "logits_raw": [ -8.718061447143555, 3.4809868335723877, 1.940075397491455 ], "probs_raw": [ 4.146997525822371e-06, 0.8235937356948853, 0.17640209197998047 ], "probs": [ 0.0001232229551533237, 0.7503045797348022, 0.24957218766212463 ], "argmax_index": 1, "argmax": "contradicted", "top2_margin": 0.5007323920726776, "near_tie": false, "gold": "contradicted", "correct": true, "host_softmax_max_abs_dp": 1.4551915228366852e-11, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "contradicted", "confidence": 0.7503045797348022, "probabilities": { "supported": 0.0001232229551533237, "contradicted": 0.7503045797348022, "insufficient": 0.24957218766212463 } } }, "usage": { "input_tokens": 68, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_caf773b23934fa7362a1", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 75, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 75, "positions": 75, "bucket": 256, "ids": [ 1, 17, 1098, 12153, 30403, 33300, 779, 52518, 884, 17980, 523, 941, 5156, 15001, 3084, 44987, 896, 1689, 1178, 1014, 47927, 936, 810, 936, 6728, 6612, 523, 18, 9886, 894, 779, 4673, 535, 15001, 811, 19382, 779, 52518, 856, 6654, 6612, 523, 19, 16, 7951, 535, 941, 4589, 48308, 779, 4673, 20, 19, 16, 18197, 13735, 535, 941, 4589, 48308, 779, 11977, 20, 19, 16, 28441, 535, 941, 4589, 2177, 1014, 6122, 4021, 20, 21 ], "markers": [ 43, 53, 64 ], "logits_raw": [ 3.9008922576904297, -2.43294620513916, 0.15340781211853027 ], "probs_raw": [ 0.9752746224403381, 0.0017313139978796244, 0.0229940302670002 ], "probs": [ 0.9262592196464539, 0.010040778666734695, 0.06370002031326294 ], "argmax_index": 0, "argmax": "supported", "top2_margin": 0.8625591993331909, "near_tie": false, "gold": "supported", "correct": true, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "supported", "confidence": 0.9262592196464539, "probabilities": { "supported": 0.9262592196464539, "contradicted": 0.010040778666734695, "insufficient": 0.06370002031326294 } } }, "usage": { "input_tokens": 75, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_cdf7f428857e1b841cb3", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 84, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 84, "positions": 84, "bucket": 256, "ids": [ 1, 17, 1098, 7330, 22558, 2314, 53859, 968, 779, 7576, 1203, 23369, 2796, 3759, 53859, 523, 46050, 931, 835, 535, 4411, 1550, 13289, 10562, 4010, 988, 879, 523, 46050, 931, 879, 535, 941, 7576, 5828, 15011, 1123, 3346, 17784, 10562, 523, 18, 15534, 779, 13719, 12467, 6612, 3346, 17784, 10562, 1490, 779, 9519, 53859, 7330, 523, 19, 16, 835, 535, 46050, 931, 835, 20, 19, 16, 28441, 535, 42749, 13719, 14350, 779, 21968, 4589, 20, 19, 16, 879, 535, 46050, 931, 879, 20, 21 ], "markers": [ 57, 65, 76 ], "logits_raw": [ 2.717790365219116, 1.3965704441070557, 2.6717827320098877 ], "probs_raw": [ 0.45007655024528503, 0.120084747672081, 0.42983871698379517 ], "probs": [ 0.4243011772632599, 0.16511568427085876, 0.41058313846588135 ], "argmax_index": 0, "argmax": "A", "top2_margin": 0.01371803879737854, "near_tie": true, "gold": "B", "correct": false, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "A", "confidence": 0.4243011772632599, "probabilities": { "A": 0.4243011772632599, "insufficient": 0.16511568427085876, "B": 0.41058313846588135 } } }, "usage": { "input_tokens": 84, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_cdfa9eed51fafcf09b72", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 82, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 82, "positions": 82, "bucket": 256, "ids": [ 1, 17, 1098, 16396, 5828, 3575, 6229, 779, 16701, 811, 51132, 41955, 523, 18, 1098, 5937, 11430, 1400, 51132, 41955, 1203, 1400, 1014, 40500, 896, 10562, 523, 11005, 30403, 2396, 2053, 1400, 51132, 523, 3354, 779, 16701, 51132, 768, 3346, 540, 19, 16, 28441, 535, 941, 21706, 2203, 2177, 1014, 27169, 3984, 779, 7330, 22558, 779, 4272, 20, 19, 16, 33776, 535, 941, 9519, 7330, 14480, 1512, 779, 4272, 20, 19, 16, 20458, 535, 941, 9519, 7330, 22558, 779, 4272, 20, 21 ], "markers": [ 42, 59, 71 ], "logits_raw": [ 1.991889238357544, 0.2586843967437744, 4.35775089263916 ], "probs_raw": [ 0.08453133702278137, 0.014938133768737316, 0.9005304574966431 ], "probs": [ 0.1490413397550583, 0.043212514370679855, 0.807746171951294 ], "argmax_index": 2, "argmax": "permitted", "top2_margin": 0.6587048321962357, "near_tie": false, "gold": "permitted", "correct": true, "host_softmax_max_abs_dp": 1.4901161193847656e-08, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "permitted", "confidence": 0.807746171951294, "probabilities": { "insufficient": 0.1490413397550583, "prohibited": 0.043212514370679855, "permitted": 0.807746171951294 } } }, "usage": { "input_tokens": 82, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_d0e36c4cb10a29b3191c", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 78, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 78, "positions": 78, "bucket": 256, "ids": [ 1, 17, 1098, 33648, 1178, 768, 37181, 6175, 1203, 1352, 8594, 3508, 938, 1014, 21706, 523, 18, 542, 37181, 6175, 22558, 9544, 8544, 811, 779, 3508, 936, 13546, 523, 3354, 779, 33648, 7157, 779, 30122, 9041, 540, 19, 16, 28441, 535, 941, 21706, 2203, 2177, 1014, 27169, 3984, 779, 7330, 22558, 779, 4272, 20, 19, 16, 20458, 535, 941, 9519, 7330, 22558, 779, 4272, 20, 19, 16, 33776, 535, 941, 9519, 7330, 14480, 1512, 779, 4272, 20, 21 ], "markers": [ 38, 55, 66 ], "logits_raw": [ 3.2459259033203125, 3.2828853130340576, 1.2678097486495972 ], "probs_raw": [ 0.45956292748451233, 0.47686588764190674, 0.06357114017009735 ], "probs": [ 0.44049862027168274, 0.4522833526134491, 0.10721801966428757 ], "argmax_index": 1, "argmax": "permitted", "top2_margin": 0.011784732341766357, "near_tie": true, "gold": "insufficient", "correct": false, "host_softmax_max_abs_dp": 2.9802322387695312e-08, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "permitted", "confidence": 0.4522833526134491, "probabilities": { "insufficient": 0.44049862027168274, "permitted": 0.4522833526134491, "prohibited": 0.10721801966428757 } } }, "usage": { "input_tokens": 78, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_d167e51bb1f499099645", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 76, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 76, "positions": 76, "bucket": 256, "ids": [ 1, 17, 1098, 10423, 57119, 12199, 19980, 523, 46050, 931, 835, 535, 941, 57119, 3498, 8177, 810, 22537, 3498, 1014, 8177, 523, 46050, 931, 879, 535, 941, 57119, 3498, 1014, 8177, 810, 22537, 3498, 8177, 523, 18, 15534, 779, 13719, 8177, 4262, 811, 779, 19670, 10423, 57119, 523, 19, 16, 835, 535, 46050, 931, 835, 20, 19, 16, 879, 535, 46050, 931, 879, 20, 19, 16, 28441, 535, 42749, 13719, 14350, 779, 21968, 4589, 20, 21 ], "markers": [ 49, 57, 65 ], "logits_raw": [ 0.7584027051925659, 0.9608060121536255, 1.1365561485290527 ], "probs_raw": [ 0.27144932746887207, 0.33234673738479614, 0.3962039053440094 ], "probs": [ 0.28854256868362427, 0.3334279954433441, 0.37802937626838684 ], "argmax_index": 2, "argmax": "insufficient", "top2_margin": 0.044601380825042725, "near_tie": false, "gold": "A", "correct": false, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "insufficient", "confidence": 0.37802937626838684, "probabilities": { "A": 0.28854256868362427, "B": 0.3334279954433441, "insufficient": 0.37802937626838684 } } }, "usage": { "input_tokens": 76, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_d41072dc35d8cd14c056", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 82, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 82, "positions": 82, "bucket": 256, "ids": [ 1, 17, 1098, 8170, 7850, 1509, 9764, 833, 4611, 1203, 1014, 11820, 37414, 9007, 4611, 523, 18, 15246, 779, 8530, 7330, 521, 779, 59534, 6835, 2059, 1509, 9764, 833, 810, 11820, 37414, 9007, 4611, 523, 3354, 1033, 8170, 6984, 936, 540, 19, 16, 20458, 535, 941, 9519, 7330, 22558, 779, 4272, 20, 19, 16, 33776, 535, 941, 9519, 7330, 14480, 1512, 779, 4272, 20, 19, 16, 28441, 535, 941, 21706, 2203, 2177, 1014, 27169, 3984, 779, 7330, 22558, 779, 4272, 20, 21 ], "markers": [ 42, 53, 65 ], "logits_raw": [ 0.1964302361011505, 0.45343083143234253, 1.8850579261779785 ], "probs_raw": [ 0.12978427112102509, 0.16781702637672424, 0.7023987770080566 ], "probs": [ 0.18042559921741486, 0.21678446233272552, 0.6027899384498596 ], "argmax_index": 2, "argmax": "insufficient", "top2_margin": 0.3860054761171341, "near_tie": false, "gold": "prohibited", "correct": false, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "insufficient", "confidence": 0.6027899384498596, "probabilities": { "permitted": 0.18042559921741486, "prohibited": 0.21678446233272552, "insufficient": 0.6027899384498596 } } }, "usage": { "input_tokens": 82, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_d51b05bfb9d6d7f09331", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 88, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 88, "positions": 88, "bucket": 256, "ids": [ 1, 17, 2797, 16701, 4117, 884, 779, 3893, 521, 1203, 1295, 17784, 25274, 933, 53859, 3129, 938, 4617, 523, 18, 1098, 5937, 11430, 1400, 51132, 41955, 1203, 1400, 1014, 40500, 896, 10562, 523, 11005, 30403, 2396, 2053, 1400, 51132, 523, 3354, 779, 16701, 51132, 768, 3346, 540, 19, 16, 33776, 535, 941, 9519, 7330, 14480, 1512, 779, 4272, 20, 19, 16, 28441, 535, 941, 21706, 2203, 2177, 1014, 27169, 3984, 779, 7330, 22558, 779, 4272, 20, 19, 16, 20458, 535, 941, 9519, 7330, 22558, 779, 4272, 20, 21 ], "markers": [ 48, 60, 77 ], "logits_raw": [ 1.244246482849121, 2.568499803543091, 2.9976813793182373 ], "probs_raw": [ 0.0949326828122139, 0.3568877577781677, 0.548179566860199 ], "probs": [ 0.14135213196277618, 0.3640234172344208, 0.49462443590164185 ], "argmax_index": 2, "argmax": "permitted", "top2_margin": 0.13060101866722107, "near_tie": false, "gold": "insufficient", "correct": false, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "permitted", "confidence": 0.49462443590164185, "probabilities": { "prohibited": 0.14135213196277618, "insufficient": 0.3640234172344208, "permitted": 0.49462443590164185 } } }, "usage": { "input_tokens": 88, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_d6de731cca52359b1c87", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 60, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 60, "positions": 60, "bucket": 256, "ids": [ 1, 17, 55945, 931, 835, 535, 941, 42271, 6602, 1352, 2154, 523, 46050, 931, 879, 535, 941, 5771, 1178, 768, 7392, 2080, 523, 18, 52739, 779, 13719, 15603, 902, 14708, 8825, 523, 19, 16, 28441, 535, 42749, 13719, 14350, 779, 21968, 4589, 20, 19, 16, 835, 535, 46050, 931, 835, 20, 19, 16, 879, 535, 46050, 931, 879, 20, 21 ], "markers": [ 33, 44, 52 ], "logits_raw": [ -0.32768940925598145, -2.296766757965088, -3.953437328338623 ], "probs_raw": [ 0.8574749231338501, 0.1196911409497261, 0.022833848372101784 ], "probs": [ 0.7575771808624268, 0.18558989465236664, 0.05683298781514168 ], "argmax_index": 0, "argmax": "insufficient", "top2_margin": 0.5719872862100601, "near_tie": false, "gold": "insufficient", "correct": true, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "insufficient", "confidence": 0.7575771808624268, "probabilities": { "insufficient": 0.7575771808624268, "A": 0.18558989465236664, "B": 0.05683298781514168 } } }, "usage": { "input_tokens": 60, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_d7e84e4245a29817da1f", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 75, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 75, "positions": 75, "bucket": 256, "ids": [ 1, 17, 55945, 931, 835, 535, 941, 12153, 21090, 811, 1039, 578, 786, 779, 7605, 8449, 2158, 10512, 1352, 6212, 523, 46050, 931, 879, 535, 941, 12153, 21090, 811, 7827, 779, 8449, 810, 6812, 768, 5156, 16400, 523, 18, 52739, 779, 13719, 14065, 875, 5156, 16400, 523, 19, 16, 28441, 535, 42749, 13719, 14350, 779, 21968, 4589, 20, 19, 16, 835, 535, 46050, 931, 835, 20, 19, 16, 879, 535, 46050, 931, 879, 20, 21 ], "markers": [ 48, 59, 67 ], "logits_raw": [ -1.3053580522537231, -1.3997042179107666, -3.440410614013672 ], "probs_raw": [ 0.4930465519428253, 0.44865643978118896, 0.058297015726566315 ], "probs": [ 0.4645945131778717, 0.4343149662017822, 0.10109055042266846 ], "argmax_index": 0, "argmax": "insufficient", "top2_margin": 0.030279546976089478, "near_tie": false, "gold": "B", "correct": false, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "insufficient", "confidence": 0.4645945131778717, "probabilities": { "insufficient": 0.4645945131778717, "A": 0.4343149662017822, "B": 0.10109055042266846 } } }, "usage": { "input_tokens": 75, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_d880e65fef5e2993fdd0", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 88, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 88, "positions": 88, "bucket": 256, "ids": [ 1, 17, 1098, 8414, 5799, 22654, 1178, 768, 12803, 13789, 810, 1236, 6034, 5231, 41095, 523, 9758, 11445, 938, 11649, 523, 18, 49290, 57077, 5799, 1919, 1288, 8414, 933, 10372, 916, 41095, 536, 13834, 4576, 1391, 58340, 779, 13789, 11287, 523, 3354, 1033, 1789, 874, 11086, 540, 19, 16, 20458, 535, 941, 9519, 7330, 22558, 779, 4272, 20, 19, 16, 28441, 535, 941, 21706, 2203, 2177, 1014, 27169, 3984, 779, 7330, 22558, 779, 4272, 20, 19, 16, 33776, 535, 941, 9519, 7330, 14480, 1512, 779, 4272, 20, 21 ], "markers": [ 48, 59, 76 ], "logits_raw": [ 1.5926122665405273, 1.731564998626709, 0.0559811033308506 ], "probs_raw": [ 0.4229806661605835, 0.4860343039035797, 0.09098505973815918 ], "probs": [ 0.4101724624633789, 0.45297500491142273, 0.13685256242752075 ], "argmax_index": 1, "argmax": "insufficient", "top2_margin": 0.04280254244804382, "near_tie": false, "gold": "permitted", "correct": false, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "insufficient", "confidence": 0.45297500491142273, "probabilities": { "permitted": 0.4101724624633789, "insufficient": 0.45297500491142273, "prohibited": 0.13685256242752075 } } }, "usage": { "input_tokens": 88, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_def9fe59d262357a673a", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 61, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 61, "positions": 61, "bucket": 256, "ids": [ 1, 17, 55945, 931, 835, 535, 941, 18779, 856, 3460, 523, 46050, 931, 879, 535, 941, 1927, 770, 856, 797, 768, 64098, 523, 18, 52739, 779, 13719, 15603, 902, 6023, 1927, 1725, 523, 19, 16, 28441, 535, 42749, 13719, 14350, 779, 21968, 4589, 20, 19, 16, 835, 535, 46050, 931, 835, 20, 19, 16, 879, 535, 46050, 931, 879, 20, 21 ], "markers": [ 34, 45, 53 ], "logits_raw": [ -0.8470367193222046, 1.5626904964447021, -4.1368608474731445 ], "probs_raw": [ 0.08218152821063995, 0.9147562980651855, 0.0030621157493442297 ], "probs": [ 0.1495330035686493, 0.83620685338974, 0.0142601877450943 ], "argmax_index": 1, "argmax": "A", "top2_margin": 0.6866738498210907, "near_tie": false, "gold": "insufficient", "correct": false, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "A", "confidence": 0.83620685338974, "probabilities": { "insufficient": 0.1495330035686493, "A": 0.83620685338974, "B": 0.0142601877450943 } } }, "usage": { "input_tokens": 61, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_e0c140e9222d1bce5647", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 76, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 76, "positions": 76, "bucket": 256, "ids": [ 1, 17, 1098, 21522, 16461, 521, 997, 1098, 21008, 3498, 25583, 1706, 938, 21706, 523, 859, 11667, 535, 7448, 25583, 1706, 1166, 4617, 2013, 1727, 5776, 2927, 18, 9886, 894, 779, 4673, 535, 779, 21522, 30261, 774, 779, 21008, 1090, 50624, 8294, 523, 19, 16, 7951, 535, 941, 4589, 48308, 779, 4673, 20, 19, 16, 18197, 13735, 535, 941, 4589, 48308, 779, 11977, 20, 19, 16, 28441, 535, 941, 4589, 2177, 1014, 6122, 4021, 20, 21 ], "markers": [ 44, 54, 65 ], "logits_raw": [ 4.24207878112793, -1.3785492181777954, 0.46038830280303955 ], "probs_raw": [ 0.9742728471755981, 0.003529172856360674, 0.022197969257831573 ], "probs": [ 0.9215284585952759, 0.01662665419280529, 0.06184489652514458 ], "argmax_index": 0, "argmax": "supported", "top2_margin": 0.8596835620701313, "near_tie": false, "gold": "supported", "correct": true, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "supported", "confidence": 0.9215284585952759, "probabilities": { "supported": 0.9215284585952759, "contradicted": 0.01662665419280529, "insufficient": 0.06184489652514458 } } }, "usage": { "input_tokens": 76, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_e1d610bd14e3d16c09b9", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 71, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 71, "positions": 71, "bucket": 256, "ids": [ 1, 17, 55945, 931, 835, 535, 941, 10046, 31494, 768, 18958, 875, 768, 4025, 17686, 1550, 811, 17575, 2723, 936, 523, 46050, 931, 879, 535, 941, 10046, 30261, 774, 26677, 768, 4025, 17686, 523, 18, 52739, 779, 13719, 36154, 768, 4025, 17686, 523, 19, 16, 835, 535, 46050, 931, 835, 20, 19, 16, 28441, 535, 42749, 13719, 14350, 779, 21968, 4589, 20, 19, 16, 879, 535, 46050, 931, 879, 20, 21 ], "markers": [ 44, 52, 63 ], "logits_raw": [ 0.24006280303001404, 0.42934364080429077, -0.9660718441009521 ], "probs_raw": [ 0.3987666666507721, 0.48186177015304565, 0.11937165260314941 ], "probs": [ 0.389519065618515, 0.4459127187728882, 0.16456826031208038 ], "argmax_index": 1, "argmax": "insufficient", "top2_margin": 0.05639365315437317, "near_tie": false, "gold": "A", "correct": false, "host_softmax_max_abs_dp": 2.9802322387695312e-08, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "insufficient", "confidence": 0.4459127187728882, "probabilities": { "A": 0.389519065618515, "insufficient": 0.4459127187728882, "B": 0.16456826031208038 } } }, "usage": { "input_tokens": 71, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_e332aa927d0c2ed044aa", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 93, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 93, "positions": 93, "bucket": 256, "ids": [ 1, 17, 1098, 7576, 6516, 903, 1653, 10562, 811, 51132, 16993, 18577, 810, 22922, 23369, 1339, 3759, 53859, 523, 903, 1653, 1757, 5689, 8376, 811, 51132, 1377, 523, 941, 7330, 4478, 902, 18135, 42183, 53859, 25274, 1295, 10562, 536, 8376, 1178, 1295, 1312, 8825, 523, 18, 9886, 894, 779, 4673, 535, 903, 1653, 1178, 10562, 811, 51132, 16993, 18577, 523, 19, 16, 18197, 13735, 535, 941, 4589, 48308, 779, 11977, 20, 19, 16, 28441, 535, 941, 4589, 2177, 1014, 6122, 4021, 20, 19, 16, 7951, 535, 941, 4589, 48308, 779, 4673, 20, 21 ], "markers": [ 61, 72, 83 ], "logits_raw": [ 0.6507751941680908, 0.9516758918762207, 1.8932214975357056 ], "probs_raw": [ 0.17196454107761383, 0.23233702778816223, 0.5956984758377075 ], "probs": [ 0.21418289840221405, 0.26554253697395325, 0.5202745199203491 ], "argmax_index": 2, "argmax": "supported", "top2_margin": 0.2547319829463959, "near_tie": false, "gold": "supported", "correct": true, "host_softmax_max_abs_dp": 1.4901161193847656e-08, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "supported", "confidence": 0.5202745199203491, "probabilities": { "contradicted": 0.21418289840221405, "insufficient": 0.26554253697395325, "supported": 0.5202745199203491 } } }, "usage": { "input_tokens": 93, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_e6155431ab7b0f8170d4", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 60, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 60, "positions": 60, "bucket": 256, "ids": [ 1, 17, 55945, 931, 835, 535, 859, 1011, 5514, 768, 52414, 1079, 27968, 523, 46050, 931, 879, 535, 941, 52414, 1079, 953, 14465, 523, 18, 52739, 779, 13719, 15603, 7850, 11746, 523, 19, 16, 28441, 535, 42749, 13719, 14350, 779, 21968, 4589, 20, 19, 16, 879, 535, 46050, 931, 879, 20, 19, 16, 835, 535, 46050, 931, 835, 20, 21 ], "markers": [ 33, 44, 52 ], "logits_raw": [ -0.8405516147613525, 0.2589983344078064, -0.3213144540786743 ], "probs_raw": [ 0.17594608664512634, 0.5283334255218506, 0.29572051763534546 ], "probs": [ 0.21540363132953644, 0.47246497869491577, 0.31213143467903137 ], "argmax_index": 1, "argmax": "B", "top2_margin": 0.1603335440158844, "near_tie": false, "gold": "B", "correct": true, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "B", "confidence": 0.47246497869491577, "probabilities": { "insufficient": 0.21540363132953644, "B": 0.47246497869491577, "A": 0.31213143467903137 } } }, "usage": { "input_tokens": 60, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_ea7ec66a5e8eac886548", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 80, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 80, "positions": 80, "bucket": 256, "ids": [ 1, 17, 1098, 22773, 2396, 2240, 3525, 768, 6643, 59534, 1203, 4943, 876, 916, 4289, 592, 17784, 875, 3507, 3893, 523, 18, 542, 6643, 1400, 874, 2047, 1403, 6643, 3893, 856, 12043, 523, 3354, 779, 6643, 874, 2047, 540, 19, 16, 20458, 535, 941, 9519, 7330, 22558, 779, 4272, 20, 19, 16, 28441, 535, 941, 21706, 2203, 2177, 1014, 27169, 3984, 779, 7330, 22558, 779, 4272, 20, 19, 16, 33776, 535, 941, 9519, 7330, 14480, 1512, 779, 4272, 20, 21 ], "markers": [ 40, 51, 68 ], "logits_raw": [ -0.49817439913749695, 1.6548174619674683, 1.9368813037872314 ], "probs_raw": [ 0.04755781963467598, 0.40950050950050354, 0.542941689491272 ], "probs": [ 0.08811099082231522, 0.410165399312973, 0.5017235279083252 ], "argmax_index": 2, "argmax": "prohibited", "top2_margin": 0.09155812859535217, "near_tie": false, "gold": "permitted", "correct": false, "host_softmax_max_abs_dp": 2.9802322387695312e-08, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "prohibited", "confidence": 0.5017235279083252, "probabilities": { "permitted": 0.08811099082231522, "insufficient": 0.410165399312973, "prohibited": 0.5017235279083252 } } }, "usage": { "input_tokens": 80, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_eac5121a4a1bf391a08c", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 58, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 58, "positions": 58, "bucket": 256, "ids": [ 1, 17, 1098, 59063, 15994, 2059, 55501, 1203, 4654, 1512, 1149, 7845, 5742, 523, 18, 9886, 894, 779, 4673, 535, 19459, 60468, 856, 4175, 523, 19, 16, 18197, 13735, 535, 941, 4589, 48308, 779, 11977, 20, 19, 16, 28441, 535, 941, 4589, 2177, 1014, 6122, 4021, 20, 19, 16, 7951, 535, 941, 4589, 48308, 779, 4673, 20, 21 ], "markers": [ 26, 37, 48 ], "logits_raw": [ 2.0061943531036377, 3.7291605472564697, -3.016935348510742 ], "probs_raw": [ 0.15133854746818542, 0.8476650714874268, 0.0009963964112102985 ], "probs": [ 0.2246410846710205, 0.7691481113433838, 0.006210819352418184 ], "argmax_index": 1, "argmax": "insufficient", "top2_margin": 0.5445070266723633, "near_tie": false, "gold": "insufficient", "correct": true, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "insufficient", "confidence": 0.7691481113433838, "probabilities": { "contradicted": 0.2246410846710205, "insufficient": 0.7691481113433838, "supported": 0.006210819352418184 } } }, "usage": { "input_tokens": 58, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_eafc22c8c40df3932a8e", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 77, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 77, "positions": 77, "bucket": 256, "ids": [ 1, 17, 1098, 13061, 6602, 779, 30174, 17150, 521, 1203, 779, 2835, 30174, 810, 26862, 2811, 952, 21148, 875, 779, 2391, 931, 938, 49033, 523, 18, 9886, 894, 779, 4673, 535, 1490, 779, 21706, 13061, 521, 779, 2391, 931, 856, 6654, 797, 8784, 523, 19, 16, 7951, 535, 941, 4589, 48308, 779, 4673, 20, 19, 16, 18197, 13735, 535, 941, 4589, 48308, 779, 11977, 20, 19, 16, 28441, 535, 941, 4589, 2177, 1014, 6122, 4021, 20, 21 ], "markers": [ 45, 55, 66 ], "logits_raw": [ -2.8015472888946533, 0.15241731703281403, 2.426419258117676 ], "probs_raw": [ 0.004840384237468243, 0.09284752607345581, 0.9023120999336243 ], "probs": [ 0.01956266537308693, 0.16137823462486267, 0.8190590739250183 ], "argmax_index": 2, "argmax": "insufficient", "top2_margin": 0.6576808393001556, "near_tie": false, "gold": "insufficient", "correct": true, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "insufficient", "confidence": 0.8190590739250183, "probabilities": { "supported": 0.01956266537308693, "contradicted": 0.16137823462486267, "insufficient": 0.8190590739250183 } } }, "usage": { "input_tokens": 77, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_ebd74aa05324a1556288", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 77, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 77, "positions": 77, "bucket": 256, "ids": [ 1, 17, 55945, 931, 835, 535, 941, 12153, 21090, 811, 1039, 578, 786, 779, 7605, 8449, 2158, 10512, 1352, 6212, 523, 46050, 931, 879, 535, 941, 12153, 21090, 811, 7827, 779, 8449, 810, 6812, 768, 5156, 16400, 523, 18, 52739, 779, 13719, 14065, 811, 12940, 779, 7605, 8449, 523, 19, 16, 28441, 535, 42749, 13719, 14350, 779, 21968, 4589, 20, 19, 16, 835, 535, 46050, 931, 835, 20, 19, 16, 879, 535, 46050, 931, 879, 20, 21 ], "markers": [ 50, 61, 69 ], "logits_raw": [ -1.8500902652740479, -1.6692605018615723, -4.712647914886475 ], "probs_raw": [ 0.44339337944984436, 0.5312789082527161, 0.02532772719860077 ], "probs": [ 0.4410553276538849, 0.5018713474273682, 0.05707336589694023 ], "argmax_index": 1, "argmax": "A", "top2_margin": 0.060816019773483276, "near_tie": false, "gold": "A", "correct": true, "host_softmax_max_abs_dp": 2.9802322387695312e-08, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "A", "confidence": 0.5018713474273682, "probabilities": { "insufficient": 0.4410553276538849, "A": 0.5018713474273682, "B": 0.05707336589694023 } } }, "usage": { "input_tokens": 77, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_ebea4681e3dcc7d92ca2", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 87, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 87, "positions": 87, "bucket": 256, "ids": [ 1, 17, 55945, 931, 835, 856, 779, 16495, 57490, 810, 856, 11649, 3222, 811, 9936, 1033, 3869, 523, 46050, 931, 879, 856, 779, 3981, 3321, 57490, 810, 856, 1236, 3222, 811, 9936, 936, 523, 18, 15246, 779, 8530, 17150, 521, 7056, 879, 19148, 879, 856, 3222, 536, 7056, 835, 1550, 1403, 879, 856, 12359, 810, 835, 856, 3222, 523, 19, 16, 28441, 535, 42749, 13719, 14350, 779, 21968, 4589, 20, 19, 16, 879, 535, 46050, 931, 879, 20, 19, 16, 835, 535, 46050, 931, 835, 20, 21 ], "markers": [ 60, 71, 79 ], "logits_raw": [ 2.0848593711853027, 1.2152581214904785, 2.261540174484253 ], "probs_raw": [ 0.3827943205833435, 0.1604362428188324, 0.4567693769931793 ], "probs": [ 0.37427595257759094, 0.2011004239320755, 0.42462357878685 ], "argmax_index": 2, "argmax": "A", "top2_margin": 0.05034762620925903, "near_tie": false, "gold": "B", "correct": false, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "A", "confidence": 0.42462357878685, "probabilities": { "insufficient": 0.37427595257759094, "B": 0.2011004239320755, "A": 0.42462357878685 } } }, "usage": { "input_tokens": 87, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_ec94ef909f7918695ab3", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 67, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 67, "positions": 67, "bucket": 256, "ids": [ 1, 17, 1098, 40348, 3575, 5405, 592, 779, 5562, 811, 4963, 810, 4943, 876, 22879, 592, 779, 7045, 5105, 5562, 523, 18, 9886, 894, 779, 4673, 535, 779, 35101, 5562, 856, 797, 5105, 523, 19, 16, 7951, 535, 941, 4589, 48308, 779, 4673, 20, 19, 16, 28441, 535, 941, 4589, 2177, 1014, 6122, 4021, 20, 19, 16, 18197, 13735, 535, 941, 4589, 48308, 779, 11977, 20, 21 ], "markers": [ 35, 45, 56 ], "logits_raw": [ -11.032297134399414, 0.18279242515563965, -0.3950861096382141 ], "probs_raw": [ 8.628144314570818e-06, 0.6405735611915588, 0.35941776633262634 ], "probs": [ 0.0001995474158320576, 0.6016390919685364, 0.3981613516807556 ], "argmax_index": 1, "argmax": "insufficient", "top2_margin": 0.20347774028778076, "near_tie": false, "gold": "contradicted", "correct": false, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "insufficient", "confidence": 0.6016390919685364, "probabilities": { "supported": 0.0001995474158320576, "insufficient": 0.6016390919685364, "contradicted": 0.3981613516807556 } } }, "usage": { "input_tokens": 67, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_edb708bd0fe5fd67d6ee", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 99, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 99, "positions": 99, "bucket": 256, "ids": [ 1, 17, 1098, 2681, 28117, 1509, 6767, 768, 8153, 777, 811, 7157, 779, 1530, 793, 523, 13629, 779, 5089, 1689, 22922, 46159, 896, 15001, 523, 3253, 2919, 6177, 9647, 523, 18, 3111, 1033, 37640, 13061, 521, 779, 1509, 8112, 6177, 856, 13827, 810, 2919, 17688, 1486, 938, 21867, 523, 3354, 779, 8153, 777, 7157, 779, 1530, 793, 540, 19, 16, 20458, 535, 941, 9519, 7330, 22558, 779, 4272, 20, 19, 16, 33776, 535, 941, 9519, 7330, 14480, 1512, 779, 4272, 20, 19, 16, 28441, 535, 941, 21706, 2203, 2177, 1014, 27169, 3984, 779, 7330, 22558, 779, 4272, 20, 21 ], "markers": [ 59, 70, 82 ], "logits_raw": [ 0.6859942078590393, 0.7516666054725647, 2.53814697265625 ], "probs_raw": [ 0.11846372485160828, 0.12650465965270996, 0.7550315856933594 ], "probs": [ 0.17232663929462433, 0.1806035041809082, 0.6470698714256287 ], "argmax_index": 2, "argmax": "insufficient", "top2_margin": 0.46646636724472046, "near_tie": false, "gold": "permitted", "correct": false, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "insufficient", "confidence": 0.6470698714256287, "probabilities": { "permitted": 0.17232663929462433, "prohibited": 0.1806035041809082, "insufficient": 0.6470698714256287 } } }, "usage": { "input_tokens": 99, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_eeae283841ce75dd25bf", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 83, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 83, "positions": 83, "bucket": 256, "ids": [ 1, 17, 1098, 3863, 12153, 22558, 9365, 779, 6332, 2154, 1203, 4943, 876, 23369, 2796, 9365, 779, 11256, 8355, 523, 18, 18693, 902, 3863, 2853, 1550, 1403, 779, 12153, 3575, 6229, 9365, 896, 2853, 523, 3354, 779, 6332, 2154, 874, 7093, 540, 19, 16, 28441, 535, 941, 21706, 2203, 2177, 1014, 27169, 3984, 779, 7330, 22558, 779, 4272, 20, 19, 16, 20458, 535, 941, 9519, 7330, 22558, 779, 4272, 20, 19, 16, 33776, 535, 941, 9519, 7330, 14480, 1512, 779, 4272, 20, 21 ], "markers": [ 43, 60, 71 ], "logits_raw": [ 1.9277441501617432, 1.6321382522583008, 3.0559921264648438 ], "probs_raw": [ 0.20685440301895142, 0.1539163440465927, 0.6392292380332947 ], "probs": [ 0.24700720608234406, 0.199987530708313, 0.5530052781105042 ], "argmax_index": 2, "argmax": "prohibited", "top2_margin": 0.3059980720281601, "near_tie": false, "gold": "permitted", "correct": false, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "prohibited", "confidence": 0.5530052781105042, "probabilities": { "insufficient": 0.24700720608234406, "permitted": 0.199987530708313, "prohibited": 0.5530052781105042 } } }, "usage": { "input_tokens": 83, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_ef3e0f8e657bf59f8a29", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 90, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 90, "positions": 90, "bucket": 256, "ids": [ 1, 17, 3111, 1033, 4889, 521, 779, 2835, 10423, 57119, 12199, 52010, 1403, 936, 19980, 916, 22537, 523, 46050, 931, 835, 535, 40831, 3498, 8177, 521, 1203, 779, 10423, 57119, 3498, 1014, 8177, 523, 46050, 931, 879, 535, 40831, 3498, 1014, 8177, 521, 1203, 779, 10423, 57119, 3498, 8177, 523, 18, 15534, 779, 13719, 8177, 4262, 811, 779, 19670, 10423, 57119, 523, 19, 16, 835, 535, 46050, 931, 835, 20, 19, 16, 28441, 535, 42749, 13719, 14350, 779, 21968, 4589, 20, 19, 16, 879, 535, 46050, 931, 879, 20, 21 ], "markers": [ 63, 71, 82 ], "logits_raw": [ -0.09687026590108871, 1.2967870235443115, 0.46955496072769165 ], "probs_raw": [ 0.1472424864768982, 0.5933225154876709, 0.2594349682331085 ], "probs": [ 0.19212587177753448, 0.5199291706085205, 0.2879450023174286 ], "argmax_index": 1, "argmax": "insufficient", "top2_margin": 0.23198416829109192, "near_tie": false, "gold": "B", "correct": false, "host_softmax_max_abs_dp": 1.4901161193847656e-08, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "insufficient", "confidence": 0.5199291706085205, "probabilities": { "A": 0.19212587177753448, "insufficient": 0.5199291706085205, "B": 0.2879450023174286 } } }, "usage": { "input_tokens": 90, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_f2b4ec4930322fe45118", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 80, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 80, "positions": 80, "bucket": 256, "ids": [ 1, 17, 1098, 6765, 22558, 10004, 2782, 768, 2984, 797, 902, 7477, 4611, 62248, 810, 4943, 876, 23369, 2796, 18109, 16711, 523, 18, 560, 965, 16711, 1550, 916, 779, 6765, 1090, 11599, 15001, 523, 3354, 16711, 874, 4997, 540, 19, 16, 28441, 535, 941, 21706, 2203, 2177, 1014, 27169, 3984, 779, 7330, 22558, 779, 4272, 20, 19, 16, 33776, 535, 941, 9519, 7330, 14480, 1512, 779, 4272, 20, 19, 16, 20458, 535, 941, 9519, 7330, 22558, 779, 4272, 20, 21 ], "markers": [ 40, 57, 69 ], "logits_raw": [ 2.063524007797241, 3.0282185077667236, 0.5558172464370728 ], "probs_raw": [ 0.26005077362060547, 0.6823695302009583, 0.057579707354307175 ], "probs": [ 0.30006879568099976, 0.597724199295044, 0.1022069901227951 ], "argmax_index": 1, "argmax": "prohibited", "top2_margin": 0.2976554036140442, "near_tie": false, "gold": "prohibited", "correct": true, "host_softmax_max_abs_dp": 7.450580596923828e-09, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "prohibited", "confidence": 0.597724199295044, "probabilities": { "insufficient": 0.30006879568099976, "prohibited": 0.597724199295044, "permitted": 0.1022069901227951 } } }, "usage": { "input_tokens": 80, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_f3d13cf184f916677def", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 95, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 95, "positions": 95, "bucket": 256, "ids": [ 1, 17, 1098, 26862, 3446, 3498, 779, 2391, 931, 856, 797, 8784, 523, 941, 51746, 30174, 3498, 936, 1178, 3191, 8784, 523, 9758, 7944, 938, 2835, 523, 941, 21706, 13061, 3498, 779, 51746, 30174, 12199, 6249, 9448, 19148, 1518, 1547, 5617, 9171, 523, 18, 9886, 894, 779, 4673, 535, 1490, 779, 21706, 13061, 521, 779, 2391, 931, 856, 6654, 797, 8784, 523, 19, 16, 18197, 13735, 535, 941, 4589, 48308, 779, 11977, 20, 19, 16, 28441, 535, 941, 4589, 2177, 1014, 6122, 4021, 20, 19, 16, 7951, 535, 941, 4589, 48308, 779, 4673, 20, 21 ], "markers": [ 63, 74, 85 ], "logits_raw": [ 1.1147890090942383, 1.3649086952209473, 3.313030481338501 ], "probs_raw": [ 0.08854779601097107, 0.11371122300624847, 0.7977409362792969 ], "probs": [ 0.14278262853622437, 0.1707146316766739, 0.6865026950836182 ], "argmax_index": 2, "argmax": "supported", "top2_margin": 0.5157880634069443, "near_tie": false, "gold": "contradicted", "correct": false, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "supported", "confidence": 0.6865026950836182, "probabilities": { "contradicted": 0.14278262853622437, "insufficient": 0.1707146316766739, "supported": 0.6865026950836182 } } }, "usage": { "input_tokens": 95, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_f40beba9088c8db8bbd6", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 68, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 68, "positions": 68, "bucket": 256, "ids": [ 1, 17, 1098, 2525, 7867, 13055, 16400, 31286, 523, 941, 23120, 44987, 1186, 1052, 1014, 4763, 1381, 25539, 811, 779, 13608, 1090, 28623, 523, 18, 9886, 894, 779, 4673, 535, 16400, 31286, 1166, 13055, 523, 19, 16, 7951, 535, 941, 4589, 48308, 779, 4673, 20, 19, 16, 28441, 535, 941, 4589, 2177, 1014, 6122, 4021, 20, 19, 16, 18197, 13735, 535, 941, 4589, 48308, 779, 11977, 20, 21 ], "markers": [ 36, 46, 57 ], "logits_raw": [ 9.068342208862305, 0.18565919995307922, -5.66766357421875 ], "probs_raw": [ 0.9998608827590942, 0.00013875206059310585, 3.9826630882089376e-07 ], "probs": [ 0.998221218585968, 0.0017519677057862282, 2.6769057512865402e-05 ], "argmax_index": 0, "argmax": "supported", "top2_margin": 0.9964692508801818, "near_tie": false, "gold": "supported", "correct": true, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "supported", "confidence": 0.998221218585968, "probabilities": { "supported": 0.998221218585968, "insufficient": 0.0017519677057862282, "contradicted": 2.6769057512865402e-05 } } }, "usage": { "input_tokens": 68, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_f41ae66b3c09630c7a42", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 68, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 68, "positions": 68, "bucket": 256, "ids": [ 1, 17, 55945, 931, 835, 535, 27022, 779, 8449, 810, 5466, 768, 16400, 523, 46050, 931, 879, 535, 2382, 578, 786, 779, 7605, 8449, 810, 3089, 1352, 6212, 523, 18, 52739, 779, 13719, 14065, 811, 12940, 779, 7605, 8449, 523, 19, 16, 28441, 535, 42749, 13719, 14350, 779, 21968, 4589, 20, 19, 16, 835, 535, 46050, 931, 835, 20, 19, 16, 879, 535, 46050, 931, 879, 20, 21 ], "markers": [ 41, 52, 60 ], "logits_raw": [ 0.4962294399738312, -0.13544343411922455, 0.3153257966041565 ], "probs_raw": [ 0.4226154386997223, 0.2247052937746048, 0.3526792526245117 ], "probs": [ 0.3975161612033844, 0.25315651297569275, 0.34932732582092285 ], "argmax_index": 0, "argmax": "insufficient", "top2_margin": 0.04818883538246155, "near_tie": false, "gold": "B", "correct": false, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "insufficient", "confidence": 0.3975161612033844, "probabilities": { "insufficient": 0.3975161612033844, "A": 0.25315651297569275, "B": 0.34932732582092285 } } }, "usage": { "input_tokens": 68, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_f46f392ef9e9e9df564b", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 73, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 73, "positions": 73, "bucket": 256, "ids": [ 1, 17, 1098, 12153, 30403, 33300, 884, 17980, 523, 941, 21706, 3084, 4654, 1512, 1113, 2919, 15001, 3807, 810, 2177, 1014, 6122, 2835, 28195, 523, 18, 9886, 894, 779, 4673, 535, 15001, 811, 19382, 779, 52518, 856, 6654, 6612, 523, 19, 16, 7951, 535, 941, 4589, 48308, 779, 4673, 20, 19, 16, 28441, 535, 941, 4589, 2177, 1014, 6122, 4021, 20, 19, 16, 18197, 13735, 535, 941, 4589, 48308, 779, 11977, 20, 21 ], "markers": [ 41, 51, 62 ], "logits_raw": [ -1.2643077373504639, 2.459071397781372, -1.124413251876831 ], "probs_raw": [ 0.022959887981414795, 0.9506327509880066, 0.02640737034380436 ], "probs": [ 0.060983750969171524, 0.8716233968734741, 0.06739288568496704 ], "argmax_index": 1, "argmax": "insufficient", "top2_margin": 0.8042305111885071, "near_tie": false, "gold": "insufficient", "correct": true, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "insufficient", "confidence": 0.8716233968734741, "probabilities": { "supported": 0.060983750969171524, "insufficient": 0.8716233968734741, "contradicted": 0.06739288568496704 } } }, "usage": { "input_tokens": 73, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_f6d68126b43f8c9ed183", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 71, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 71, "positions": 71, "bucket": 256, "ids": [ 1, 17, 55945, 931, 835, 535, 941, 6634, 8269, 896, 779, 7850, 33037, 1927, 7410, 523, 46050, 931, 879, 535, 941, 6634, 15994, 33037, 29067, 7441, 1203, 3498, 1295, 1927, 953, 2685, 523, 18, 52739, 779, 13719, 15603, 902, 6023, 1927, 1725, 523, 19, 16, 28441, 535, 42749, 13719, 14350, 779, 21968, 4589, 20, 19, 16, 835, 535, 46050, 931, 835, 20, 19, 16, 879, 535, 46050, 931, 879, 20, 21 ], "markers": [ 44, 55, 63 ], "logits_raw": [ -1.3054568767547607, -1.7994532585144043, -3.502875328063965 ], "probs_raw": [ 0.580965518951416, 0.3544952869415283, 0.06453920900821686 ], "probs": [ 0.5233494639396667, 0.36773747205734253, 0.10891313850879669 ], "argmax_index": 0, "argmax": "insufficient", "top2_margin": 0.15561199188232422, "near_tie": false, "gold": "A", "correct": false, "host_softmax_max_abs_dp": 2.9802322387695312e-08, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "insufficient", "confidence": 0.5233494639396667, "probabilities": { "insufficient": 0.5233494639396667, "A": 0.36773747205734253, "B": 0.10891313850879669 } } }, "usage": { "input_tokens": 71, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_f756e1cb04e11766341e", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 96, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 96, "positions": 96, "bucket": 256, "ids": [ 1, 17, 3111, 1033, 4889, 521, 779, 2835, 10423, 57119, 12199, 52010, 1403, 936, 19980, 916, 22537, 523, 46050, 931, 835, 535, 40831, 3498, 8177, 521, 1203, 779, 10423, 57119, 3498, 1014, 8177, 523, 46050, 931, 879, 535, 40831, 3498, 1014, 8177, 521, 1203, 779, 10423, 57119, 3498, 8177, 523, 18, 15246, 779, 8530, 13061, 521, 22537, 12199, 19980, 523, 22269, 779, 13719, 8177, 4262, 811, 22537, 523, 19, 16, 28441, 535, 42749, 13719, 14350, 779, 21968, 4589, 20, 19, 16, 835, 535, 46050, 931, 835, 20, 19, 16, 879, 535, 46050, 931, 879, 20, 21 ], "markers": [ 69, 80, 88 ], "logits_raw": [ -0.13149817287921906, -2.9378244876861572, -0.3618924617767334 ], "probs_raw": [ 0.5391861796379089, 0.03258117288351059, 0.42823266983032227 ], "probs": [ 0.5042976140975952, 0.0679316371679306, 0.4277706742286682 ], "argmax_index": 0, "argmax": "insufficient", "top2_margin": 0.076526939868927, "near_tie": false, "gold": "A", "correct": false, "host_softmax_max_abs_dp": 2.9802322387695312e-08, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "insufficient", "confidence": 0.5042976140975952, "probabilities": { "insufficient": 0.5042976140975952, "A": 0.0679316371679306, "B": 0.4277706742286682 } } }, "usage": { "input_tokens": 96, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_fb73a85d386cf59b0c51", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 71, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 71, "positions": 71, "bucket": 256, "ids": [ 1, 17, 1098, 22773, 12800, 768, 6177, 1203, 1352, 3893, 12757, 856, 58754, 523, 18, 31678, 3893, 856, 6767, 1550, 1720, 3507, 25252, 980, 17784, 523, 3354, 936, 6305, 540, 19, 16, 28441, 535, 941, 21706, 2203, 2177, 1014, 27169, 3984, 779, 7330, 22558, 779, 4272, 20, 19, 16, 20458, 535, 941, 9519, 7330, 22558, 779, 4272, 20, 19, 16, 33776, 535, 941, 9519, 7330, 14480, 1512, 779, 4272, 20, 21 ], "markers": [ 31, 48, 59 ], "logits_raw": [ 1.5704766511917114, -1.7801976203918457, 2.1247355937957764 ], "probs_raw": [ 0.3602679967880249, 0.0126312505453825, 0.6271007061004639 ], "probs": [ 0.38803747296333313, 0.035431068390607834, 0.5765314698219299 ], "argmax_index": 2, "argmax": "prohibited", "top2_margin": 0.1884939968585968, "near_tie": false, "gold": "insufficient", "correct": false, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "prohibited", "confidence": 0.5765314698219299, "probabilities": { "insufficient": 0.38803747296333313, "permitted": 0.035431068390607834, "prohibited": 0.5765314698219299 } } }, "usage": { "input_tokens": 71, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_fda0ca94652c6d739ecf", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 84, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 84, "positions": 84, "bucket": 256, "ids": [ 1, 17, 1098, 37181, 6175, 4943, 876, 13748, 774, 779, 30122, 9041, 810, 13546, 1550, 779, 1232, 33388, 920, 2543, 10074, 523, 18, 542, 37181, 6175, 22558, 9544, 8544, 811, 779, 3508, 936, 13546, 523, 3354, 779, 33648, 7157, 779, 30122, 9041, 540, 19, 16, 33776, 535, 941, 9519, 7330, 14480, 1512, 779, 4272, 20, 19, 16, 20458, 535, 941, 9519, 7330, 22558, 779, 4272, 20, 19, 16, 28441, 535, 941, 21706, 2203, 2177, 1014, 27169, 3984, 779, 7330, 22558, 779, 4272, 20, 21 ], "markers": [ 44, 56, 67 ], "logits_raw": [ 2.2747561931610107, 1.4293830394744873, 2.278024196624756 ], "probs_raw": [ 0.41107088327407837, 0.17651264369487762, 0.4124164879322052 ], "probs": [ 0.3923072814941406, 0.21446861326694489, 0.3932241201400757 ], "argmax_index": 2, "argmax": "insufficient", "top2_margin": 0.0009168386459350586, "near_tie": true, "gold": "prohibited", "correct": false, "host_softmax_max_abs_dp": 1.4901161193847656e-08, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "insufficient", "confidence": 0.3932241201400757, "probabilities": { "prohibited": 0.3923072814941406, "permitted": 0.21446861326694489, "insufficient": 0.3932241201400757 } } }, "usage": { "input_tokens": 84, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_fdb9fc214aea2710763a", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 58, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 58, "positions": 58, "bucket": 256, "ids": [ 1, 17, 55945, 931, 835, 535, 941, 13608, 6602, 768, 2154, 523, 46050, 931, 879, 535, 941, 11019, 856, 12534, 523, 18, 52739, 779, 13719, 3575, 4651, 41071, 3665, 523, 19, 16, 879, 535, 46050, 931, 879, 20, 19, 16, 835, 535, 46050, 931, 835, 20, 19, 16, 28441, 535, 42749, 13719, 14350, 779, 21968, 4589, 20, 21 ], "markers": [ 31, 39, 47 ], "logits_raw": [ 0.5282842516899109, -4.471530914306641, 1.901053786277771 ], "probs_raw": [ 0.20189766585826874, 0.0013606270076707006, 0.7967416644096375 ], "probs": [ 0.2706928849220276, 0.0076097361743450165, 0.7216973900794983 ], "argmax_index": 2, "argmax": "insufficient", "top2_margin": 0.4510045051574707, "near_tie": false, "gold": "insufficient", "correct": true, "host_softmax_max_abs_dp": 2.9802322387695312e-08, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "insufficient", "confidence": 0.7216973900794983, "probabilities": { "B": 0.2706928849220276, "A": 0.0076097361743450165, "insufficient": 0.7216973900794983 } } }, "usage": { "input_tokens": 58, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_fdd902cc6e65fad7cb4f", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 91, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 91, "positions": 91, "bucket": 256, "ids": [ 1, 17, 1098, 7330, 22558, 2314, 53859, 968, 779, 7576, 1203, 23369, 2796, 1523, 3759, 53859, 523, 46050, 931, 835, 535, 859, 5068, 3346, 17784, 10562, 5828, 988, 779, 7576, 523, 46050, 931, 879, 535, 859, 5068, 936, 1550, 988, 835, 521, 810, 1052, 1295, 1312, 8825, 523, 18, 15534, 779, 13719, 12467, 6612, 3346, 17784, 10562, 1490, 779, 9519, 53859, 7330, 523, 19, 16, 879, 535, 46050, 931, 879, 20, 19, 16, 835, 535, 46050, 931, 835, 20, 19, 16, 28441, 535, 42749, 13719, 14350, 779, 21968, 4589, 20, 21 ], "markers": [ 64, 72, 80 ], "logits_raw": [ 0.6088454723358154, 2.894845724105835, 1.3413251638412476 ], "probs_raw": [ 0.07742483913898468, 0.7615134716033936, 0.16106167435646057 ], "probs": [ 0.12809723615646362, 0.6557409167289734, 0.21616187691688538 ], "argmax_index": 1, "argmax": "A", "top2_margin": 0.439579039812088, "near_tie": false, "gold": "A", "correct": true, "host_softmax_max_abs_dp": 1.4901161193847656e-08, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "A", "confidence": 0.6557409167289734, "probabilities": { "B": 0.12809723615646362, "A": 0.6557409167289734, "insufficient": 0.21616187691688538 } } }, "usage": { "input_tokens": 91, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "semif_ffee861474a9665a093a", "mode": "text", "source": "semif_authored144", "native": true, "prefix": 0, "usage_input_tokens": 77, "questions": [ { "qid": "decision", "type": "choice", "K": 3, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 77, "positions": 77, "bucket": 256, "ids": [ 1, 17, 1098, 3434, 919, 14350, 768, 2154, 1203, 1295, 2681, 933, 1336, 6249, 523, 18, 542, 12518, 1090, 2230, 6835, 4624, 797, 779, 20610, 810, 1178, 1295, 21701, 27434, 523, 3354, 936, 874, 12800, 540, 19, 16, 33776, 535, 941, 9519, 7330, 14480, 1512, 779, 4272, 20, 19, 16, 20458, 535, 941, 9519, 7330, 22558, 779, 4272, 20, 19, 16, 28441, 535, 941, 21706, 2203, 2177, 1014, 27169, 3984, 779, 7330, 22558, 779, 4272, 20, 21 ], "markers": [ 37, 49, 60 ], "logits_raw": [ 0.34799861907958984, 4.5191450119018555, 2.7875254154205322 ], "probs_raw": [ 0.012943761423230171, 0.8386222124099731, 0.14843404293060303 ], "probs": [ 0.03788952901959419, 0.745668888092041, 0.2164415717124939 ], "argmax_index": 1, "argmax": "permitted", "top2_margin": 0.5292273163795471, "near_tie": false, "gold": "insufficient", "correct": false, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "decision": { "type": "choice", "choice": "permitted", "confidence": 0.745668888092041, "probabilities": { "prohibited": 0.03788952901959419, "permitted": 0.745668888092041, "insufficient": 0.2164415717124939 } } }, "usage": { "input_tokens": 77, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "tv4_000", "mode": "text", "source": "transfer_v4_dev_head", "native": true, "prefix": 0, "usage_input_tokens": 103, "questions": [ { "qid": "answer", "type": "choice", "K": 4, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 103, "positions": 103, "bucket": 256, "ids": [ 1, 17, 38685, 42505, 6881, 997, 53983, 3609, 1878, 997, 39420, 6881, 997, 28828, 803, 779, 2773, 2172, 1014, 25764, 906, 768, 13974, 803, 779, 28857, 7169, 797, 779, 730, 898, 881, 3935, 8523, 602, 18, 28828, 7832, 16066, 13041, 779, 3952, 540, 19, 16, 768, 535, 941, 10557, 7169, 26807, 988, 1994, 768, 3674, 811, 768, 3134, 10282, 25624, 523, 20, 19, 16, 794, 535, 1592, 3761, 797, 779, 28857, 7169, 523, 20, 19, 16, 781, 535, 835, 10636, 797, 779, 10557, 7169, 523, 20, 19, 16, 778, 535, 835, 4320, 797, 779, 2080, 810, 5681, 803, 14773, 39451, 523, 20, 21 ], "markers": [ 44, 63, 75, 87 ], "logits_raw": [ -5.553777694702148, -7.747239112854004, -5.589899063110352, -5.4578471183776855 ], "probs_raw": [ 0.3147885501384735, 0.035108376294374466, 0.3036208748817444, 0.3464823067188263 ], "probs": [ 0.3072996139526367, 0.06413247436285019, 0.29947182536125183, 0.32909613847732544 ], "argmax_index": 3, "argmax": "d", "top2_margin": 0.02179652452468872, "near_tie": false, "gold": "c", "correct": false, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "answer": { "type": "choice", "choice": "d", "confidence": 0.32909613847732544, "probabilities": { "a": 0.3072996139526367, "b": 0.06413247436285019, "c": 0.29947182536125183, "d": 0.32909613847732544 } } }, "usage": { "input_tokens": 103, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "tv4_001", "mode": "text", "source": "transfer_v4_dev_head", "native": true, "prefix": 0, "usage_input_tokens": 74, "questions": [ { "qid": "answer", "type": "choice", "K": 4, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 74, "positions": 74, "bucket": 256, "ids": [ 1, 17, 38685, 42505, 6881, 997, 50988, 4331, 1878, 997, 39420, 6881, 997, 3493, 856, 6365, 2807, 4791, 14232, 4027, 8523, 602, 18, 28828, 7832, 16066, 13041, 779, 3952, 540, 19, 16, 768, 535, 42144, 3685, 803, 35231, 6104, 523, 20, 19, 16, 794, 535, 15858, 1227, 41134, 852, 523, 20, 19, 16, 781, 535, 5148, 44771, 3685, 803, 35231, 6104, 523, 20, 19, 16, 778, 535, 63401, 17712, 30769, 10254, 523, 20, 21 ], "markers": [ 31, 42, 52, 64 ], "logits_raw": [ 0.6107369661331177, -4.443951606750488, 2.522935152053833, -6.117058753967285 ], "probs_raw": [ 0.12860864400863647, 0.0008204402402043343, 0.8704168796539307, 0.00015396632079500705 ], "probs": [ 0.2018294632434845, 0.0054557411931455135, 0.7910635471343994, 0.0016512047732248902 ], "argmax_index": 2, "argmax": "c", "top2_margin": 0.5892340838909149, "near_tie": false, "gold": "c", "correct": true, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "answer": { "type": "choice", "choice": "c", "confidence": 0.7910635471343994, "probabilities": { "a": 0.2018294632434845, "b": 0.0054557411931455135, "c": 0.7910635471343994, "d": 0.0016512047732248902 } } }, "usage": { "input_tokens": 74, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "tv4_002", "mode": "text", "source": "transfer_v4_dev_head", "native": true, "prefix": 0, "usage_input_tokens": 107, "questions": [ { "qid": "answer", "type": "choice", "K": 4, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 107, "positions": 107, "bucket": 256, "ids": [ 1, 17, 38685, 42505, 6881, 997, 17756, 2254, 4167, 578, 2499, 1504, 1878, 997, 39420, 6881, 997, 28828, 803, 779, 2773, 856, 5405, 8523, 602, 18, 28828, 7832, 16066, 13041, 779, 3952, 540, 19, 16, 768, 535, 1030, 779, 1787, 2685, 521, 1113, 23032, 938, 9452, 523, 20, 19, 16, 794, 535, 1030, 779, 3290, 2685, 521, 1113, 23032, 938, 9452, 523, 20, 19, 16, 781, 535, 1030, 779, 1787, 2685, 521, 7018, 856, 1014, 3222, 811, 8529, 7408, 811, 3807, 797, 6154, 523, 20, 19, 16, 778, 535, 1030, 779, 3290, 2685, 521, 7018, 856, 3222, 811, 8529, 7408, 811, 3807, 797, 6154, 523, 20, 21 ], "markers": [ 34, 49, 64, 86 ], "logits_raw": [ -3.5198915004730225, -1.3380160331726074, -1.1915581226348877, -0.26958832144737244 ], "probs_raw": [ 0.021776100620627403, 0.1929997056722641, 0.22344081103801727, 0.5617833733558655 ], "probs": [ 0.04711969569325447, 0.2239195704460144, 0.24861547350883484, 0.4803452491760254 ], "argmax_index": 3, "argmax": "d", "top2_margin": 0.23172977566719055, "near_tie": false, "gold": "a", "correct": false, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "answer": { "type": "choice", "choice": "d", "confidence": 0.4803452491760254, "probabilities": { "a": 0.04711969569325447, "b": 0.2239195704460144, "c": 0.24861547350883484, "d": 0.4803452491760254 } } }, "usage": { "input_tokens": 107, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "tv4_003", "mode": "text", "source": "transfer_v4_dev_head", "native": true, "prefix": 0, "usage_input_tokens": 87, "questions": [ { "qid": "answer", "type": "choice", "K": 4, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 87, "positions": 87, "bucket": 256, "ids": [ 1, 17, 38685, 42505, 6881, 997, 28009, 19179, 1878, 997, 39420, 6881, 997, 5711, 4092, 896, 6993, 7239, 1914, 1569, 3476, 803, 62151, 856, 896, 511, 602, 18, 28828, 7832, 16066, 13041, 779, 3952, 540, 19, 16, 768, 535, 6405, 62311, 834, 1817, 1277, 963, 768, 12280, 3209, 20, 19, 16, 794, 535, 2216, 1112, 1014, 3896, 1842, 1620, 856, 3529, 884, 20, 19, 16, 781, 535, 1470, 1400, 874, 768, 15513, 1723, 875, 22344, 20, 19, 16, 778, 535, 6504, 7485, 1052, 2920, 28195, 20, 21 ], "markers": [ 36, 50, 64, 77 ], "logits_raw": [ -2.4365646839141846, -1.0311179161071777, -3.46833872795105, -1.6626391410827637 ], "probs_raw": [ 0.13154460489749908, 0.5363531708717346, 0.046879105269908905, 0.2852230668067932 ], "probs": [ 0.1681860387325287, 0.4589926302433014, 0.08048214763402939, 0.2923392355442047 ], "argmax_index": 1, "argmax": "b", "top2_margin": 0.16665339469909668, "near_tie": false, "gold": "a", "correct": false, "host_softmax_max_abs_dp": 2.9802322387695312e-08, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "answer": { "type": "choice", "choice": "b", "confidence": 0.4589926302433014, "probabilities": { "a": 0.1681860387325287, "b": 0.4589926302433014, "c": 0.08048214763402939, "d": 0.2923392355442047 } } }, "usage": { "input_tokens": 87, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "tv4_004", "mode": "text", "source": "transfer_v4_dev_head", "native": true, "prefix": 0, "usage_input_tokens": 75, "questions": [ { "qid": "answer", "type": "choice", "K": 4, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 75, "positions": 75, "bucket": 256, "ids": [ 1, 17, 38685, 42505, 6881, 997, 28009, 19179, 1878, 997, 39420, 6881, 997, 544, 22525, 20800, 938, 1519, 1316, 511, 602, 18, 28828, 7832, 16066, 13041, 779, 3952, 540, 19, 16, 768, 535, 43391, 5747, 2076, 779, 1739, 24452, 20, 19, 16, 794, 535, 12220, 2731, 3209, 810, 6993, 7239, 20, 19, 16, 781, 535, 9195, 730, 2096, 933, 1209, 1779, 2731, 20, 19, 16, 778, 535, 14566, 1014, 4763, 8107, 2731, 3209, 20, 21 ], "markers": [ 30, 41, 52, 64 ], "logits_raw": [ -1.9950649738311768, -0.7058923244476318, 4.306713104248047, 0.4114217162132263 ], "probs_raw": [ 0.0017816874897107482, 0.006467134226113558, 0.9719834327697754, 0.01976766437292099 ], "probs": [ 0.010075385682284832, 0.025304943323135376, 0.9084072113037109, 0.05621250718832016 ], "argmax_index": 2, "argmax": "c", "top2_margin": 0.8521947041153908, "near_tie": false, "gold": "c", "correct": true, "host_softmax_max_abs_dp": 1.1920928955078125e-07, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "answer": { "type": "choice", "choice": "c", "confidence": 0.9084072113037109, "probabilities": { "a": 0.010075385682284832, "b": 0.025304943323135376, "c": 0.9084072113037109, "d": 0.05621250718832016 } } }, "usage": { "input_tokens": 75, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "tv4_005", "mode": "text", "source": "transfer_v4_dev_head", "native": true, "prefix": 0, "usage_input_tokens": 87, "questions": [ { "qid": "answer", "type": "choice", "K": 4, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 87, "positions": 87, "bucket": 256, "ids": [ 1, 17, 38685, 42505, 6881, 997, 5606, 30049, 1878, 997, 39420, 6881, 997, 1098, 22152, 5939, 1178, 15161, 28327, 9251, 55053, 968, 1144, 803, 779, 2773, 8523, 602, 18, 28828, 7832, 16066, 13041, 779, 3952, 540, 19, 16, 768, 535, 41305, 29349, 947, 957, 20467, 730, 533, 7695, 20, 19, 16, 794, 535, 5001, 3476, 803, 58190, 20, 19, 16, 781, 535, 40494, 24612, 14275, 810, 18357, 988, 2630, 4492, 768, 5511, 61129, 37185, 20, 19, 16, 778, 535, 5001, 63011, 13415, 33955, 1146, 19288, 20, 21 ], "markers": [ 37, 50, 59, 76 ], "logits_raw": [ -0.5547366142272949, 0.12310024350881577, -2.19450044631958, -2.8999176025390625 ], "probs_raw": [ 0.3067985475063324, 0.6042742133140564, 0.05952684208750725, 0.029400456696748734 ], "probs": [ 0.3205031454563141, 0.520139217376709, 0.09933988004922867, 0.06001778692007065 ], "argmax_index": 1, "argmax": "b", "top2_margin": 0.1996360719203949, "near_tie": false, "gold": "b", "correct": true, "host_softmax_max_abs_dp": 2.9802322387695312e-08, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "answer": { "type": "choice", "choice": "b", "confidence": 0.520139217376709, "probabilities": { "a": 0.3205031454563141, "b": 0.520139217376709, "c": 0.09933988004922867, "d": 0.06001778692007065 } } }, "usage": { "input_tokens": 87, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "tv4_006", "mode": "text", "source": "transfer_v4_dev_head", "native": true, "prefix": 0, "usage_input_tokens": 243, "questions": [ { "qid": "answer", "type": "choice", "K": 4, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 243, "positions": 243, "bucket": 256, "ids": [ 1, 17, 38685, 42505, 6881, 997, 30327, 16964, 9152, 1878, 997, 39420, 6881, 997, 542, 730, 5489, 6194, 10835, 6394, 4834, 811, 779, 16329, 1929, 803, 902, 730, 533, 33344, 3129, 803, 13467, 20738, 521, 19236, 521, 810, 8550, 17828, 523, 7074, 1033, 2926, 521, 1689, 1236, 1178, 1320, 6448, 59335, 774, 521, 13467, 57554, 521, 810, 1178, 1381, 9730, 834, 27252, 963, 1336, 875, 34585, 5019, 523, 7438, 779, 3650, 730, 527, 6272, 521, 1689, 1178, 1320, 2204, 19941, 803, 54670, 19877, 810, 3290, 2122, 803, 8311, 896, 1052, 44405, 791, 1450, 988, 5392, 523, 3365, 24037, 856, 730, 2346, 40619, 521, 810, 3186, 4809, 856, 730, 7984, 524, 2475, 8061, 904, 580, 523, 20737, 14435, 4427, 1295, 32219, 523, 25650, 5742, 14435, 4427, 768, 35459, 15365, 810, 1737, 34270, 4531, 523, 2799, 3498, 896, 1689, 856, 1295, 5039, 9368, 797, 4637, 896, 1689, 1593, 811, 6401, 523, 2799, 1178, 60532, 2542, 852, 2630, 768, 2984, 523, 3365, 51879, 9403, 856, 730, 1632, 836, 51275, 521, 810, 18941, 7097, 1215, 769, 9403, 856, 730, 9300, 21955, 30147, 523, 20463, 803, 779, 2773, 856, 779, 1529, 6994, 4764, 3801, 797, 3157, 8523, 602, 18, 28828, 7832, 16066, 13041, 779, 3952, 540, 19, 16, 768, 535, 889, 1174, 14191, 814, 6738, 20, 19, 16, 794, 535, 12656, 14641, 51112, 6738, 20, 19, 16, 781, 535, 918, 1335, 2086, 13001, 17292, 8886, 6738, 20, 19, 16, 778, 535, 2334, 1974, 813, 976, 6738, 20, 21 ], "markers": [ 202, 212, 221, 233 ], "logits_raw": [ -1.3662583827972412, -2.0687501430511475, -1.8015267848968506, -0.1909313052892685 ], "probs_raw": [ 0.18581685423851013, 0.09204428642988205, 0.12024041265249252, 0.6018983721733093 ], "probs": [ 0.21488778293132782, 0.13009950518608093, 0.1574624925851822, 0.4975501596927643 ], "argmax_index": 3, "argmax": "d", "top2_margin": 0.28266237676143646, "near_tie": false, "gold": "d", "correct": true, "host_softmax_max_abs_dp": 1.4901161193847656e-08, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "answer": { "type": "choice", "choice": "d", "confidence": 0.4975501596927643, "probabilities": { "a": 0.21488778293132782, "b": 0.13009950518608093, "c": 0.1574624925851822, "d": 0.4975501596927643 } } }, "usage": { "input_tokens": 243, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "tv4_007", "mode": "text", "source": "transfer_v4_dev_head", "native": true, "prefix": 0, "usage_input_tokens": 92, "questions": [ { "qid": "answer", "type": "choice", "K": 4, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 92, "positions": 92, "bucket": 256, "ids": [ 1, 17, 38685, 42505, 6881, 997, 780, 776, 5797, 1878, 997, 39420, 6881, 997, 61490, 591, 2115, 856, 779, 1996, 803, 511, 602, 18, 28828, 7832, 16066, 13041, 779, 3952, 540, 19, 16, 768, 535, 5287, 57918, 7401, 810, 7809, 1720, 11192, 2115, 523, 20, 19, 16, 794, 535, 54375, 11150, 20779, 30667, 7401, 810, 7809, 2076, 11192, 2115, 523, 20, 19, 16, 781, 535, 5287, 57918, 7401, 810, 7809, 2076, 11192, 2115, 523, 20, 19, 16, 778, 535, 54375, 11150, 20779, 30667, 7401, 810, 7809, 1720, 11192, 2115, 523, 20, 21 ], "markers": [ 32, 46, 62, 76 ], "logits_raw": [ -1.0777621269226074, 0.14282402396202087, -1.3859832286834717, -1.6446017026901245 ], "probs_raw": [ 0.17570854723453522, 0.5955068469047546, 0.12910236418247223, 0.09968220442533493 ], "probs": [ 0.20572371780872345, 0.4919832646846771, 0.16506841778755188, 0.13722461462020874 ], "argmax_index": 1, "argmax": "b", "top2_margin": 0.2862595468759537, "near_tie": false, "gold": "b", "correct": true, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "answer": { "type": "choice", "choice": "b", "confidence": 0.4919832646846771, "probabilities": { "a": 0.20572371780872345, "b": 0.4919832646846771, "c": 0.16506841778755188, "d": 0.13722461462020874 } } }, "usage": { "input_tokens": 92, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "tv4_008", "mode": "text", "source": "transfer_v4_dev_head", "native": true, "prefix": 0, "usage_input_tokens": 106, "questions": [ { "qid": "answer", "type": "choice", "K": 4, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 106, "positions": 106, "bucket": 256, "ids": [ 1, 17, 38685, 42505, 6881, 997, 17756, 2254, 16290, 1878, 997, 39420, 6881, 997, 28828, 803, 779, 2773, 6306, 856, 7253, 875, 12467, 3658, 779, 1883, 12709, 797, 768, 12825, 7322, 803, 1883, 797, 3053, 2779, 586, 13761, 803, 768, 3226, 8523, 602, 18, 28828, 7832, 16066, 13041, 779, 3952, 540, 19, 16, 768, 535, 779, 3636, 49121, 5401, 803, 1883, 20, 19, 16, 794, 535, 779, 3636, 44652, 801, 803, 5401, 803, 1883, 20, 19, 16, 781, 535, 779, 1569, 5480, 7614, 803, 1883, 20, 19, 16, 778, 535, 779, 3636, 44652, 801, 5401, 803, 59697, 797, 779, 3053, 2779, 586, 2009, 7680, 20, 21 ], "markers": [ 52, 63, 76, 87 ], "logits_raw": [ -2.249392509460449, 0.9455831050872803, -7.214566707611084, 1.1811388731002808 ], "probs_raw": [ 0.017758958041667938, 0.4334886968135834, 0.00012389960465952754, 0.5486285090446472 ], "probs": [ 0.04459759220480919, 0.4370166063308716, 0.0012851412175223231, 0.5171006321907043 ], "argmax_index": 3, "argmax": "d", "top2_margin": 0.08008402585983276, "near_tie": false, "gold": "b", "correct": false, "host_softmax_max_abs_dp": 3.725290298461914e-09, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "answer": { "type": "choice", "choice": "d", "confidence": 0.5171006321907043, "probabilities": { "a": 0.04459759220480919, "b": 0.4370166063308716, "c": 0.0012851412175223231, "d": 0.5171006321907043 } } }, "usage": { "input_tokens": 106, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "tv4_009", "mode": "text", "source": "transfer_v4_dev_head", "native": true, "prefix": 0, "usage_input_tokens": 339, "questions": [ { "qid": "answer", "type": "choice", "K": 4, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 339, "positions": 339, "bucket": 512, "ids": [ 1, 17, 38685, 42505, 6881, 997, 17756, 2254, 1081, 3129, 1878, 997, 39420, 6881, 997, 2443, 3952, 11335, 811, 779, 2773, 2203, 13909, 587, 18588, 550, 48949, 1014, 768, 3223, 521, 2084, 17423, 1665, 521, 896, 2908, 779, 1509, 12037, 803, 4367, 843, 7895, 856, 2166, 3522, 2759, 8619, 810, 3676, 1779, 1109, 1720, 1101, 2507, 15808, 3279, 810, 3226, 1327, 803, 779, 59362, 2892, 810, 24408, 1413, 521, 896, 1023, 803, 5690, 1738, 3467, 1052, 779, 22148, 811, 2021, 5572, 3789, 800, 797, 1559, 38501, 810, 48079, 6454, 906, 938, 3191, 906, 4763, 946, 53363, 24597, 803, 1377, 523, 2439, 1701, 1701, 1701, 859, 64096, 2507, 6905, 896, 779, 1394, 2396, 3314, 881, 810, 2166, 856, 896, 1023, 803, 5690, 1400, 5611, 810, 1192, 62975, 1607, 3015, 1082, 779, 26695, 570, 843, 888, 1023, 1188, 13690, 518, 2059, 916, 779, 59362, 1008, 810, 779, 24408, 797, 1192, 803, 4367, 810, 1312, 7519, 906, 4763, 1044, 3762, 18559, 13909, 587, 18588, 4483, 27104, 1288, 1413, 1166, 797, 1081, 896, 12856, 811, 10268, 779, 10669, 803, 1494, 3374, 1144, 23089, 811, 874, 797, 1973, 2184, 1177, 521, 1023, 1499, 1014, 1113, 1033, 2158, 1052, 1607, 14908, 3800, 570, 779, 12762, 800, 803, 1518, 13967, 1144, 803, 27569, 810, 2084, 1514, 1007, 2344, 30171, 1081, 521, 906, 968, 779, 2006, 36159, 896, 1861, 6283, 4650, 1529, 15676, 876, 13909, 511, 569, 587, 1424, 35247, 53673, 2417, 3058, 521, 889, 1449, 33262, 1798, 36116, 9874, 779, 31196, 803, 4367, 810, 779, 15602, 2100, 583, 14692, 521, 730, 11909, 527, 13699, 4486, 2773, 779, 6157, 803, 7793, 53673, 2417, 3058, 521, 5690, 953, 7962, 3222, 811, 511, 602, 18, 28828, 7832, 16066, 13041, 779, 3952, 540, 19, 16, 768, 535, 8323, 779, 5271, 810, 24408, 12708, 1419, 52096, 523, 20, 19, 16, 794, 535, 45389, 2611, 4863, 803, 6202, 797, 779, 45034, 3935, 523, 20, 19, 16, 781, 535, 6122, 2993, 19087, 3256, 779, 13874, 53236, 803, 3859, 4367, 523, 20, 19, 16, 778, 535, 7730, 779, 14127, 6547, 22892, 523, 20, 21 ], "markers": [ 283, 297, 312, 328 ], "logits_raw": [ -4.4359869956970215, -3.294753313064575, 1.9535304307937622, -4.301299571990967 ], "probs_raw": [ 0.0016643258277326822, 0.005210384260863066, 0.9912209510803223, 0.0019042867934331298 ], "probs": [ 0.009964779019355774, 0.02251727692782879, 0.9565467834472656, 0.010971151292324066 ], "argmax_index": 2, "argmax": "c", "top2_margin": 0.9340295065194368, "near_tie": false, "gold": "c", "correct": true, "host_softmax_max_abs_dp": 5.960464477539063e-08, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "answer": { "type": "choice", "choice": "c", "confidence": 0.9565467834472656, "probabilities": { "a": 0.009964779019355774, "b": 0.02251727692782879, "c": 0.9565467834472656, "d": 0.010971151292324066 } } }, "usage": { "input_tokens": 339, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "tv4_010", "mode": "text", "source": "transfer_v4_dev_head", "native": true, "prefix": 0, "usage_input_tokens": 84, "questions": [ { "qid": "answer", "type": "choice", "K": 4, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 84, "positions": 84, "bucket": 256, "ids": [ 1, 17, 38685, 42505, 6881, 997, 586, 3762, 965, 12993, 1878, 997, 39420, 6881, 997, 28828, 28763, 6142, 768, 2129, 598, 5508, 2160, 34918, 6557, 797, 779, 12681, 1452, 31915, 43084, 3504, 535, 941, 923, 62401, 803, 19821, 516, 8523, 602, 18, 28828, 7832, 16066, 13041, 779, 3952, 540, 19, 16, 768, 535, 12747, 920, 582, 6998, 3750, 20, 19, 16, 794, 535, 34483, 31191, 20, 19, 16, 781, 535, 984, 1380, 882, 1491, 2774, 20, 19, 16, 778, 535, 30360, 20227, 20, 21 ], "markers": [ 50, 60, 67, 77 ], "logits_raw": [ -1.631797432899475, -2.068960428237915, -1.538092017173767, -1.4344347715377808 ], "probs_raw": [ 0.2523794174194336, 0.1630033254623413, 0.2771722078323364, 0.3074450194835663 ], "probs": [ 0.25301024317741394, 0.18514660000801086, 0.2705257534980774, 0.2913174331188202 ], "argmax_index": 3, "argmax": "d", "top2_margin": 0.020791679620742798, "near_tie": false, "gold": "d", "correct": true, "host_softmax_max_abs_dp": 1.4901161193847656e-08, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "answer": { "type": "choice", "choice": "d", "confidence": 0.2913174331188202, "probabilities": { "a": 0.25301024317741394, "b": 0.18514660000801086, "c": 0.2705257534980774, "d": 0.2913174331188202 } } }, "usage": { "input_tokens": 84, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "tv4_011", "mode": "text", "source": "transfer_v4_dev_head", "native": true, "prefix": 0, "usage_input_tokens": 125, "questions": [ { "qid": "answer", "type": "choice", "K": 4, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 125, "positions": 125, "bucket": 256, "ids": [ 1, 17, 38685, 42505, 6881, 997, 2293, 3810, 5354, 4575, 1878, 997, 39420, 6881, 997, 28828, 803, 779, 2773, 15450, 1337, 17924, 13236, 12825, 4167, 2783, 20312, 856, 843, 1072, 518, 3347, 540, 569, 587, 550, 523, 38679, 875, 17924, 61359, 3675, 1209, 1394, 7441, 1399, 2025, 875, 12825, 61359, 13909, 587, 2667, 523, 31498, 11236, 4167, 769, 41913, 938, 946, 10305, 13033, 13909, 587, 7986, 523, 31498, 11236, 4167, 769, 41913, 3976, 1052, 768, 3437, 1782, 3955, 810, 5321, 2374, 905, 2460, 21321, 2927, 602, 18, 28828, 7832, 16066, 13041, 779, 3952, 540, 19, 16, 768, 535, 859, 1550, 20, 19, 16, 794, 535, 3504, 1550, 20, 19, 16, 781, 535, 6572, 1550, 20, 19, 16, 778, 535, 3504, 810, 6572, 1550, 20, 21 ], "markers": [ 95, 102, 109, 116 ], "logits_raw": [ -3.118961811065674, -3.313667058944702, -3.6159772872924805, -1.2328732013702393 ], "probs_raw": [ 0.11080395430326462, 0.09120020270347595, 0.06740687042474747, 0.730588972568512 ], "probs": [ 0.15580348670482635, 0.13557299971580505, 0.10924122482538223, 0.5993822813034058 ], "argmax_index": 3, "argmax": "d", "top2_margin": 0.4435787945985794, "near_tie": false, "gold": "b", "correct": false, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "answer": { "type": "choice", "choice": "d", "confidence": 0.5993822813034058, "probabilities": { "a": 0.15580348670482635, "b": 0.13557299971580505, "c": 0.10924122482538223, "d": 0.5993822813034058 } } }, "usage": { "input_tokens": 125, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "tv4_012", "mode": "text", "source": "transfer_v4_dev_head", "native": true, "prefix": 0, "usage_input_tokens": 97, "questions": [ { "qid": "answer", "type": "choice", "K": 4, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 97, "positions": 97, "bucket": 256, "ids": [ 1, 17, 38685, 42505, 6881, 997, 17756, 2254, 4167, 578, 2499, 1504, 1878, 997, 39420, 6881, 997, 6626, 31243, 5068, 768, 9920, 797, 1450, 8292, 521, 1689, 4302, 28204, 1209, 7302, 16830, 55324, 810, 13962, 1944, 592, 4245, 523, 1890, 31243, 521, 1944, 592, 4245, 511, 602, 18, 28828, 7832, 16066, 13041, 779, 3952, 540, 19, 16, 768, 535, 58969, 2718, 779, 3036, 803, 6154, 20, 19, 16, 794, 535, 938, 768, 7411, 2546, 20, 19, 16, 781, 535, 938, 768, 35281, 2184, 916, 7302, 16830, 20, 19, 16, 778, 535, 938, 902, 20451, 2184, 20, 21 ], "markers": [ 56, 67, 76, 88 ], "logits_raw": [ 1.5743370056152344, -2.8246326446533203, -0.30558764934539795, -1.4633177518844604 ], "probs_raw": [ 0.8245118856430054, 0.010133247822523117, 0.12582182884216309, 0.03953303396701813 ], "probs": [ 0.70499187707901, 0.030442405492067337, 0.18406431376934052, 0.08050142228603363 ], "argmax_index": 0, "argmax": "a", "top2_margin": 0.5209275633096695, "near_tie": false, "gold": "d", "correct": false, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "answer": { "type": "choice", "choice": "a", "confidence": 0.70499187707901, "probabilities": { "a": 0.70499187707901, "b": 0.030442405492067337, "c": 0.18406431376934052, "d": 0.08050142228603363 } } }, "usage": { "input_tokens": 97, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "tv4_013", "mode": "text", "source": "transfer_v4_dev_head", "native": true, "prefix": 0, "usage_input_tokens": 170, "questions": [ { "qid": "answer", "type": "choice", "K": 4, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 170, "positions": 170, "bucket": 256, "ids": [ 1, 17, 38685, 42505, 6881, 997, 51584, 6368, 1878, 997, 39420, 6881, 997, 1286, 2642, 900, 2774, 810, 39279, 581, 1172, 4066, 884, 3392, 10598, 6180, 10725, 521, 3921, 2177, 1390, 7075, 3618, 779, 22515, 4348, 988, 3822, 8523, 602, 18, 28828, 7832, 16066, 13041, 779, 3952, 540, 19, 16, 768, 535, 2342, 7075, 22792, 779, 9397, 988, 1993, 779, 11811, 561, 25900, 521, 1144, 856, 5175, 811, 779, 22515, 4348, 523, 20, 19, 16, 794, 535, 2342, 7075, 3807, 779, 784, 810, 9117, 1002, 1860, 896, 938, 1593, 521, 1101, 902, 20639, 1126, 4584, 2111, 779, 9397, 1172, 3347, 784, 810, 9117, 4396, 20, 19, 16, 781, 535, 2342, 7075, 7226, 6229, 779, 781, 15479, 1885, 1994, 1908, 36446, 521, 5430, 22813, 2844, 779, 14008, 1685, 902, 20639, 1126, 1172, 6136, 810, 779, 22515, 6825, 20, 19, 16, 778, 535, 2342, 7075, 16038, 768, 7226, 3685, 803, 1394, 811, 779, 1908, 25704, 2888, 811, 779, 37708, 810, 9707, 968, 779, 1390, 7075, 7226, 2536, 877, 521, 1144, 20, 21 ], "markers": [ 49, 74, 106, 138 ], "logits_raw": [ -0.35539257526397705, 0.6858941316604614, -0.02686423808336258, 0.8678652048110962 ], "probs_raw": [ 0.11600898206233978, 0.32863715291023254, 0.1611274927854538, 0.39422640204429626 ], "probs": [ 0.14782999455928802, 0.31103119254112244, 0.18693166971206665, 0.35420718789100647 ], "argmax_index": 3, "argmax": "d", "top2_margin": 0.04317599534988403, "near_tie": false, "gold": "c", "correct": false, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "answer": { "type": "choice", "choice": "d", "confidence": 0.35420718789100647, "probabilities": { "a": 0.14782999455928802, "b": 0.31103119254112244, "c": 0.18693166971206665, "d": 0.35420718789100647 } } }, "usage": { "input_tokens": 170, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "tv4_014", "mode": "text", "source": "transfer_v4_dev_head", "native": true, "prefix": 0, "usage_input_tokens": 98, "questions": [ { "qid": "answer", "type": "choice", "K": 4, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 98, "positions": 98, "bucket": 256, "ids": [ 1, 17, 38685, 42505, 6881, 997, 17756, 2254, 8496, 21671, 2499, 1504, 1878, 997, 39420, 6881, 997, 28828, 803, 779, 2773, 4744, 1178, 1295, 1862, 884, 27565, 8523, 602, 18, 28828, 7832, 16066, 13041, 779, 3952, 540, 19, 16, 768, 535, 2297, 7909, 768, 730, 2081, 532, 5204, 14785, 988, 768, 6709, 523, 20, 19, 16, 794, 535, 941, 5455, 803, 34736, 4228, 3809, 768, 6671, 523, 20, 19, 16, 781, 535, 6066, 6190, 1825, 768, 4072, 6073, 896, 5535, 1901, 12374, 1523, 16711, 523, 20, 19, 16, 778, 535, 835, 8117, 1675, 1264, 13505, 523, 20, 21 ], "markers": [ 38, 55, 69, 87 ], "logits_raw": [ -1.4280264377593994, -4.1390509605407715, -3.740838050842285, -2.7201414108276367 ], "probs_raw": [ 0.6943767070770264, 0.04615429788827896, 0.06873118132352829, 0.190737783908844 ], "probs": [ 0.5769809484481812, 0.08319791406393051, 0.11057352274656296, 0.22924762964248657 ], "argmax_index": 0, "argmax": "a", "top2_margin": 0.3477333188056946, "near_tie": false, "gold": "a", "correct": true, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "answer": { "type": "choice", "choice": "a", "confidence": 0.5769809484481812, "probabilities": { "a": 0.5769809484481812, "b": 0.08319791406393051, "c": 0.11057352274656296, "d": 0.22924762964248657 } } }, "usage": { "input_tokens": 98, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "tv4_015", "mode": "text", "source": "transfer_v4_dev_head", "native": true, "prefix": 0, "usage_input_tokens": 191, "questions": [ { "qid": "answer", "type": "choice", "K": 4, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 191, "positions": 191, "bucket": 256, "ids": [ 1, 17, 38685, 42505, 6881, 997, 30327, 16964, 3036, 1878, 997, 39420, 6881, 997, 542, 8117, 5006, 953, 12260, 2681, 988, 768, 6999, 523, 1405, 1320, 1381, 11249, 8693, 963, 779, 6999, 810, 953, 1943, 57778, 1048, 523, 15504, 8167, 12900, 13525, 779, 5006, 810, 38995, 1889, 523, 3145, 950, 17436, 811, 5642, 1523, 803, 1149, 4639, 521, 779, 8167, 3151, 779, 5006, 797, 779, 1892, 12799, 803, 768, 8167, 1681, 523, 941, 8167, 3191, 779, 5006, 797, 779, 1892, 12799, 803, 779, 8167, 1681, 875, 8174, 1235, 8831, 521, 810, 1757, 5789, 1889, 523, 2182, 779, 5006, 60153, 768, 4673, 2723, 779, 8167, 875, 4831, 46383, 521, 950, 1188, 1529, 3849, 511, 602, 18, 28828, 7832, 16066, 13041, 779, 3952, 540, 19, 16, 768, 535, 64155, 521, 10967, 779, 8167, 2047, 768, 6612, 18626, 523, 20, 19, 16, 794, 535, 64155, 521, 1288, 950, 1348, 12034, 811, 779, 53600, 1929, 950, 953, 57778, 1048, 523, 20, 19, 16, 781, 535, 1014, 64155, 521, 1929, 8167, 4584, 874, 4908, 50359, 875, 4831, 46383, 523, 20, 19, 16, 778, 535, 1014, 64155, 521, 1929, 779, 5006, 14759, 1295, 5808, 523, 20, 21 ], "markers": [ 123, 138, 158, 176 ], "logits_raw": [ 0.18301290273666382, -1.1472127437591553, 0.13043098151683807, -0.8381863832473755 ], "probs_raw": [ 0.3885974586009979, 0.10275200009346008, 0.36869218945503235, 0.13995833694934845 ], "probs": [ 0.3531140089035034, 0.13653221726417542, 0.3400966227054596, 0.17025713622570038 ], "argmax_index": 0, "argmax": "a", "top2_margin": 0.013017386198043823, "near_tie": true, "gold": "a", "correct": true, "host_softmax_max_abs_dp": 2.9802322387695312e-08, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "answer": { "type": "choice", "choice": "a", "confidence": 0.3531140089035034, "probabilities": { "a": 0.3531140089035034, "b": 0.13653221726417542, "c": 0.3400966227054596, "d": 0.17025713622570038 } } }, "usage": { "input_tokens": 191, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "tv4_016", "mode": "text", "source": "transfer_v4_dev_head", "native": true, "prefix": 0, "usage_input_tokens": 93, "questions": [ { "qid": "answer", "type": "choice", "K": 4, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 93, "positions": 93, "bucket": 256, "ids": [ 1, 17, 38685, 42505, 6881, 997, 54717, 61895, 1088, 1878, 997, 39420, 6881, 997, 28828, 32630, 4174, 856, 779, 1529, 45796, 916, 768, 7243, 2193, 893, 44352, 803, 22355, 8523, 602, 18, 28828, 7832, 16066, 13041, 779, 3952, 540, 19, 16, 768, 535, 27619, 14289, 27952, 6368, 523, 20, 19, 16, 794, 535, 27619, 14289, 856, 768, 2780, 40380, 2511, 968, 1144, 811, 3618, 933, 4859, 12383, 523, 20, 19, 16, 781, 535, 27619, 14289, 27952, 4297, 6155, 523, 20, 19, 16, 778, 535, 27619, 14289, 1753, 6812, 779, 12383, 523, 20, 21 ], "markers": [ 40, 50, 70, 81 ], "logits_raw": [ -3.956939458847046, -2.1060941219329834, -4.5931596755981445, -2.959867238998413 ], "probs_raw": [ 0.09429672360420227, 0.6002172827720642, 0.04991023987531662, 0.25557568669319153 ], "probs": [ 0.13468477129936218, 0.5052562355995178, 0.08549527823925018, 0.274563729763031 ], "argmax_index": 1, "argmax": "b", "top2_margin": 0.23069250583648682, "near_tie": false, "gold": "d", "correct": false, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "answer": { "type": "choice", "choice": "b", "confidence": 0.5052562355995178, "probabilities": { "a": 0.13468477129936218, "b": 0.5052562355995178, "c": 0.08549527823925018, "d": 0.274563729763031 } } }, "usage": { "input_tokens": 93, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "tv4_017", "mode": "text", "source": "transfer_v4_dev_head", "native": true, "prefix": 0, "usage_input_tokens": 87, "questions": [ { "qid": "answer", "type": "choice", "K": 4, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 87, "positions": 87, "bucket": 256, "ids": [ 1, 17, 38685, 42505, 6881, 997, 43046, 1878, 997, 39420, 6881, 997, 20581, 1316, 938, 29286, 1753, 23606, 587, 511, 602, 18, 28828, 7832, 16066, 13041, 779, 3952, 540, 19, 16, 768, 535, 29564, 768, 1847, 5337, 811, 779, 2280, 803, 779, 2832, 20, 19, 16, 794, 535, 38084, 779, 2043, 810, 2337, 1209, 6886, 20, 19, 16, 781, 535, 49782, 1149, 3426, 2915, 13105, 521, 3761, 1149, 4099, 2119, 810, 4859, 1149, 8660, 1394, 20, 19, 16, 778, 535, 38084, 1209, 8123, 810, 11346, 20, 21 ], "markers": [ 30, 45, 57, 77 ], "logits_raw": [ -1.8940390348434448, -2.765936851501465, 6.657193660736084, -0.15114456415176392 ], "probs_raw": [ 0.00019304058514535427, 8.072132186498493e-05, 0.9986231327056885, 0.0011030061868950725 ], "probs": [ 0.002199452603235841, 0.0011798404157161713, 0.9889820218086243, 0.007638671435415745 ], "argmax_index": 2, "argmax": "c", "top2_margin": 0.9813433503732085, "near_tie": false, "gold": "c", "correct": true, "host_softmax_max_abs_dp": 4.656612873077393e-10, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "answer": { "type": "choice", "choice": "c", "confidence": 0.9889820218086243, "probabilities": { "a": 0.002199452603235841, "b": 0.0011798404157161713, "c": 0.9889820218086243, "d": 0.007638671435415745 } } }, "usage": { "input_tokens": 87, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "tv4_018", "mode": "text", "source": "transfer_v4_dev_head", "native": true, "prefix": 0, "usage_input_tokens": 77, "questions": [ { "qid": "answer", "type": "choice", "K": 4, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 77, "positions": 77, "bucket": 256, "ids": [ 1, 17, 38685, 42505, 6881, 997, 17756, 2254, 18849, 1878, 997, 39420, 6881, 997, 28828, 803, 779, 2773, 11335, 811, 779, 7691, 803, 3785, 902, 10077, 6872, 811, 4859, 19553, 54924, 8523, 602, 18, 28828, 7832, 16066, 13041, 779, 3952, 540, 19, 16, 768, 535, 5993, 5134, 1668, 5654, 803, 17916, 20, 19, 16, 794, 535, 47836, 25556, 800, 20, 19, 16, 781, 535, 7996, 1580, 783, 16454, 20, 19, 16, 778, 535, 4885, 4377, 20, 21 ], "markers": [ 42, 53, 61, 70 ], "logits_raw": [ -0.7562735080718994, -1.111168622970581, 1.7879376411437988, -3.8880667686462402 ], "probs_raw": [ 0.0690699890255928, 0.04843512922525406, 0.8794806599617004, 0.003014178015291691 ], "probs": [ 0.12439634650945663, 0.09654004871845245, 0.7657828330993652, 0.013280742801725864 ], "argmax_index": 2, "argmax": "c", "top2_margin": 0.6413864865899086, "near_tie": false, "gold": "c", "correct": true, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "answer": { "type": "choice", "choice": "c", "confidence": 0.7657828330993652, "probabilities": { "a": 0.12439634650945663, "b": 0.09654004871845245, "c": 0.7657828330993652, "d": 0.013280742801725864 } } }, "usage": { "input_tokens": 77, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "tv4_019", "mode": "text", "source": "transfer_v4_dev_head", "native": true, "prefix": 0, "usage_input_tokens": 145, "questions": [ { "qid": "answer", "type": "choice", "K": 4, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 145, "positions": 145, "bucket": 256, "ids": [ 1, 17, 38685, 42505, 6881, 997, 586, 8988, 21469, 1878, 997, 39420, 6881, 997, 3111, 1144, 803, 1518, 1547, 21469, 2177, 779, 2120, 3338, 843, 18275, 5582, 859, 524, 1477, 524, 3462, 518, 1112, 3503, 8870, 61611, 7871, 521, 4262, 811, 17027, 11241, 8079, 797, 779, 2641, 906, 803, 730, 1718, 525, 540, 3108, 31933, 730, 526, 2171, 859, 953, 19049, 811, 8498, 896, 1727, 3801, 31396, 1320, 1014, 17502, 1626, 1210, 7927, 523, 3108, 31933, 730, 527, 2171, 859, 2588, 730, 20397, 810, 4468, 896, 1727, 9652, 953, 14209, 37168, 963, 1450, 2681, 2927, 602, 18, 28828, 7832, 16066, 13041, 779, 3952, 540, 19, 16, 768, 535, 923, 35736, 521, 923, 35736, 20, 19, 16, 794, 535, 923, 35736, 521, 4524, 7871, 20, 19, 16, 781, 535, 4524, 7871, 521, 923, 35736, 20, 19, 16, 778, 535, 4524, 7871, 521, 4524, 7871, 20, 21 ], "markers": [ 105, 115, 125, 135 ], "logits_raw": [ -0.6472692489624023, -2.074370861053467, -1.1838890314102173, -1.2575846910476685 ], "probs_raw": [ 0.4223143458366394, 0.10135693848133087, 0.24693623185157776, 0.22939249873161316 ], "probs": [ 0.3718819320201874, 0.1341748833656311, 0.2534708082675934, 0.24047236144542694 ], "argmax_index": 0, "argmax": "a", "top2_margin": 0.118411123752594, "near_tie": false, "gold": "a", "correct": true, "host_softmax_max_abs_dp": 2.9802322387695312e-08, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "answer": { "type": "choice", "choice": "a", "confidence": 0.3718819320201874, "probabilities": { "a": 0.3718819320201874, "b": 0.1341748833656311, "c": 0.2534708082675934, "d": 0.24047236144542694 } } }, "usage": { "input_tokens": 145, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "tv4_020", "mode": "text", "source": "transfer_v4_dev_head", "native": true, "prefix": 0, "usage_input_tokens": 100, "questions": [ { "qid": "answer", "type": "choice", "K": 4, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 100, "positions": 100, "bucket": 256, "ids": [ 1, 17, 38685, 42505, 6881, 997, 30327, 16964, 18849, 1878, 997, 39420, 6881, 997, 18588, 9467, 811, 904, 13011, 810, 2342, 20052, 1008, 1090, 3237, 1430, 11167, 2801, 521, 43175, 18588, 870, 3071, 4171, 36020, 18588, 18588, 856, 768, 1947, 803, 779, 19485, 1172, 535, 18588, 511, 602, 18, 28828, 7832, 16066, 13041, 779, 3952, 540, 19, 16, 768, 535, 5937, 803, 6450, 2917, 523, 20, 19, 16, 794, 535, 16798, 523, 20, 19, 16, 781, 535, 5051, 810, 41519, 811, 11282, 10912, 523, 20, 19, 16, 778, 535, 15209, 1337, 779, 5564, 810, 2536, 803, 1336, 523, 20, 21 ], "markers": [ 57, 67, 74, 86 ], "logits_raw": [ -0.6914830803871155, 0.4765990376472473, 1.2798019647598267, -2.3174314498901367 ], "probs_raw": [ 0.08626310527324677, 0.27740657329559326, 0.6193602681159973, 0.016970114782452583 ], "probs": [ 0.12978744506835938, 0.2989580035209656, 0.5306280255317688, 0.040626559406518936 ], "argmax_index": 2, "argmax": "c", "top2_margin": 0.23167002201080322, "near_tie": false, "gold": "c", "correct": true, "host_softmax_max_abs_dp": 2.9802322387695312e-08, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "answer": { "type": "choice", "choice": "c", "confidence": 0.5306280255317688, "probabilities": { "a": 0.12978744506835938, "b": 0.2989580035209656, "c": 0.5306280255317688, "d": 0.040626559406518936 } } }, "usage": { "input_tokens": 100, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "tv4_021", "mode": "text", "source": "transfer_v4_dev_head", "native": true, "prefix": 0, "usage_input_tokens": 75, "questions": [ { "qid": "answer", "type": "choice", "K": 4, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 75, "positions": 75, "bucket": 256, "ids": [ 1, 17, 38685, 42505, 6881, 997, 50988, 4331, 1878, 997, 39420, 6881, 997, 542, 16426, 2258, 1011, 874, 1593, 875, 17912, 602, 18, 28828, 7832, 16066, 13041, 779, 3952, 540, 19, 16, 768, 535, 14201, 23551, 878, 5077, 523, 20, 19, 16, 794, 535, 1569, 6660, 878, 5077, 523, 20, 19, 16, 781, 535, 51192, 4149, 2363, 779, 878, 4315, 523, 20, 19, 16, 778, 535, 1485, 11478, 803, 779, 878, 4315, 523, 20, 21 ], "markers": [ 31, 41, 51, 63 ], "logits_raw": [ -1.173835039138794, -0.4603675603866577, -1.0826773643493652, -0.9287775754928589 ], "probs_raw": [ 0.18469972908496857, 0.37698256969451904, 0.20232780277729034, 0.23598989844322205 ], "probs": [ 0.20311520993709564, 0.33812984824180603, 0.21678167581558228, 0.24197322130203247 ], "argmax_index": 1, "argmax": "b", "top2_margin": 0.09615662693977356, "near_tie": false, "gold": "c", "correct": false, "host_softmax_max_abs_dp": 2.9802322387695312e-08, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "answer": { "type": "choice", "choice": "b", "confidence": 0.33812984824180603, "probabilities": { "a": 0.20311520993709564, "b": 0.33812984824180603, "c": 0.21678167581558228, "d": 0.24197322130203247 } } }, "usage": { "input_tokens": 75, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "tv4_022", "mode": "text", "source": "transfer_v4_dev_head", "native": true, "prefix": 0, "usage_input_tokens": 148, "questions": [ { "qid": "answer", "type": "choice", "K": 4, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 148, "positions": 148, "bucket": 256, "ids": [ 1, 17, 38685, 42505, 6881, 997, 17756, 2254, 17263, 1878, 997, 39420, 6881, 997, 542, 14060, 2312, 9494, 779, 2773, 2203, 1337, 2204, 24190, 803, 62523, 523, 32989, 856, 1058, 2198, 13264, 514, 513, 1209, 11932, 1399, 37436, 810, 1178, 1058, 1698, 12185, 514, 513, 2442, 7460, 1399, 3735, 867, 523, 3735, 867, 856, 1058, 1698, 12185, 514, 513, 2442, 11932, 1399, 32989, 810, 1178, 1058, 2625, 12185, 3088, 600, 526, 2945, 528, 602, 13264, 514, 513, 1209, 7460, 1399, 37436, 523, 37436, 7025, 1763, 569, 513, 526, 523, 1056, 513, 1094, 6232, 803, 7460, 523, 3747, 856, 779, 2080, 803, 37843, 1094, 6232, 803, 7460, 803, 3735, 867, 8523, 602, 18, 28828, 7832, 16066, 13041, 779, 3952, 540, 19, 16, 768, 535, 730, 2475, 20, 19, 16, 794, 535, 730, 6505, 20, 19, 16, 781, 535, 730, 2198, 20, 19, 16, 778, 535, 730, 16511, 20, 21 ], "markers": [ 120, 127, 134, 141 ], "logits_raw": [ -1.871951937675476, -2.1734390258789062, -2.9600579738616943, -2.754812240600586 ], "probs_raw": [ 0.4015790820121765, 0.2970550060272217, 0.13527347147464752, 0.1660923957824707 ], "probs": [ 0.35737961530685425, 0.28813666105270386, 0.16427196562290192, 0.19021180272102356 ], "argmax_index": 0, "argmax": "a", "top2_margin": 0.06924295425415039, "near_tie": false, "gold": "a", "correct": true, "host_softmax_max_abs_dp": 2.9802322387695312e-08, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "answer": { "type": "choice", "choice": "a", "confidence": 0.35737961530685425, "probabilities": { "a": 0.35737961530685425, "b": 0.28813666105270386, "c": 0.16427196562290192, "d": 0.19021180272102356 } } }, "usage": { "input_tokens": 148, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "tv4_023", "mode": "text", "source": "transfer_v4_dev_head", "native": true, "prefix": 0, "usage_input_tokens": 104, "questions": [ { "qid": "answer", "type": "choice", "K": 4, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 104, "positions": 104, "bucket": 256, "ids": [ 1, 17, 38685, 42505, 6881, 997, 586, 8988, 38419, 1878, 997, 39420, 6881, 997, 1030, 59507, 886, 1090, 3160, 521, 779, 1509, 810, 43997, 3801, 7069, 768, 5516, 803, 1494, 5962, 3775, 856, 768, 4484, 12487, 896, 1494, 13010, 856, 1014, 10349, 3077, 933, 1883, 1203, 4286, 511, 602, 18, 28828, 7832, 16066, 13041, 779, 3952, 540, 19, 16, 768, 535, 1429, 13985, 2425, 803, 11244, 523, 20, 19, 16, 794, 535, 902, 6807, 15372, 1587, 5158, 3054, 523, 20, 19, 16, 781, 535, 1429, 29501, 2425, 803, 11244, 523, 20, 19, 16, 778, 535, 779, 25369, 803, 1113, 4624, 3311, 523, 20, 21 ], "markers": [ 58, 69, 81, 92 ], "logits_raw": [ -2.9687066078186035, 0.7563834190368652, -3.7341668605804443, -6.6945295333862305 ], "probs_raw": [ 0.023275204002857208, 0.9653382897377014, 0.01082576159387827, 0.0005607756902463734 ], "probs": [ 0.06266120076179504, 0.8966939449310303, 0.03626834228634834, 0.004376488737761974 ], "argmax_index": 1, "argmax": "b", "top2_margin": 0.8340327441692352, "near_tie": false, "gold": "a", "correct": false, "host_softmax_max_abs_dp": 4.656612873077393e-10, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "answer": { "type": "choice", "choice": "b", "confidence": 0.8966939449310303, "probabilities": { "a": 0.06266120076179504, "b": 0.8966939449310303, "c": 0.03626834228634834, "d": 0.004376488737761974 } } }, "usage": { "input_tokens": 104, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "tv4_024", "mode": "text", "source": "transfer_v4_dev_head", "native": true, "prefix": 0, "usage_input_tokens": 119, "questions": [ { "qid": "answer", "type": "choice", "K": 4, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 119, "positions": 119, "bucket": 256, "ids": [ 1, 17, 38685, 42505, 6881, 997, 53983, 3609, 1878, 997, 39420, 6881, 997, 3493, 856, 779, 2536, 803, 768, 2375, 4734, 1785, 1786, 893, 28714, 797, 6368, 3609, 8523, 602, 18, 28828, 7832, 16066, 13041, 779, 3952, 540, 19, 16, 768, 535, 2989, 11787, 1081, 811, 2846, 779, 1336, 1327, 803, 4665, 3910, 5720, 592, 3191, 1230, 803, 6257, 2006, 36159, 884, 6368, 523, 20, 19, 16, 794, 535, 2989, 14057, 779, 4077, 803, 2410, 810, 1352, 6079, 521, 797, 5574, 11336, 1399, 6257, 3606, 523, 20, 19, 16, 781, 535, 2989, 8759, 779, 4588, 896, 10338, 14773, 810, 4109, 938, 779, 5592, 9715, 18572, 6368, 4008, 523, 20, 19, 16, 778, 535, 2595, 803, 1518, 523, 20, 21 ], "markers": [ 39, 66, 88, 110 ], "logits_raw": [ 1.9829537868499756, 1.969934344291687, -0.5511482954025269, 3.171776294708252 ], "probs_raw": [ 0.18692950904369354, 0.18451154232025146, 0.014829661697149277, 0.6137293577194214 ], "probs": [ 0.22260868549346924, 0.22054797410964966, 0.03642337769269943, 0.5204199552536011 ], "argmax_index": 3, "argmax": "d", "top2_margin": 0.29781126976013184, "near_tie": false, "gold": "d", "correct": true, "host_softmax_max_abs_dp": 1.4901161193847656e-08, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "answer": { "type": "choice", "choice": "d", "confidence": 0.5204199552536011, "probabilities": { "a": 0.22260868549346924, "b": 0.22054797410964966, "c": 0.03642337769269943, "d": 0.5204199552536011 } } }, "usage": { "input_tokens": 119, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "tv4_025", "mode": "text", "source": "transfer_v4_dev_head", "native": true, "prefix": 0, "usage_input_tokens": 192, "questions": [ { "qid": "answer", "type": "choice", "K": 4, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 192, "positions": 192, "bucket": 256, "ids": [ 1, 17, 38685, 42505, 6881, 997, 30327, 16964, 3036, 1878, 997, 39420, 6881, 997, 542, 45301, 20271, 968, 768, 6394, 797, 730, 1086, 526, 60843, 768, 13733, 803, 2496, 875, 768, 15879, 803, 1235, 23589, 521, 43152, 803, 1144, 953, 28984, 521, 43175, 1095, 1727, 6709, 875, 2105, 521, 1203, 1288, 58154, 856, 1760, 4997, 884, 779, 13733, 803, 2496, 521, 1757, 811, 1727, 30221, 810, 1210, 950, 14352, 521, 810, 1288, 875, 1523, 4092, 779, 2862, 2712, 2448, 60843, 811, 1727, 30221, 856, 1014, 6612, 521, 1757, 859, 21225, 779, 2084, 811, 840, 522, 9381, 779, 13733, 803, 2496, 810, 2337, 1892, 1727, 5401, 523, 43175, 1030, 730, 1086, 529, 521, 779, 6709, 5942, 14643, 931, 2076, 779, 18302, 950, 1320, 21440, 1738, 874, 33533, 523, 12747, 856, 18887, 811, 779, 40951, 803, 779, 13838, 8523, 602, 18, 28828, 7832, 16066, 13041, 779, 3952, 540, 19, 16, 768, 535, 941, 6709, 1090, 950, 14352, 523, 20, 19, 16, 794, 535, 941, 30221, 523, 20, 19, 16, 781, 535, 941, 6394, 523, 20, 19, 16, 778, 535, 889, 47348, 15395, 1685, 779, 6709, 1090, 950, 14352, 810, 779, 6394, 523, 20, 21 ], "markers": [ 147, 158, 166, 174 ], "logits_raw": [ -0.358635812997818, -0.6387908458709717, 0.28748929500579834, 0.5898016095161438 ], "probs_raw": [ 0.16011600196361542, 0.12099432200193405, 0.3055225908756256, 0.4133670926094055 ], "probs": [ 0.18607696890830994, 0.15232786536216736, 0.2952173948287964, 0.3663778007030487 ], "argmax_index": 3, "argmax": "d", "top2_margin": 0.07116040587425232, "near_tie": false, "gold": "a", "correct": false, "host_softmax_max_abs_dp": 2.9802322387695312e-08, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "answer": { "type": "choice", "choice": "d", "confidence": 0.3663778007030487, "probabilities": { "a": 0.18607696890830994, "b": 0.15232786536216736, "c": 0.2952173948287964, "d": 0.3663778007030487 } } }, "usage": { "input_tokens": 192, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "tv4_026", "mode": "text", "source": "transfer_v4_dev_head", "native": true, "prefix": 0, "usage_input_tokens": 105, "questions": [ { "qid": "answer", "type": "choice", "K": 4, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 105, "positions": 105, "bucket": 256, "ids": [ 1, 17, 38685, 42505, 6881, 997, 54717, 61895, 1088, 1878, 997, 39420, 6881, 997, 50902, 32600, 10723, 843, 7377, 560, 518, 856, 2492, 4965, 811, 843, 777, 916, 518, 2473, 59491, 523, 20463, 803, 779, 2773, 15450, 856, 55495, 8523, 602, 18, 28828, 7832, 16066, 13041, 779, 3952, 540, 19, 16, 768, 535, 9758, 938, 11483, 916, 779, 1452, 26196, 797, 4272, 12158, 20, 19, 16, 794, 535, 9758, 938, 5898, 21127, 786, 893, 810, 775, 1570, 1021, 20, 19, 16, 781, 535, 9758, 6378, 768, 19467, 37344, 523, 20, 19, 16, 778, 535, 9758, 5308, 811, 1640, 6033, 2563, 779, 3036, 523, 20, 21 ], "markers": [ 50, 65, 80, 91 ], "logits_raw": [ -8.456859588623047, -7.910188674926758, -4.601129531860352, -2.9162330627441406 ], "probs_raw": [ 0.0032805362716317177, 0.005667101591825485, 0.15504834055900574, 0.8360040187835693 ], "probs": [ 0.014177599921822548, 0.0209506768733263, 0.2227291464805603, 0.7421425580978394 ], "argmax_index": 3, "argmax": "d", "top2_margin": 0.519413411617279, "near_tie": false, "gold": "c", "correct": false, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "answer": { "type": "choice", "choice": "d", "confidence": 0.7421425580978394, "probabilities": { "a": 0.014177599921822548, "b": 0.0209506768733263, "c": 0.2227291464805603, "d": 0.7421425580978394 } } }, "usage": { "input_tokens": 105, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "tv4_027", "mode": "text", "source": "transfer_v4_dev_head", "native": true, "prefix": 0, "usage_input_tokens": 94, "questions": [ { "qid": "answer", "type": "choice", "K": 4, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 94, "positions": 94, "bucket": 256, "ids": [ 1, 17, 38685, 42505, 6881, 997, 17756, 2254, 8496, 21671, 2499, 1504, 1878, 997, 39420, 6881, 997, 561, 788, 20167, 810, 17683, 802, 511, 602, 18, 28828, 7832, 16066, 13041, 779, 3952, 540, 19, 16, 768, 535, 1725, 797, 3730, 10798, 11128, 523, 20, 19, 16, 794, 535, 5650, 9920, 810, 5650, 3730, 779, 3685, 803, 779, 2546, 4997, 52797, 1939, 523, 20, 19, 16, 781, 535, 3730, 779, 3685, 803, 779, 2546, 4997, 52797, 1939, 523, 20, 19, 16, 778, 535, 9920, 779, 3685, 803, 779, 2546, 4997, 52797, 1939, 523, 20, 21 ], "markers": [ 34, 45, 64, 79 ], "logits_raw": [ -1.4790819883346558, 3.206943988800049, -0.3642081022262573, -1.1758590936660767 ], "probs_raw": [ 0.008785426616668701, 0.9525288343429565, 0.02678837813436985, 0.011897370219230652 ], "probs": [ 0.03040599264204502, 0.8644078373908997, 0.06742636859416962, 0.03775973990559578 ], "argmax_index": 1, "argmax": "b", "top2_margin": 0.79698146879673, "near_tie": false, "gold": "c", "correct": false, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "answer": { "type": "choice", "choice": "b", "confidence": 0.8644078373908997, "probabilities": { "a": 0.03040599264204502, "b": 0.8644078373908997, "c": 0.06742636859416962, "d": 0.03775973990559578 } } }, "usage": { "input_tokens": 94, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "tv4_028", "mode": "text", "source": "transfer_v4_dev_head", "native": true, "prefix": 0, "usage_input_tokens": 319, "questions": [ { "qid": "answer", "type": "choice", "K": 4, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 319, "positions": 319, "bucket": 512, "ids": [ 1, 17, 38685, 42505, 6881, 997, 30327, 16964, 3036, 1878, 997, 39420, 6881, 997, 542, 64188, 856, 779, 9829, 27355, 803, 768, 5792, 12800, 968, 902, 12235, 4129, 2083, 2844, 779, 2105, 803, 1210, 7950, 521, 2166, 40918, 523, 941, 5792, 6034, 768, 32787, 7376, 896, 6527, 32886, 587, 960, 856, 4025, 1108, 797, 779, 3308, 896, 3295, 803, 779, 2083, 3107, 43175, 27012, 5828, 521, 810, 19848, 803, 1113, 1312, 5817, 521, 988, 40314, 8462, 1862, 791, 22333, 1626, 9774, 17844, 810, 45127, 3255, 523, 43175, 1098, 64188, 1090, 7950, 1827, 1450, 3295, 797, 779, 4598, 1047, 18660, 803, 1450, 4799, 523, 941, 3295, 12393, 988, 30080, 5457, 968, 779, 21384, 803, 768, 6527, 1985, 37249, 1859, 10231, 42605, 523, 941, 64188, 953, 20890, 810, 23548, 12608, 875, 779, 12494, 803, 1210, 7950, 523, 4162, 768, 42212, 8727, 521, 779, 64188, 953, 5930, 4477, 803, 779, 7786, 523, 4162, 779, 12235, 4129, 17436, 811, 4025, 779, 64188, 1090, 12235, 4673, 521, 779, 64188, 60509, 1033, 5614, 4272, 523, 941, 38203, 8753, 14340, 779, 5792, 521, 13621, 774, 779, 11445, 10674, 779, 3295, 803, 779, 2083, 3107, 521, 1563, 871, 896, 779, 3295, 953, 2638, 779, 5792, 23984, 875, 6527, 32886, 587, 960, 16655, 521, 810, 16365, 20524, 30525, 523, 2805, 8727, 521, 779, 12235, 4129, 1090, 26917, 48207, 811, 14993, 41240, 592, 988, 779, 64188, 1090, 12494, 8727, 523, 941, 64188, 1090, 26917, 7007, 811, 779, 11258, 803, 1559, 4589, 523, 941, 8727, 16277, 1753, 7330, 779, 899, 1016, 3501, 4589, 511, 602, 18, 28828, 7832, 16066, 13041, 779, 3952, 540, 19, 16, 768, 535, 6416, 41748, 906, 768, 31137, 2723, 2862, 523, 20, 19, 16, 794, 535, 6416, 41748, 906, 5211, 33763, 523, 20, 19, 16, 781, 535, 797, 51039, 41748, 521, 1929, 803, 3749, 15850, 1170, 31355, 523, 20, 19, 16, 778, 535, 797, 51039, 41748, 521, 1929, 803, 6527, 2250, 36115, 598, 523, 20, 21 ], "markers": [ 263, 276, 287, 303 ], "logits_raw": [ -1.302972435951233, 1.3318623304367065, -1.1114861965179443, -1.2292307615280151 ], "probs_raw": [ 0.058043066412210464, 0.80917888879776, 0.07029295712709427, 0.06248502433300018 ], "probs": [ 0.10237134993076324, 0.6723430156707764, 0.11737720668315887, 0.1079084724187851 ], "argmax_index": 1, "argmax": "b", "top2_margin": 0.5549658089876175, "near_tie": false, "gold": "b", "correct": true, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "answer": { "type": "choice", "choice": "b", "confidence": 0.6723430156707764, "probabilities": { "a": 0.10237134993076324, "b": 0.6723430156707764, "c": 0.11737720668315887, "d": 0.1079084724187851 } } }, "usage": { "input_tokens": 319, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "tv4_029", "mode": "text", "source": "transfer_v4_dev_head", "native": true, "prefix": 0, "usage_input_tokens": 74, "questions": [ { "qid": "answer", "type": "choice", "K": 4, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 74, "positions": 74, "bucket": 256, "ids": [ 1, 17, 38685, 42505, 6881, 997, 17756, 2254, 16290, 1878, 997, 39420, 6881, 997, 37512, 2574, 586, 3838, 7809, 797, 859, 5045, 785, 17868, 4027, 3504, 523, 49936, 582, 6572, 523, 50168, 38822, 592, 511, 602, 18, 28828, 7832, 16066, 13041, 779, 3952, 540, 19, 16, 768, 535, 859, 1550, 20, 19, 16, 794, 535, 3504, 1550, 20, 19, 16, 781, 535, 6572, 1550, 20, 19, 16, 778, 535, 859, 810, 6572, 20, 21 ], "markers": [ 45, 52, 59, 66 ], "logits_raw": [ -0.2669294476509094, -0.4548293352127075, -1.1632142066955566, 1.690333366394043 ], "probs_raw": [ 0.10733412951231003, 0.08894754946231842, 0.04380122572183609, 0.759917140007019 ], "probs": [ 0.155056893825531, 0.1355808526277542, 0.0817398950457573, 0.6276223063468933 ], "argmax_index": 3, "argmax": "d", "top2_margin": 0.4725654125213623, "near_tie": false, "gold": "d", "correct": true, "host_softmax_max_abs_dp": 7.450580596923828e-09, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "answer": { "type": "choice", "choice": "d", "confidence": 0.6276223063468933, "probabilities": { "a": 0.155056893825531, "b": 0.1355808526277542, "c": 0.0817398950457573, "d": 0.6276223063468933 } } }, "usage": { "input_tokens": 74, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "tv4_030", "mode": "text", "source": "transfer_v4_dev_head", "native": true, "prefix": 0, "usage_input_tokens": 91, "questions": [ { "qid": "answer", "type": "choice", "K": 4, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 91, "positions": 91, "bucket": 256, "ids": [ 1, 17, 38685, 42505, 6881, 997, 52220, 6028, 598, 1878, 997, 39420, 6881, 997, 549, 2995, 21179, 29375, 906, 17912, 602, 18, 28828, 7832, 16066, 13041, 779, 3952, 540, 19, 16, 768, 535, 768, 22028, 811, 33353, 20845, 523, 20, 19, 16, 794, 535, 768, 22028, 811, 22710, 11104, 523, 20, 19, 16, 781, 535, 12738, 6421, 4272, 933, 4183, 6602, 768, 5589, 1888, 768, 29647, 803, 2396, 575, 852, 523, 20, 19, 16, 778, 535, 12738, 6421, 4272, 933, 4183, 14635, 1337, 779, 2238, 11364, 875, 1113, 523, 20, 21 ], "markers": [ 30, 41, 52, 73 ], "logits_raw": [ -0.01371368020772934, -3.478639602661133, -2.0193262100219727, 0.3967164158821106 ], "probs_raw": [ 0.3740668296813965, 0.011699061840772629, 0.050341106951236725, 0.5638930201530457 ], "probs": [ 0.375444233417511, 0.03159448876976967, 0.08960627764463425, 0.5033549666404724 ], "argmax_index": 3, "argmax": "d", "top2_margin": 0.12791073322296143, "near_tie": false, "gold": "c", "correct": false, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "answer": { "type": "choice", "choice": "d", "confidence": 0.5033549666404724, "probabilities": { "a": 0.375444233417511, "b": 0.03159448876976967, "c": 0.08960627764463425, "d": 0.5033549666404724 } } }, "usage": { "input_tokens": 91, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "tv4_031", "mode": "text", "source": "transfer_v4_dev_head", "native": true, "prefix": 0, "usage_input_tokens": 79, "questions": [ { "qid": "answer", "type": "choice", "K": 4, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 79, "positions": 79, "bucket": 256, "ids": [ 1, 17, 38685, 42505, 6881, 997, 17756, 2254, 18849, 1878, 997, 39420, 6881, 997, 9467, 811, 33620, 2371, 521, 768, 3444, 8035, 1090, 21925, 10373, 14808, 2363, 511, 602, 18, 28828, 7832, 16066, 13041, 779, 3952, 540, 19, 16, 768, 535, 26368, 3384, 8199, 523, 19628, 20, 19, 16, 794, 535, 9552, 8199, 523, 3671, 20694, 20, 19, 16, 781, 535, 36218, 8199, 523, 30852, 810, 36025, 20, 19, 16, 778, 535, 5294, 8199, 523, 20451, 960, 20, 21 ], "markers": [ 38, 48, 58, 69 ], "logits_raw": [ -1.1906007528305054, 2.4247899055480957, -2.300215005874634, -2.701061248779297 ], "probs_raw": [ 0.025828873738646507, 0.9599524736404419, 0.008515420369803905, 0.005703229922801256 ], "probs": [ 0.06655921787023544, 0.8806865215301514, 0.030127935111522675, 0.022626303136348724 ], "argmax_index": 1, "argmax": "b", "top2_margin": 0.8141273036599159, "near_tie": false, "gold": "a", "correct": false, "host_softmax_max_abs_dp": 7.450580596923828e-09, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "answer": { "type": "choice", "choice": "b", "confidence": 0.8806865215301514, "probabilities": { "a": 0.06655921787023544, "b": 0.8806865215301514, "c": 0.030127935111522675, "d": 0.022626303136348724 } } }, "usage": { "input_tokens": 79, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "tv4_032", "mode": "text", "source": "transfer_v4_dev_head", "native": true, "prefix": 0, "usage_input_tokens": 145, "questions": [ { "qid": "answer", "type": "choice", "K": 4, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 145, "positions": 145, "bucket": 256, "ids": [ 1, 17, 38685, 42505, 6881, 997, 3034, 41336, 1878, 997, 39420, 6881, 997, 3493, 856, 779, 5521, 1685, 902, 13954, 5562, 810, 768, 7889, 5562, 8523, 602, 18, 28828, 7832, 16066, 13041, 779, 3952, 540, 19, 16, 768, 535, 1592, 13954, 5562, 856, 2538, 968, 7250, 521, 2158, 768, 7889, 5562, 19000, 988, 730, 1664, 811, 730, 1838, 1779, 523, 20, 19, 16, 794, 535, 835, 7889, 5562, 856, 10491, 988, 12255, 14506, 14820, 6793, 521, 2158, 902, 13954, 5562, 856, 1550, 3973, 916, 4985, 10621, 523, 20, 19, 16, 781, 535, 835, 7889, 5562, 11335, 811, 768, 1423, 933, 768, 3795, 803, 1779, 521, 2158, 902, 13954, 5562, 11136, 768, 17540, 2802, 2565, 523, 20, 19, 16, 778, 535, 1592, 13954, 5562, 11335, 811, 768, 1423, 933, 768, 3795, 803, 1779, 521, 2158, 768, 7889, 5562, 11136, 768, 17540, 2802, 2565, 523, 20, 21 ], "markers": [ 36, 62, 89, 117 ], "logits_raw": [ -2.510103464126587, -1.3151054382324219, 0.091059111058712, 0.8483920097351074 ], "probs_raw": [ 0.021492037922143936, 0.07100004702806473, 0.2896997630596161, 0.6178081631660461 ], "probs": [ 0.04813937842845917, 0.11303896456956863, 0.30865028500556946, 0.5301713943481445 ], "argmax_index": 3, "argmax": "d", "top2_margin": 0.22152110934257507, "near_tie": false, "gold": "d", "correct": true, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "answer": { "type": "choice", "choice": "d", "confidence": 0.5301713943481445, "probabilities": { "a": 0.04813937842845917, "b": 0.11303896456956863, "c": 0.30865028500556946, "d": 0.5301713943481445 } } }, "usage": { "input_tokens": 145, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "tv4_033", "mode": "text", "source": "transfer_v4_dev_head", "native": true, "prefix": 0, "usage_input_tokens": 308, "questions": [ { "qid": "answer", "type": "choice", "K": 4, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 308, "positions": 308, "bucket": 512, "ids": [ 1, 17, 38685, 42505, 6881, 997, 17756, 2254, 18807, 780, 3129, 1878, 997, 39420, 6881, 997, 2443, 3952, 11335, 811, 779, 2773, 2203, 13909, 587, 18588, 2299, 1090, 2743, 510, 1311, 1090, 18805, 963, 779, 7952, 9276, 1090, 6756, 3823, 511, 768, 27549, 2448, 54907, 963, 1123, 906, 950, 9657, 523, 859, 8107, 779, 44726, 889, 49994, 797, 1235, 43053, 523, 7356, 859, 2062, 779, 6282, 963, 768, 38455, 523, 7356, 43175, 1098, 5707, 6252, 803, 56312, 2274, 3742, 1846, 14231, 521, 4053, 730, 527, 1607, 2444, 529, 3787, 18588, 7356, 1311, 953, 902, 34760, 811, 768, 3847, 810, 768, 4311, 5271, 2217, 523, 7356, 3408, 510, 7356, 20385, 22720, 1203, 48109, 3085, 1048, 521, 859, 4549, 1892, 811, 1607, 3462, 15600, 1090, 6756, 570, 810, 21791, 1419, 779, 6756, 803, 29281, 52460, 1088, 1206, 521, 1494, 10292, 523, 43175, 3493, 856, 5242, 8600, 540, 18588, 950, 5751, 1123, 523, 43175, 2299, 1090, 22062, 521, 19398, 510, 18588, 43175, 2299, 1188, 874, 1113, 2084, 1757, 568, 3823, 511, 52460, 1088, 1206, 25723, 3787, 569, 587, 559, 62261, 889, 983, 7569, 2771, 521, 4162, 730, 1838, 21045, 521, 781, 523, 730, 1848, 530, 13699, 6314, 779, 12543, 521, 1235, 1400, 10953, 896, 889, 983, 7569, 2771, 33871, 896, 511, 602, 18, 28828, 7832, 16066, 13041, 779, 3952, 540, 19, 16, 768, 535, 19178, 2274, 875, 2104, 797, 730, 2444, 529, 953, 2581, 23073, 916, 2507, 14801, 19555, 1586, 797, 5242, 20, 19, 16, 794, 535, 19178, 2274, 875, 2104, 797, 730, 2444, 529, 953, 2581, 23073, 916, 2507, 14801, 19555, 1586, 1113, 3648, 3200, 20, 19, 16, 781, 535, 19178, 2274, 875, 2104, 797, 730, 2444, 529, 953, 2581, 23073, 916, 2507, 37875, 1113, 3648, 3200, 20, 19, 16, 778, 535, 19178, 2274, 875, 2104, 797, 730, 2444, 529, 953, 2581, 23073, 916, 2507, 37875, 797, 5242, 20, 21 ], "markers": [ 218, 241, 265, 287 ], "logits_raw": [ 0.10337462276220322, -2.55706524848938, -0.875851035118103, -0.2083251178264618 ], "probs_raw": [ 0.45919573307037354, 0.03210579603910446, 0.17247475683689117, 0.3362237215042114 ], "probs": [ 0.408709853887558, 0.06110251322388649, 0.2030612677335739, 0.3271264135837555 ], "argmax_index": 0, "argmax": "a", "top2_margin": 0.08158344030380249, "near_tie": false, "gold": "d", "correct": false, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "answer": { "type": "choice", "choice": "a", "confidence": 0.408709853887558, "probabilities": { "a": 0.408709853887558, "b": 0.06110251322388649, "c": 0.2030612677335739, "d": 0.3271264135837555 } } }, "usage": { "input_tokens": 308, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "tv4_034", "mode": "text", "source": "transfer_v4_dev_head", "native": true, "prefix": 0, "usage_input_tokens": 87, "questions": [ { "qid": "answer", "type": "choice", "K": 4, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 87, "positions": 87, "bucket": 256, "ids": [ 1, 17, 38685, 42505, 6881, 997, 17756, 2254, 18849, 1878, 997, 39420, 6881, 997, 542, 2514, 803, 7524, 797, 768, 5392, 2033, 938, 811, 874, 48586, 803, 5392, 875, 2759, 3378, 523, 4162, 1149, 2341, 35210, 36744, 4915, 521, 7524, 1400, 6305, 15603, 1144, 803, 779, 2773, 8523, 602, 18, 28828, 7832, 16066, 13041, 779, 3952, 540, 19, 16, 768, 535, 63183, 20, 19, 16, 794, 535, 1097, 1380, 20, 19, 16, 781, 535, 39393, 803, 24710, 20, 19, 16, 778, 535, 7329, 1617, 10021, 20, 21 ], "markers": [ 58, 64, 71, 79 ], "logits_raw": [ -2.793691396713257, -2.0432310104370117, -1.2673485279083252, 2.3498144149780273 ], "probs_raw": [ 0.005585509818047285, 0.011829973198473454, 0.02570071816444397, 0.9568836688995361 ], "probs": [ 0.022172139957547188, 0.037898775190114975, 0.06596741825342178, 0.8739617466926575 ], "argmax_index": 3, "argmax": "d", "top2_margin": 0.8079943284392357, "near_tie": false, "gold": "d", "correct": true, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "answer": { "type": "choice", "choice": "d", "confidence": 0.8739617466926575, "probabilities": { "a": 0.022172139957547188, "b": 0.037898775190114975, "c": 0.06596741825342178, "d": 0.8739617466926575 } } }, "usage": { "input_tokens": 87, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "tv4_035", "mode": "text", "source": "transfer_v4_dev_head", "native": true, "prefix": 0, "usage_input_tokens": 120, "questions": [ { "qid": "answer", "type": "choice", "K": 4, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 120, "positions": 120, "bucket": 256, "ids": [ 1, 17, 38685, 42505, 6881, 997, 578, 2499, 813, 8731, 1878, 997, 39420, 6881, 997, 3383, 1413, 1166, 768, 43186, 1862, 797, 4741, 521, 1620, 1499, 874, 779, 7128, 803, 779, 6196, 3616, 13818, 875, 906, 2801, 896, 34162, 875, 896, 43186, 8523, 602, 18, 28828, 7832, 16066, 13041, 779, 3952, 540, 19, 16, 768, 535, 1311, 1499, 7860, 1209, 6214, 875, 6384, 43368, 1399, 875, 4948, 6531, 803, 779, 2043, 17165, 20, 19, 16, 794, 535, 1311, 1499, 874, 13820, 14977, 1337, 9617, 20, 19, 16, 781, 535, 1311, 1499, 7860, 2442, 6214, 875, 6384, 43368, 1399, 875, 4948, 6531, 803, 779, 2043, 17165, 20, 19, 16, 778, 535, 1311, 1499, 874, 9617, 875, 1113, 4948, 43368, 20, 21 ], "markers": [ 53, 74, 86, 107 ], "logits_raw": [ 1.3277885913848877, -0.30665910243988037, -1.4691928625106812, -2.5176525115966797 ], "probs_raw": [ 0.7828211188316345, 0.1526971459388733, 0.04774731770157814, 0.016734344884753227 ], "probs": [ 0.661874532699585, 0.20592838525772095, 0.08975527435541153, 0.04244174808263779 ], "argmax_index": 0, "argmax": "a", "top2_margin": 0.455946147441864, "near_tie": false, "gold": "a", "correct": true, "host_softmax_max_abs_dp": 7.450580596923828e-09, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "answer": { "type": "choice", "choice": "a", "confidence": 0.661874532699585, "probabilities": { "a": 0.661874532699585, "b": 0.20592838525772095, "c": 0.08975527435541153, "d": 0.04244174808263779 } } }, "usage": { "input_tokens": 120, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "tv4_036", "mode": "text", "source": "transfer_v4_dev_head", "native": true, "prefix": 0, "usage_input_tokens": 150, "questions": [ { "qid": "answer", "type": "choice", "K": 4, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 150, "positions": 150, "bucket": 256, "ids": [ 1, 17, 38685, 42505, 6881, 997, 30327, 16964, 21167, 1878, 997, 39420, 6881, 997, 553, 26717, 13807, 730, 1086, 525, 19190, 803, 24008, 1969, 37891, 523, 2624, 9051, 896, 1166, 15399, 797, 1423, 730, 526, 963, 1058, 1155, 523, 1838, 1094, 5611, 523, 1030, 1423, 730, 529, 10765, 772, 5068, 768, 730, 530, 514, 2222, 522, 40864, 1108, 7533, 959, 803, 24008, 1969, 2624, 9051, 523, 1030, 1423, 730, 530, 779, 9051, 13305, 730, 527, 20432, 522, 526, 523, 1030, 779, 2835, 1423, 10765, 772, 4997, 730, 7765, 19190, 523, 3747, 856, 10765, 772, 1090, 6524, 797, 779, 730, 7765, 19190, 803, 9051, 4997, 8523, 602, 18, 28828, 7832, 16066, 13041, 779, 3952, 540, 19, 16, 768, 535, 1058, 529, 521, 1873, 20, 19, 16, 794, 535, 1058, 533, 521, 1873, 20, 19, 16, 781, 535, 1058, 533, 521, 6019, 20, 19, 16, 778, 535, 730, 10173, 1056, 20, 21 ], "markers": [ 115, 124, 133, 142 ], "logits_raw": [ -2.7359061241149902, -1.6822254657745361, -1.4564356803894043, -5.317482948303223 ], "probs_raw": [ 0.13265100121498108, 0.38046807050704956, 0.47684529423713684, 0.010035661049187183 ], "probs": [ 0.17315810918807983, 0.36756089329719543, 0.4318934977054596, 0.027387524023652077 ], "argmax_index": 2, "argmax": "c", "top2_margin": 0.06433260440826416, "near_tie": false, "gold": "a", "correct": false, "host_softmax_max_abs_dp": 1.862645149230957e-09, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "answer": { "type": "choice", "choice": "c", "confidence": 0.4318934977054596, "probabilities": { "a": 0.17315810918807983, "b": 0.36756089329719543, "c": 0.4318934977054596, "d": 0.027387524023652077 } } }, "usage": { "input_tokens": 150, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "tv4_037", "mode": "text", "source": "transfer_v4_dev_head", "native": true, "prefix": 0, "usage_input_tokens": 76, "questions": [ { "qid": "answer", "type": "choice", "K": 4, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 76, "positions": 76, "bucket": 256, "ids": [ 1, 17, 38685, 42505, 6881, 997, 5606, 30049, 1878, 997, 39420, 6881, 997, 28828, 803, 779, 2773, 856, 779, 1529, 3429, 15488, 803, 779, 29749, 8046, 1678, 8523, 602, 18, 28828, 7832, 16066, 13041, 779, 3952, 540, 19, 16, 768, 535, 44437, 6754, 20, 19, 16, 794, 535, 837, 2998, 20, 19, 16, 781, 535, 28718, 837, 2998, 20, 19, 16, 778, 535, 1976, 9529, 11755, 803, 1181, 55092, 19453, 810, 8305, 837, 2998, 20, 21 ], "markers": [ 38, 45, 52, 60 ], "logits_raw": [ -0.564216136932373, -0.7476570010185242, -0.27579110860824585, -0.22039031982421875 ], "probs_raw": [ 0.21848095953464508, 0.18186379969120026, 0.2915244400501251, 0.3081308901309967 ], "probs": [ 0.22808265686035156, 0.20007041096687317, 0.28026649355888367, 0.2915804088115692 ], "argmax_index": 3, "argmax": "d", "top2_margin": 0.011313915252685547, "near_tie": true, "gold": "d", "correct": true, "host_softmax_max_abs_dp": 2.9802322387695312e-08, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "answer": { "type": "choice", "choice": "d", "confidence": 0.2915804088115692, "probabilities": { "a": 0.22808265686035156, "b": 0.20007041096687317, "c": 0.28026649355888367, "d": 0.2915804088115692 } } }, "usage": { "input_tokens": 76, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "tv4_038", "mode": "text", "source": "transfer_v4_dev_head", "native": true, "prefix": 0, "usage_input_tokens": 143, "questions": [ { "qid": "answer", "type": "choice", "K": 4, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 143, "positions": 143, "bucket": 256, "ids": [ 1, 17, 38685, 42505, 6881, 997, 578, 2499, 813, 8731, 1878, 997, 39420, 6881, 997, 3493, 856, 779, 2120, 5521, 1685, 779, 16865, 2718, 59024, 843, 7094, 518, 810, 34258, 3995, 770, 1711, 843, 15612, 518, 10524, 811, 6232, 7152, 6976, 8523, 602, 18, 28828, 7832, 16066, 13041, 779, 3952, 540, 19, 16, 768, 535, 835, 7094, 856, 768, 3437, 10805, 4355, 811, 6232, 7152, 6976, 2158, 27967, 856, 768, 3769, 4355, 20, 19, 16, 794, 535, 27967, 6630, 15637, 779, 40129, 3996, 810, 8530, 43237, 1101, 896, 1413, 856, 6195, 25923, 1490, 779, 3996, 16454, 803, 779, 27967, 1927, 20, 19, 16, 781, 535, 941, 27967, 1927, 42773, 902, 22156, 23819, 875, 62228, 777, 6566, 11201, 13294, 797, 779, 1927, 20436, 20, 19, 16, 778, 535, 27967, 6630, 1052, 2184, 2410, 797, 2211, 7515, 11441, 40129, 6630, 1112, 1014, 523, 20, 21 ], "markers": [ 52, 74, 102, 124 ], "logits_raw": [ -0.7698924541473389, 0.4720029830932617, -0.3455861806869507, -0.4429304003715515 ], "probs_raw": [ 0.13554823398590088, 0.46929097175598145, 0.20718948543071747, 0.1879713535308838 ], "probs": [ 0.16541790962219238, 0.401660680770874, 0.22398363053798676, 0.20893773436546326 ], "argmax_index": 1, "argmax": "b", "top2_margin": 0.17767705023288727, "near_tie": false, "gold": "c", "correct": false, "host_softmax_max_abs_dp": 1.4901161193847656e-08, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "answer": { "type": "choice", "choice": "b", "confidence": 0.401660680770874, "probabilities": { "a": 0.16541790962219238, "b": 0.401660680770874, "c": 0.22398363053798676, "d": 0.20893773436546326 } } }, "usage": { "input_tokens": 143, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "tv4_039", "mode": "text", "source": "transfer_v4_dev_head", "native": true, "prefix": 0, "usage_input_tokens": 81, "questions": [ { "qid": "answer", "type": "choice", "K": 4, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 81, "positions": 81, "bucket": 256, "ids": [ 1, 17, 38685, 42505, 6881, 997, 17756, 2254, 4167, 578, 2499, 1504, 1878, 997, 39420, 6881, 997, 1098, 3426, 18768, 988, 775, 1008, 1666, 856, 15372, 1884, 1403, 1186, 938, 15399, 3032, 511, 602, 18, 28828, 7832, 16066, 13041, 779, 3952, 540, 19, 16, 768, 535, 22482, 18768, 856, 9617, 20, 19, 16, 794, 535, 22482, 7691, 30322, 22482, 2780, 20, 19, 16, 781, 535, 14060, 41938, 856, 9617, 20, 19, 16, 778, 535, 16791, 975, 9806, 856, 10836, 20, 21 ], "markers": [ 43, 52, 62, 71 ], "logits_raw": [ 1.7035473585128784, -1.886009931564331, -2.6342694759368896, -5.541640281677246 ], "probs_raw": [ 0.9602558612823486, 0.026513198390603065, 0.01254576351493597, 0.0006852386868558824 ], "probs": [ 0.8867239356040955, 0.06826366484165192, 0.039999522268772125, 0.005012943875044584 ], "argmax_index": 0, "argmax": "a", "top2_margin": 0.8184602707624435, "near_tie": false, "gold": "a", "correct": true, "host_softmax_max_abs_dp": 5.960464477539063e-08, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "answer": { "type": "choice", "choice": "a", "confidence": 0.8867239356040955, "probabilities": { "a": 0.8867239356040955, "b": 0.06826366484165192, "c": 0.039999522268772125, "d": 0.005012943875044584 } } }, "usage": { "input_tokens": 81, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "tv4_040", "mode": "text", "source": "transfer_v4_dev_head", "native": true, "prefix": 0, "usage_input_tokens": 219, "questions": [ { "qid": "answer", "type": "choice", "K": 4, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 219, "positions": 219, "bucket": 256, "ids": [ 1, 17, 38685, 42505, 6881, 997, 30327, 16964, 18849, 1878, 997, 39420, 6881, 997, 4083, 6984, 768, 1868, 988, 15202, 523, 13202, 1316, 856, 1943, 30133, 1929, 1450, 730, 1536, 6194, 10835, 9545, 31520, 779, 20481, 803, 1450, 2238, 6709, 2204, 3378, 6027, 810, 856, 1943, 1709, 853, 57994, 523, 941, 10761, 27973, 2094, 51707, 521, 4746, 2094, 6791, 521, 810, 1011, 2094, 5910, 40698, 523, 2297, 1052, 6235, 3758, 3822, 916, 21570, 810, 797, 7376, 10373, 11125, 4135, 523, 2876, 521, 1413, 856, 1295, 1235, 3581, 797, 779, 3681, 1316, 856, 1209, 9468, 1399, 1010, 938, 523, 1783, 902, 19316, 42724, 521, 1010, 1188, 17912, 602, 18, 28828, 7832, 16066, 13041, 779, 3952, 540, 19, 16, 768, 535, 11667, 811, 2130, 15202, 523, 13202, 1172, 9545, 797, 6738, 2908, 1010, 6695, 1320, 1429, 3758, 7376, 10373, 11125, 8938, 523, 20, 19, 16, 794, 535, 1836, 15202, 523, 13202, 1337, 1180, 5388, 803, 3758, 810, 2944, 1450, 11245, 1288, 1689, 11467, 1010, 811, 3176, 6738, 811, 1450, 9545, 523, 20, 19, 16, 781, 535, 1836, 15202, 523, 13202, 896, 1010, 4584, 2130, 1450, 9545, 1929, 803, 1180, 5388, 803, 3758, 523, 20, 19, 16, 778, 535, 2130, 15202, 523, 13202, 1172, 9545, 797, 6738, 1550, 3032, 779, 10373, 1178, 11446, 933, 3032, 1010, 28062, 8530, 4135, 523, 20, 21 ], "markers": [ 116, 142, 171, 193 ], "logits_raw": [ -2.3792548179626465, 0.8844364881515503, -0.7305554747581482, -2.2510766983032227 ], "probs_raw": [ 0.029866043478250504, 0.7808738350868225, 0.1553097814321518, 0.03395038843154907 ], "probs": [ 0.06395912170410156, 0.6582744121551514, 0.2076745629310608, 0.0700918585062027 ], "argmax_index": 1, "argmax": "b", "top2_margin": 0.4505998492240906, "near_tie": false, "gold": "d", "correct": false, "host_softmax_max_abs_dp": 1.4901161193847656e-08, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "answer": { "type": "choice", "choice": "b", "confidence": 0.6582744121551514, "probabilities": { "a": 0.06395912170410156, "b": 0.6582744121551514, "c": 0.2076745629310608, "d": 0.0700918585062027 } } }, "usage": { "input_tokens": 219, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "tv4_041", "mode": "text", "source": "transfer_v4_dev_head", "native": true, "prefix": 0, "usage_input_tokens": 226, "questions": [ { "qid": "answer", "type": "choice", "K": 4, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 226, "positions": 226, "bucket": 256, "ids": [ 1, 17, 38685, 42505, 6881, 997, 30327, 16964, 3036, 1878, 997, 39420, 6881, 997, 542, 16779, 953, 11915, 884, 768, 11513, 12991, 13668, 811, 1825, 768, 4153, 7909, 523, 941, 4153, 55154, 953, 3469, 5019, 521, 1101, 779, 16779, 6806, 811, 7498, 5329, 797, 779, 34517, 6707, 7921, 1929, 950, 4602, 768, 4605, 9837, 523, 10238, 53927, 768, 1681, 33138, 779, 11513, 12991, 810, 768, 10231, 7989, 1230, 523, 941, 16779, 1928, 3700, 791, 1391, 810, 10357, 811, 779, 54322, 523, 2904, 50322, 1720, 1994, 10231, 810, 2076, 950, 5942, 521, 779, 16779, 940, 21494, 521, 43175, 1098, 47715, 2172, 936, 510, 18588, 941, 47715, 953, 1459, 13735, 810, 856, 2166, 884, 8727, 875, 779, 16779, 1090, 12494, 523, 18276, 884, 1518, 11445, 6955, 521, 779, 6416, 1271, 3912, 803, 779, 16779, 1090, 8294, 1753, 874, 7877, 968, 511, 602, 18, 28828, 7832, 16066, 13041, 779, 3952, 540, 19, 16, 768, 535, 779, 28292, 521, 1720, 768, 27519, 22430, 968, 779, 16277, 523, 20, 19, 16, 794, 535, 779, 28292, 521, 2630, 768, 27519, 22430, 968, 779, 16277, 523, 20, 19, 16, 781, 535, 779, 16277, 521, 916, 11497, 811, 779, 28292, 896, 936, 1400, 2260, 779, 21712, 31137, 1288, 936, 25195, 896, 779, 16779, 8291, 950, 953, 21712, 523, 20, 19, 16, 778, 535, 779, 16277, 521, 2630, 12224, 988, 779, 28292, 523, 20, 21 ], "markers": [ 149, 165, 181, 212 ], "logits_raw": [ -0.2598228454589844, -0.670441746711731, 0.7866662740707397, -2.2362213134765625 ], "probs_raw": [ 0.21507954597473145, 0.14264927804470062, 0.6124681830406189, 0.02980300597846508 ], "probs": [ 0.24382537603378296, 0.18184082210063934, 0.5149134993553162, 0.059420328587293625 ], "argmax_index": 2, "argmax": "c", "top2_margin": 0.2710881233215332, "near_tie": false, "gold": "d", "correct": false, "host_softmax_max_abs_dp": 1.4901161193847656e-08, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "answer": { "type": "choice", "choice": "c", "confidence": 0.5149134993553162, "probabilities": { "a": 0.24382537603378296, "b": 0.18184082210063934, "c": 0.5149134993553162, "d": 0.059420328587293625 } } }, "usage": { "input_tokens": 226, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "tv4_042", "mode": "text", "source": "transfer_v4_dev_head", "native": true, "prefix": 0, "usage_input_tokens": 73, "questions": [ { "qid": "answer", "type": "choice", "K": 4, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 73, "positions": 73, "bucket": 256, "ids": [ 1, 17, 38685, 42505, 6881, 997, 3600, 1021, 3981, 36158, 1878, 997, 39420, 6881, 997, 1098, 3981, 3384, 803, 36025, 968, 10039, 856, 768, 3680, 2432, 803, 511, 602, 18, 28828, 7832, 16066, 13041, 779, 3952, 540, 19, 16, 768, 535, 12499, 3384, 803, 10807, 20, 19, 16, 794, 535, 2097, 949, 3294, 16798, 20, 19, 16, 781, 535, 10439, 5222, 779, 3952, 20, 19, 16, 778, 535, 13104, 1437, 23223, 20, 21 ], "markers": [ 38, 47, 56, 65 ], "logits_raw": [ 1.2054716348648071, -2.7291550636291504, -2.9371767044067383, -3.3536458015441895 ], "probs_raw": [ 0.9561097621917725, 0.018694812431931496, 0.01518373005092144, 0.010011709295213223 ], "probs": [ 0.8691593408584595, 0.052293624728918076, 0.04507270082831383, 0.03347426652908325 ], "argmax_index": 0, "argmax": "a", "top2_margin": 0.8168657161295414, "near_tie": false, "gold": "a", "correct": true, "host_softmax_max_abs_dp": 3.725290298461914e-09, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "answer": { "type": "choice", "choice": "a", "confidence": 0.8691593408584595, "probabilities": { "a": 0.8691593408584595, "b": 0.052293624728918076, "c": 0.04507270082831383, "d": 0.03347426652908325 } } }, "usage": { "input_tokens": 73, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "tv4_043", "mode": "text", "source": "transfer_v4_dev_head", "native": true, "prefix": 0, "usage_input_tokens": 72, "questions": [ { "qid": "answer", "type": "choice", "K": 4, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 72, "positions": 72, "bucket": 256, "ids": [ 1, 17, 38685, 42505, 6881, 997, 30735, 14977, 10028, 1878, 997, 39420, 6881, 997, 31081, 15909, 797, 784, 14681, 2841, 803, 768, 13989, 10321, 31067, 938, 2588, 906, 511, 602, 18, 28828, 7832, 16066, 13041, 779, 3952, 540, 19, 16, 768, 535, 4021, 20744, 32637, 933, 7604, 32637, 20, 19, 16, 794, 535, 20744, 32637, 20, 19, 16, 781, 535, 7604, 32637, 20, 19, 16, 778, 535, 1623, 1143, 32637, 20, 21 ], "markers": [ 39, 50, 57, 64 ], "logits_raw": [ -1.8836767673492432, -2.334036350250244, -3.524329423904419, -3.608792781829834 ], "probs_raw": [ 0.497659832239151, 0.31720784306526184, 0.0964730903506279, 0.08865928649902344 ], "probs": [ 0.4298710525035858, 0.31161749362945557, 0.13315387070178986, 0.12535753846168518 ], "argmax_index": 0, "argmax": "a", "top2_margin": 0.11825355887413025, "near_tie": false, "gold": "b", "correct": false, "host_softmax_max_abs_dp": 2.9802322387695312e-08, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "answer": { "type": "choice", "choice": "a", "confidence": 0.4298710525035858, "probabilities": { "a": 0.4298710525035858, "b": 0.31161749362945557, "c": 0.13315387070178986, "d": 0.12535753846168518 } } }, "usage": { "input_tokens": 72, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "tv4_044", "mode": "text", "source": "transfer_v4_dev_head", "native": true, "prefix": 0, "usage_input_tokens": 89, "questions": [ { "qid": "answer", "type": "choice", "K": 4, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 89, "positions": 89, "bucket": 256, "ids": [ 1, 17, 38685, 42505, 6881, 997, 17756, 2254, 3134, 810, 13032, 1878, 997, 39420, 6881, 997, 1098, 20069, 3345, 1678, 803, 9170, 800, 5614, 3714, 10380, 953, 10328, 968, 779, 39629, 1678, 906, 768, 1725, 803, 779, 511, 602, 18, 28828, 7832, 16066, 13041, 779, 3952, 540, 19, 16, 768, 535, 53311, 59812, 20, 19, 16, 794, 535, 14366, 10335, 23335, 4120, 803, 730, 1848, 526, 20, 19, 16, 781, 535, 2947, 21889, 40468, 4120, 20, 19, 16, 778, 535, 54130, 13501, 4120, 803, 730, 3699, 528, 20, 21 ], "markers": [ 48, 55, 68, 77 ], "logits_raw": [ -2.3090124130249023, -2.2886719703674316, -0.1633472740650177, 1.2341519594192505 ], "probs_raw": [ 0.022151127457618713, 0.02260630577802658, 0.1893417239189148, 0.7659007906913757 ], "probs": [ 0.052051614969968796, 0.05281345546245575, 0.24104060232639313, 0.654094398021698 ], "argmax_index": 3, "argmax": "d", "top2_margin": 0.41305379569530487, "near_tie": false, "gold": "d", "correct": true, "host_softmax_max_abs_dp": 5.960464477539063e-08, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "answer": { "type": "choice", "choice": "d", "confidence": 0.654094398021698, "probabilities": { "a": 0.052051614969968796, "b": 0.05281345546245575, "c": 0.24104060232639313, "d": 0.654094398021698 } } }, "usage": { "input_tokens": 89, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "tv4_045", "mode": "text", "source": "transfer_v4_dev_head", "native": true, "prefix": 0, "usage_input_tokens": 106, "questions": [ { "qid": "answer", "type": "choice", "K": 4, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 106, "positions": 106, "bucket": 256, "ids": [ 1, 17, 38685, 42505, 6881, 997, 578, 2499, 813, 8731, 1878, 997, 39420, 6881, 997, 25551, 779, 2773, 2801, 875, 1058, 598, 2884, 513, 23606, 587, 13699, 513, 598, 2884, 1051, 13264, 4634, 2011, 13264, 5762, 766, 2011, 832, 2884, 4927, 587, 13699, 28828, 1235, 803, 779, 2773, 1529, 20298, 11854, 779, 1996, 875, 1058, 598, 2884, 513, 8523, 602, 18, 28828, 7832, 16066, 13041, 779, 3952, 540, 19, 16, 768, 535, 835, 6232, 7152, 1996, 20, 19, 16, 794, 535, 835, 37970, 1996, 20, 19, 16, 781, 535, 835, 15249, 3547, 13974, 1996, 20, 19, 16, 778, 535, 835, 7226, 5043, 916, 26436, 20, 21 ], "markers": [ 69, 78, 86, 96 ], "logits_raw": [ 1.1159785985946655, -0.7625792026519775, -0.41189315915107727, -1.6540266275405884 ], "probs_raw": [ 0.6980953812599182, 0.10667619109153748, 0.15148460865020752, 0.04374382644891739 ], "probs": [ 0.5762596130371094, 0.15060095489025116, 0.19347375631332397, 0.07966570556163788 ], "argmax_index": 0, "argmax": "a", "top2_margin": 0.3827858567237854, "near_tie": false, "gold": "c", "correct": false, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "answer": { "type": "choice", "choice": "a", "confidence": 0.5762596130371094, "probabilities": { "a": 0.5762596130371094, "b": 0.15060095489025116, "c": 0.19347375631332397, "d": 0.07966570556163788 } } }, "usage": { "input_tokens": 106, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "tv4_046", "mode": "text", "source": "transfer_v4_dev_head", "native": true, "prefix": 0, "usage_input_tokens": 78, "questions": [ { "qid": "answer", "type": "choice", "K": 4, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 78, "positions": 78, "bucket": 256, "ids": [ 1, 17, 38685, 42505, 6881, 997, 17756, 2254, 16290, 1878, 997, 39420, 6881, 997, 2443, 1996, 8806, 811, 779, 9257, 3893, 803, 1547, 14471, 1005, 595, 931, 521, 1547, 26171, 521, 810, 1547, 55984, 549, 2927, 602, 18, 28828, 7832, 16066, 13041, 779, 3952, 540, 19, 16, 768, 535, 47256, 2293, 2916, 20, 19, 16, 794, 535, 8670, 2574, 586, 3838, 20, 19, 16, 781, 535, 12656, 957, 852, 20, 19, 16, 778, 535, 41778, 8192, 20, 21 ], "markers": [ 46, 54, 63, 71 ], "logits_raw": [ -0.21103397011756897, -4.946765422821045, -1.4108130931854248, -0.5645290017127991 ], "probs_raw": [ 0.4969521462917328, 0.004361264873296022, 0.14971217513084412, 0.34897446632385254 ], "probs": [ 0.4473859369754791, 0.015188070014119148, 0.18987658619880676, 0.3475493788719177 ], "argmax_index": 0, "argmax": "a", "top2_margin": 0.0998365581035614, "near_tie": false, "gold": "a", "correct": true, "host_softmax_max_abs_dp": 5.960464477539063e-08, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "answer": { "type": "choice", "choice": "a", "confidence": 0.4473859369754791, "probabilities": { "a": 0.4473859369754791, "b": 0.015188070014119148, "c": 0.18987658619880676, "d": 0.3475493788719177 } } }, "usage": { "input_tokens": 78, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "tv4_047", "mode": "text", "source": "transfer_v4_dev_head", "native": true, "prefix": 0, "usage_input_tokens": 79, "questions": [ { "qid": "answer", "type": "choice", "K": 4, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 79, "positions": 79, "bucket": 256, "ids": [ 1, 17, 38685, 42505, 6881, 997, 30327, 16964, 18849, 1878, 997, 39420, 6881, 997, 542, 6394, 5867, 1512, 4100, 44888, 811, 1450, 2153, 5822, 1689, 13564, 20972, 4381, 4100, 8255, 523, 941, 13789, 20314, 968, 1033, 6394, 9784, 768, 15488, 803, 1450, 511, 602, 18, 28828, 7832, 16066, 13041, 779, 3952, 540, 19, 16, 768, 535, 1267, 29871, 20, 19, 16, 794, 535, 16234, 7695, 20, 19, 16, 781, 535, 49249, 20, 19, 16, 778, 535, 38939, 20, 21 ], "markers": [ 53, 60, 67, 73 ], "logits_raw": [ -3.3293027877807617, 2.4848766326904297, -5.358247756958008, 1.0101518630981445 ], "probs_raw": [ 0.0024224012158811092, 0.8115440607070923, 0.000318483158480376, 0.18571512401103973 ], "probs": [ 0.011484759859740734, 0.7309231758117676, 0.002695730421692133, 0.25489628314971924 ], "argmax_index": 1, "argmax": "b", "top2_margin": 0.47602689266204834, "near_tie": false, "gold": "c", "correct": false, "host_softmax_max_abs_dp": 5.960464477539063e-08, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "answer": { "type": "choice", "choice": "b", "confidence": 0.7309231758117676, "probabilities": { "a": 0.011484759859740734, "b": 0.7309231758117676, "c": 0.002695730421692133, "d": 0.25489628314971924 } } }, "usage": { "input_tokens": 79, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "tv4_048", "mode": "text", "source": "transfer_v4_dev_head", "native": true, "prefix": 0, "usage_input_tokens": 136, "questions": [ { "qid": "answer", "type": "choice", "K": 4, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 136, "positions": 136, "bucket": 256, "ids": [ 1, 17, 38685, 42505, 6881, 997, 586, 8988, 21469, 1878, 997, 39420, 6881, 997, 3111, 1144, 803, 1518, 1547, 21469, 2177, 779, 2120, 3338, 843, 18275, 5582, 859, 524, 1477, 524, 3462, 518, 1112, 3503, 8870, 61611, 7871, 521, 4262, 811, 17027, 11241, 8079, 797, 779, 2641, 906, 803, 730, 1718, 525, 540, 3108, 31933, 730, 526, 2171, 859, 3595, 5665, 1419, 768, 9953, 1101, 859, 1738, 1927, 936, 875, 6471, 523, 3108, 31933, 730, 527, 2171, 859, 17453, 779, 30704, 1680, 1429, 6248, 29293, 2927, 602, 18, 28828, 7832, 16066, 13041, 779, 3952, 540, 19, 16, 768, 535, 923, 35736, 521, 923, 35736, 20, 19, 16, 794, 535, 923, 35736, 521, 4524, 7871, 20, 19, 16, 781, 535, 4524, 7871, 521, 923, 35736, 20, 19, 16, 778, 535, 4524, 7871, 521, 4524, 7871, 20, 21 ], "markers": [ 96, 106, 116, 126 ], "logits_raw": [ 1.6919503211975098, -0.03276406228542328, -0.9112541675567627, -1.3428339958190918 ], "probs_raw": [ 0.7690266966819763, 0.13705897331237793, 0.0569356232881546, 0.036978673189878464 ], "probs": [ 0.6402612328529358, 0.18676437437534332, 0.09971431642770767, 0.07326005399227142 ], "argmax_index": 0, "argmax": "a", "top2_margin": 0.45349685847759247, "near_tie": false, "gold": "d", "correct": false, "host_softmax_max_abs_dp": 5.960464477539063e-08, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "answer": { "type": "choice", "choice": "a", "confidence": 0.6402612328529358, "probabilities": { "a": 0.6402612328529358, "b": 0.18676437437534332, "c": 0.09971431642770767, "d": 0.07326005399227142 } } }, "usage": { "input_tokens": 136, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "tv4_049", "mode": "text", "source": "transfer_v4_dev_head", "native": true, "prefix": 0, "usage_input_tokens": 96, "questions": [ { "qid": "answer", "type": "choice", "K": 4, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 96, "positions": 96, "bucket": 256, "ids": [ 1, 17, 38685, 42505, 6881, 997, 5146, 3717, 1878, 997, 39420, 6881, 997, 28828, 8294, 856, 3347, 1337, 3560, 2514, 2098, 8523, 602, 18, 28828, 7832, 16066, 13041, 779, 3952, 540, 19, 16, 768, 535, 941, 7524, 1753, 874, 7795, 963, 7226, 523, 20, 19, 16, 794, 535, 1311, 856, 902, 2623, 803, 768, 9885, 2098, 6690, 875, 27877, 523, 20, 19, 16, 781, 535, 941, 2203, 1011, 874, 33518, 811, 8487, 768, 5814, 3443, 523, 20, 19, 16, 778, 535, 941, 7747, 10361, 29581, 1818, 896, 1011, 874, 1593, 811, 1705, 21622, 523, 20, 21 ], "markers": [ 32, 45, 62, 78 ], "logits_raw": [ -6.835418224334717, -0.9563892483711243, -3.4500579833984375, -0.31408247351646423 ], "probs_raw": [ 0.0009367892635054886, 0.3348664939403534, 0.027662096545100212, 0.636534571647644 ], "probs": [ 0.005424226634204388, 0.3615814447402954, 0.0608958899974823, 0.5720984935760498 ], "argmax_index": 3, "argmax": "d", "top2_margin": 0.2105170488357544, "near_tie": false, "gold": "d", "correct": true, "host_softmax_max_abs_dp": 5.960464477539063e-08, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "answer": { "type": "choice", "choice": "d", "confidence": 0.5720984935760498, "probabilities": { "a": 0.005424226634204388, "b": 0.3615814447402954, "c": 0.0608958899974823, "d": 0.5720984935760498 } } }, "usage": { "input_tokens": 96, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "tv4_050", "mode": "text", "source": "transfer_v4_dev_head", "native": true, "prefix": 0, "usage_input_tokens": 78, "questions": [ { "qid": "answer", "type": "choice", "K": 4, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 78, "positions": 78, "bucket": 256, "ids": [ 1, 17, 38685, 42505, 6881, 997, 578, 12453, 17263, 1878, 997, 39420, 6881, 997, 64005, 779, 10805, 15702, 1417, 523, 5843, 1051, 15702, 527, 523, 528, 598, 523, 11205, 1180, 5642, 2927, 602, 18, 28828, 7832, 16066, 13041, 779, 3952, 540, 19, 16, 768, 535, 730, 532, 523, 529, 20, 19, 16, 794, 535, 15702, 532, 523, 529, 20, 19, 16, 781, 535, 15702, 913, 523, 2925, 20, 19, 16, 778, 535, 730, 1536, 523, 8333, 20, 21 ], "markers": [ 42, 51, 60, 69 ], "logits_raw": [ -0.026801295578479767, -0.5347694158554077, -2.2902116775512695, -0.3073588013648987 ], "probs_raw": [ 0.4063265323638916, 0.24449355900287628, 0.04225601628422737, 0.30692383646965027 ], "probs": [ 0.36864981055259705, 0.25646352767944336, 0.07318627089262009, 0.3017003834247589 ], "argmax_index": 0, "argmax": "a", "top2_margin": 0.06694942712783813, "near_tie": false, "gold": "a", "correct": true, "host_softmax_max_abs_dp": 2.9802322387695312e-08, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "answer": { "type": "choice", "choice": "a", "confidence": 0.36864981055259705, "probabilities": { "a": 0.36864981055259705, "b": 0.25646352767944336, "c": 0.07318627089262009, "d": 0.3017003834247589 } } }, "usage": { "input_tokens": 78, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "tv4_051", "mode": "text", "source": "transfer_v4_dev_head", "native": true, "prefix": 0, "usage_input_tokens": 145, "questions": [ { "qid": "answer", "type": "choice", "K": 4, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 145, "positions": 145, "bucket": 256, "ids": [ 1, 17, 38685, 42505, 6881, 997, 3034, 41336, 1878, 997, 39420, 6881, 997, 1098, 8757, 797, 6131, 803, 3528, 769, 2796, 3525, 810, 1993, 9903, 6486, 968, 963, 3653, 730, 528, 523, 528, 3847, 1779, 6027, 856, 7951, 968, 17912, 602, 18, 28828, 7832, 16066, 13041, 779, 3952, 540, 19, 16, 768, 535, 4169, 16120, 884, 6134, 13665, 896, 1166, 2047, 968, 1203, 824, 2644, 916, 9556, 803, 9903, 523, 20, 19, 16, 794, 535, 779, 36448, 803, 4163, 920, 42437, 11490, 3545, 33450, 868, 6108, 810, 18583, 969, 916, 3609, 803, 4624, 2222, 54127, 57203, 896, 13473, 6690, 1517, 523, 20, 19, 16, 781, 535, 4048, 9903, 39052, 521, 902, 20247, 18438, 521, 810, 1434, 2385, 1833, 963, 779, 903, 1071, 584, 14415, 730, 528, 3475, 797, 3227, 16485, 2468, 521, 24973, 20, 19, 16, 778, 535, 1113, 803, 779, 3537, 523, 20, 21 ], "markers": [ 50, 72, 103, 135 ], "logits_raw": [ -2.4747769832611084, 1.4043216705322266, -0.4620096981525421, 2.141174077987671 ], "probs_raw": [ 0.006331176962703466, 0.30630606412887573, 0.047382522374391556, 0.6399801969528198 ], "probs": [ 0.020735442638397217, 0.3312358558177948, 0.08732527494430542, 0.560703456401825 ], "argmax_index": 3, "argmax": "d", "top2_margin": 0.22946760058403015, "near_tie": false, "gold": "d", "correct": true, "host_softmax_max_abs_dp": 7.450580596923828e-09, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "answer": { "type": "choice", "choice": "d", "confidence": 0.560703456401825, "probabilities": { "a": 0.020735442638397217, "b": 0.3312358558177948, "c": 0.08732527494430542, "d": 0.560703456401825 } } }, "usage": { "input_tokens": 145, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "tv4_052", "mode": "text", "source": "transfer_v4_dev_head", "native": true, "prefix": 0, "usage_input_tokens": 143, "questions": [ { "qid": "answer", "type": "choice", "K": 4, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 143, "positions": 143, "bucket": 256, "ids": [ 1, 17, 38685, 42505, 6881, 997, 17756, 2254, 18166, 1878, 997, 39420, 6881, 997, 1098, 9257, 14951, 793, 10805, 5559, 1403, 7864, 803, 33128, 529, 21369, 810, 4059, 20099, 528, 938, 10372, 3658, 856, 511, 602, 18, 28828, 7832, 16066, 13041, 779, 3952, 540, 19, 16, 768, 535, 4059, 46777, 63868, 518, 2011, 2642, 25960, 63868, 518, 17493, 4059, 21369, 5381, 518, 20, 19, 16, 794, 535, 33128, 529, 46777, 63868, 518, 2011, 4059, 46777, 63868, 518, 17493, 4059, 22386, 549, 529, 518, 528, 520, 527, 13792, 590, 518, 20, 19, 16, 781, 535, 2642, 25960, 63868, 518, 2011, 15081, 528, 25960, 63868, 518, 17493, 15081, 528, 21369, 13792, 590, 518, 20, 19, 16, 778, 535, 33128, 529, 21369, 13792, 590, 518, 2011, 15081, 528, 25960, 63868, 518, 17493, 33128, 529, 20099, 528, 13792, 590, 518, 2011, 2642, 25960, 63868, 518, 20, 21 ], "markers": [ 45, 64, 91, 113 ], "logits_raw": [ 0.5273202657699585, -0.3178127408027649, -2.177464246749878, 0.6166142225265503 ], "probs_raw": [ 0.3861321210861206, 0.16584382951259613, 0.02582634799182415, 0.4221976399421692 ], "probs": [ 0.3626495897769928, 0.19828923046588898, 0.05252596735954285, 0.38653525710105896 ], "argmax_index": 3, "argmax": "d", "top2_margin": 0.023885667324066162, "near_tie": false, "gold": "a", "correct": false, "host_softmax_max_abs_dp": 1.4901161193847656e-08, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "answer": { "type": "choice", "choice": "d", "confidence": 0.38653525710105896, "probabilities": { "a": 0.3626495897769928, "b": 0.19828923046588898, "c": 0.05252596735954285, "d": 0.38653525710105896 } } }, "usage": { "input_tokens": 143, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "tv4_053", "mode": "text", "source": "transfer_v4_dev_head", "native": true, "prefix": 0, "usage_input_tokens": 343, "questions": [ { "qid": "answer", "type": "choice", "K": 4, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 343, "positions": 343, "bucket": 512, "ids": [ 1, 17, 38685, 42505, 6881, 997, 30327, 16964, 3036, 1878, 997, 39420, 6881, 997, 2797, 12153, 16217, 768, 7021, 18479, 3436, 523, 1405, 8091, 950, 7050, 811, 42766, 810, 9670, 1210, 3436, 875, 12614, 523, 835, 48463, 5944, 779, 21008, 810, 15596, 779, 18479, 3436, 523, 941, 48463, 5689, 779, 12153, 896, 1689, 953, 1943, 9368, 797, 18649, 779, 3436, 521, 1203, 896, 1689, 4602, 811, 1673, 768, 16763, 988, 779, 7493, 811, 17830, 779, 11931, 523, 941, 9888, 43360, 768, 6827, 521, 10423, 968, 2059, 9888, 521, 1144, 9519, 896, 779, 48463, 11860, 811, 11931, 779, 18479, 3436, 988, 779, 12153, 875, 1058, 7896, 521, 1873, 523, 941, 4985, 10449, 11874, 768, 19505, 41622, 779, 12614, 1499, 1014, 2337, 1862, 43175, 885, 2620, 779, 48463, 856, 7021, 797, 1450, 23956, 811, 4387, 768, 16763, 988, 768, 7493, 811, 17830, 779, 11931, 803, 779, 3436, 523, 43175, 50380, 521, 779, 48463, 2047, 1295, 4516, 811, 5308, 811, 4387, 768, 16763, 988, 768, 7493, 1101, 896, 779, 12614, 1738, 874, 2524, 1884, 523, 4162, 2993, 4493, 521, 779, 12153, 25752, 14503, 2723, 779, 48463, 875, 43510, 803, 6827, 523, 20463, 803, 779, 2773, 1188, 874, 779, 12153, 1090, 31712, 8757, 797, 2258, 803, 1210, 4272, 2723, 779, 48463, 875, 43510, 803, 6827, 8523, 602, 18, 28828, 7832, 16066, 13041, 779, 3952, 540, 19, 16, 768, 535, 941, 27471, 803, 768, 16763, 988, 768, 7493, 953, 1014, 2359, 791, 906, 768, 997, 25556, 511, 811, 779, 48463, 1090, 15774, 811, 7909, 779, 3436, 523, 20, 19, 16, 794, 535, 6463, 27471, 768, 16763, 988, 768, 7493, 953, 768, 4576, 811, 779, 48463, 1090, 15774, 811, 7909, 779, 3436, 521, 779, 4576, 1753, 874, 4016, 6086, 1929, 20, 19, 16, 781, 535, 6463, 27471, 768, 16763, 988, 768, 7493, 953, 768, 4576, 811, 779, 48463, 1090, 15774, 811, 7909, 779, 3436, 521, 936, 1753, 874, 54604, 3449, 988, 779, 20, 19, 16, 778, 535, 941, 48463, 2276, 4660, 902, 32056, 16114, 811, 1825, 768, 16336, 4516, 811, 4387, 768, 16763, 988, 768, 7493, 523, 20, 21 ], "markers": [ 222, 254, 286, 318 ], "logits_raw": [ 0.5962756872177124, -0.33312368392944336, 0.5012220740318298, -0.5879086852073669 ], "probs_raw": [ 0.38312551379203796, 0.15125441551208496, 0.3483853042125702, 0.11723475158214569 ], "probs": [ 0.34742024540901184, 0.17886479198932648, 0.3246133625507355, 0.14910155534744263 ], "argmax_index": 0, "argmax": "a", "top2_margin": 0.022806882858276367, "near_tie": false, "gold": "d", "correct": false, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "answer": { "type": "choice", "choice": "a", "confidence": 0.34742024540901184, "probabilities": { "a": 0.34742024540901184, "b": 0.17886479198932648, "c": 0.3246133625507355, "d": 0.14910155534744263 } } }, "usage": { "input_tokens": 343, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "tv4_054", "mode": "text", "source": "transfer_v4_dev_head", "native": true, "prefix": 0, "usage_input_tokens": 109, "questions": [ { "qid": "answer", "type": "choice", "K": 4, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 109, "positions": 109, "bucket": 256, "ids": [ 1, 17, 38685, 42505, 6881, 997, 2293, 3810, 18166, 1878, 997, 39420, 6881, 997, 59061, 931, 779, 12277, 26698, 21232, 803, 730, 4613, 35117, 523, 1607, 550, 1051, 730, 532, 41562, 736, 527, 570, 511, 602, 18, 28828, 7832, 16066, 13041, 779, 3952, 540, 19, 16, 768, 535, 730, 527, 523, 10782, 3053, 730, 1155, 34193, 2933, 962, 775, 20, 19, 16, 794, 535, 730, 528, 523, 25785, 3053, 730, 1155, 34193, 2933, 962, 775, 20, 19, 16, 781, 535, 730, 529, 523, 4420, 3053, 730, 1155, 34193, 2933, 962, 775, 20, 19, 16, 778, 535, 730, 530, 523, 33052, 3053, 730, 1155, 34193, 2933, 962, 775, 20, 21 ], "markers": [ 45, 61, 77, 93 ], "logits_raw": [ -3.5054848194122314, -2.425651788711548, -0.6409085988998413, -3.510509490966797 ], "probs_raw": [ 0.044482409954071045, 0.13096457719802856, 0.7802935242652893, 0.04425947368144989 ], "probs": [ 0.08404666930437088, 0.18176917731761932, 0.6504386067390442, 0.08374553918838501 ], "argmax_index": 2, "argmax": "c", "top2_margin": 0.46866942942142487, "near_tie": false, "gold": "c", "correct": true, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "answer": { "type": "choice", "choice": "c", "confidence": 0.6504386067390442, "probabilities": { "a": 0.08404666930437088, "b": 0.18176917731761932, "c": 0.6504386067390442, "d": 0.08374553918838501 } } }, "usage": { "input_tokens": 109, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "tv4_055", "mode": "text", "source": "transfer_v4_dev_head", "native": true, "prefix": 0, "usage_input_tokens": 169, "questions": [ { "qid": "answer", "type": "choice", "K": 4, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 169, "positions": 169, "bucket": 256, "ids": [ 1, 17, 38685, 42505, 6881, 997, 30327, 16964, 21167, 1878, 997, 39420, 6881, 997, 43211, 853, 3862, 523, 5582, 768, 26856, 30174, 1678, 523, 941, 2773, 938, 30174, 23301, 875, 779, 5805, 803, 5250, 535, 730, 526, 524, 526, 47044, 30174, 730, 1086, 1056, 8202, 963, 1058, 1532, 730, 526, 524, 898, 873, 43418, 730, 5091, 1056, 8202, 963, 1058, 1386, 730, 526, 524, 1968, 873, 43418, 730, 6019, 1056, 8202, 963, 1058, 1417, 730, 526, 524, 2999, 35791, 963, 1058, 898, 1094, 6232, 730, 4291, 1056, 8202, 47751, 5582, 779, 903, 8526, 556, 2511, 811, 6187, 779, 2536, 803, 1352, 30174, 523, 3747, 3685, 1753, 47751, 2312, 906, 2780, 803, 11391, 4997, 884, 1352, 8292, 8294, 875, 779, 5805, 803, 5250, 8523, 602, 18, 28828, 7832, 16066, 13041, 779, 3952, 540, 19, 16, 768, 535, 1058, 22557, 521, 1873, 20, 19, 16, 794, 535, 1058, 14671, 521, 1873, 20, 19, 16, 781, 535, 1058, 23975, 521, 1873, 20, 19, 16, 778, 535, 730, 2096, 1873, 525, 20, 21 ], "markers": [ 133, 142, 151, 160 ], "logits_raw": [ -3.6334569454193115, -0.35530930757522583, -3.424314260482788, -5.927770137786865 ], "probs_raw": [ 0.034649983048439026, 0.9191460013389587, 0.04271029308438301, 0.003493780270218849 ], "probs": [ 0.07840510457754135, 0.8153303861618042, 0.09103896468877792, 0.015225550159811974 ], "argmax_index": 1, "argmax": "b", "top2_margin": 0.7242914214730263, "near_tie": false, "gold": "c", "correct": false, "host_softmax_max_abs_dp": 0.0, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "answer": { "type": "choice", "choice": "b", "confidence": 0.8153303861618042, "probabilities": { "a": 0.07840510457754135, "b": 0.8153303861618042, "c": 0.09103896468877792, "d": 0.015225550159811974 } } }, "usage": { "input_tokens": 169, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "tv4_056", "mode": "text", "source": "transfer_v4_dev_head", "native": true, "prefix": 0, "usage_input_tokens": 81, "questions": [ { "qid": "answer", "type": "choice", "K": 4, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 81, "positions": 81, "bucket": 256, "ids": [ 1, 17, 38685, 42505, 6881, 997, 17756, 2254, 31116, 1878, 997, 39420, 6881, 997, 28828, 803, 779, 2773, 856, 14353, 768, 5992, 803, 902, 13793, 9147, 9985, 843, 10153, 567, 518, 8523, 602, 18, 28828, 7832, 16066, 13041, 779, 3952, 540, 19, 16, 768, 535, 14710, 3890, 35495, 884, 30118, 20, 19, 16, 794, 535, 27965, 5168, 28443, 15774, 11867, 20, 19, 16, 781, 535, 2315, 34987, 988, 30101, 20698, 20, 19, 16, 778, 535, 1087, 2930, 11244, 14272, 20, 21 ], "markers": [ 42, 52, 62, 72 ], "logits_raw": [ -7.277989864349365, -2.838202953338623, -4.520140171051025, -0.9320415258407593 ], "probs_raw": [ 0.0014887587167322636, 0.12618252635002136, 0.02347162738442421, 0.8488571047782898 ], "probs": [ 0.007995810359716415, 0.1906471699476242, 0.057337407022714615, 0.744019627571106 ], "argmax_index": 3, "argmax": "d", "top2_margin": 0.5533724576234818, "near_tie": false, "gold": "d", "correct": true, "host_softmax_max_abs_dp": 5.960464477539063e-08, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "answer": { "type": "choice", "choice": "d", "confidence": 0.744019627571106, "probabilities": { "a": 0.007995810359716415, "b": 0.1906471699476242, "c": 0.057337407022714615, "d": 0.744019627571106 } } }, "usage": { "input_tokens": 81, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "tv4_057", "mode": "text", "source": "transfer_v4_dev_head", "native": true, "prefix": 0, "usage_input_tokens": 73, "questions": [ { "qid": "answer", "type": "choice", "K": 4, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 73, "positions": 73, "bucket": 256, "ids": [ 1, 17, 38685, 42505, 6881, 997, 28009, 19179, 1878, 997, 39420, 6881, 997, 28828, 803, 779, 2773, 856, 14353, 1235, 803, 779, 1209, 9146, 5668, 4637, 968, 2025, 1316, 938, 19620, 8523, 602, 18, 28828, 7832, 16066, 13041, 779, 3952, 540, 19, 16, 768, 535, 42154, 2644, 20, 19, 16, 794, 535, 26596, 20, 19, 16, 781, 535, 5148, 54599, 779, 2681, 20, 19, 16, 778, 535, 5754, 4289, 979, 2362, 20, 21 ], "markers": [ 42, 49, 55, 64 ], "logits_raw": [ -8.583589553833008, -5.527565956115723, -6.5222697257995605, -5.621074676513672 ], "probs_raw": [ 0.020224157720804214, 0.4296201169490814, 0.15888769924640656, 0.39126795530319214 ], "probs": [ 0.044379282742738724, 0.39378514885902405, 0.19349488615989685, 0.36834073066711426 ], "argmax_index": 1, "argmax": "b", "top2_margin": 0.02544441819190979, "near_tie": false, "gold": "c", "correct": false, "host_softmax_max_abs_dp": 2.9802322387695312e-08, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "answer": { "type": "choice", "choice": "b", "confidence": 0.39378514885902405, "probabilities": { "a": 0.044379282742738724, "b": 0.39378514885902405, "c": 0.19349488615989685, "d": 0.36834073066711426 } } }, "usage": { "input_tokens": 73, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "tv4_058", "mode": "text", "source": "transfer_v4_dev_head", "native": true, "prefix": 0, "usage_input_tokens": 147, "questions": [ { "qid": "answer", "type": "choice", "K": 4, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 147, "positions": 147, "bucket": 256, "ids": [ 1, 17, 38685, 42505, 6881, 997, 586, 8988, 21469, 1878, 997, 39420, 6881, 997, 3111, 1144, 803, 1518, 1547, 21469, 2177, 779, 2120, 3338, 843, 18275, 5582, 859, 524, 1477, 524, 3462, 518, 1112, 3503, 8870, 61611, 7871, 521, 4262, 811, 17027, 11241, 8079, 797, 779, 2641, 906, 803, 730, 1718, 525, 540, 3108, 31933, 730, 526, 2171, 859, 8940, 1727, 6709, 1673, 1230, 803, 779, 3622, 1720, 950, 5807, 1210, 58110, 523, 3108, 31933, 730, 527, 2171, 859, 4925, 1886, 45201, 2601, 11468, 797, 779, 3681, 12420, 521, 2158, 7250, 953, 963, 779, 12420, 3662, 2927, 602, 18, 28828, 7832, 16066, 13041, 779, 3952, 540, 19, 16, 768, 535, 923, 35736, 521, 923, 35736, 20, 19, 16, 794, 535, 923, 35736, 521, 4524, 7871, 20, 19, 16, 781, 535, 4524, 7871, 521, 923, 35736, 20, 19, 16, 778, 535, 4524, 7871, 521, 4524, 7871, 20, 21 ], "markers": [ 107, 117, 127, 137 ], "logits_raw": [ 0.6942271590232849, -2.231766939163208, -0.7739002704620361, -1.7253468036651611 ], "probs_raw": [ 0.7283706665039062, 0.039048947393894196, 0.16778486967086792, 0.06479549407958984 ], "probs": [ 0.605466902256012, 0.07487740367650986, 0.21214313805103302, 0.10751249641180038 ], "argmax_index": 0, "argmax": "a", "top2_margin": 0.39332376420497894, "near_tie": false, "gold": "c", "correct": false, "host_softmax_max_abs_dp": 1.4901161193847656e-08, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "answer": { "type": "choice", "choice": "a", "confidence": 0.605466902256012, "probabilities": { "a": 0.605466902256012, "b": 0.07487740367650986, "c": 0.21214313805103302, "d": 0.10751249641180038 } } }, "usage": { "input_tokens": 147, "output_tokens": 0 } }, "oracle_file": "records_ref.json" }, { "id": "tv4_059", "mode": "text", "source": "transfer_v4_dev_head", "native": true, "prefix": 0, "usage_input_tokens": 115, "questions": [ { "qid": "answer", "type": "choice", "K": 4, "calibrate": true, "temperature_key": "choice:3-5", "T": 1.3998981714248657, "prefix": 0, "n_ids": 115, "positions": 115, "bucket": 256, "ids": [ 1, 17, 38685, 42505, 6881, 997, 17756, 2254, 18849, 1878, 997, 39420, 6881, 997, 28828, 803, 779, 2773, 856, 2492, 3347, 803, 6448, 18328, 8523, 602, 18, 28828, 7832, 16066, 13041, 779, 3952, 540, 19, 16, 768, 535, 7817, 938, 1209, 3849, 811, 18328, 2203, 896, 856, 880, 27354, 806, 916, 1149, 5344, 5812, 7321, 523, 20, 19, 16, 794, 535, 7817, 1188, 1014, 18328, 2203, 10967, 936, 856, 11611, 797, 3822, 6448, 523, 20, 19, 16, 781, 535, 7817, 1052, 1295, 4978, 811, 18328, 2203, 896, 1178, 1381, 1741, 40927, 523, 20, 19, 16, 778, 535, 7817, 1188, 18328, 34045, 2378, 28450, 3037, 1209, 11675, 1399, 1186, 1188, 8753, 803, 5743, 523, 20, 21 ], "markers": [ 35, 58, 76, 94 ], "logits_raw": [ 3.9577229022979736, -3.6439967155456543, -2.666415214538574, -0.6292939186096191 ], "probs_raw": [ 0.9881318807601929, 0.0004936622572131455, 0.0013121648225933313, 0.01006233412772417 ], "probs": [ 0.951524019241333, 0.004169808700680733, 0.008382895030081272, 0.03592329099774361 ], "argmax_index": 0, "argmax": "a", "top2_margin": 0.9156007282435894, "near_tie": false, "gold": "a", "correct": true, "host_softmax_max_abs_dp": 4.656612873077393e-10, "batch": { "call": 0, "index": 0, "B": 1 } } ], "response": { "answers": { "answer": { "type": "choice", "choice": "a", "confidence": 0.951524019241333, "probabilities": { "a": 0.951524019241333, "b": 0.004169808700680733, "c": 0.008382895030081272, "d": 0.03592329099774361 } } }, "usage": { "input_tokens": 115, "output_tokens": 0 } }, "oracle_file": "records_ref.json" } ] }