{ "binding": { "generator_sha256": "ecfb91fa37a292def9eccb730bdcad1367cde9276ef1288e071f7ef7df3d039a", "ranker_sha256": "d6626acd4513a47bc5bda4cdd805db2b5731d3e251c04ea2579dc0522d5f2a05", "corpus": { "version": 3, "seed": 9821, "source_manifest": { "version": 1, "source": "data/processed/stories-512", "source_manifest": { "dataset": "roneneldan/TinyStories", "revision": "f54c09fd23315a6f9c86f9dc80f725de7d8f9c64", "license": "cdla-sharing-1.0", "source": "https://huggingface.co/datasets/roneneldan/TinyStories", "synthetic": true, "selection": "First unique documents in pinned source order; alternating validation/test documents", "tokenizer_train_documents": 50000, "vocab_size": 8192, "max_length": 512, "tokenizer_sha256": "c5478e3379227640a0cbf9c50d3814373d843e9418c146e43de8cc9412dd1283", "splits": { "train": { "documents": 500000, "windows": 514199, "text_tokens": 107857284, "complete_story_windows": 485910, "documents_sha256": "b4876f6394b78ca3560848f22195b8e125c2872621c857148b77f93e5264cc16", "array_sha256": "b1fde06202d29af0d1146276d8be1bb4e5ebac1e7d767f6d58a8ca43a1aa16f1" }, "validation": { "documents": 1000, "windows": 1008, "text_tokens": 187038, "complete_story_windows": 992, "documents_sha256": "c6c4866335d6d101e152a43617c39bc6b2f25f05be5c841b6a67a6765e6e7b08", "array_sha256": "92d568c80995bf436d18ab320b4644c96eed757f47a263270be4303f566fd703" }, "test": { "documents": 1000, "windows": 1012, "text_tokens": 188300, "complete_story_windows": 988, "documents_sha256": "03d78b126a718f6be7aed5fdb09cd492f4462e0aaf717ca48f81d7349a43087d", "array_sha256": "fb581db76905e21e777b1746b2c3ca5d3cb2c8b5462e15910c8b4e6d644250c9" } }, "tokenizer_source": "data/processed/stories-v2/tokenizer.json" }, "seed": 4801, "spans_per_window": 8, "max_words": 12, "max_span_tokens": 16, "tokenizer_sha256": "c5478e3379227640a0cbf9c50d3814373d843e9418c146e43de8cc9412dd1283", "splits": { "test": { "rejected_quality": 1, "rejected_duplicate": 0, "rejected_alignment": 0, "windows": 1011, "candidate_spans": 8088, "files": { "test.npy": "419c2dfe9a0fcf03f1203914c2df0400ea4601dff19806f9e3fdfd9441c403ad", "test_spans.npy": "41d835ed26a1786f6405b6d2435ed2f4c2dca22eb794b79d1a3d4cf32bca6e86", "test_source_rows.npy": "aa465ccaa78b501f63bdebe5769efb165cf7f28a8fabce3451e116d09dd8891a" } }, "validation": { "rejected_quality": 5, "rejected_duplicate": 0, "rejected_alignment": 0, "windows": 1003, "candidate_spans": 8024, "files": { "validation.npy": "f9eaf135ca2c8d291cf3d277357ed3f1dd7921305b9e51ebc79d66660f5640d8", "validation_spans.npy": "612b4b04db23a3778a08a29dfd042a8e38ccf583c9a17952e3e2ee23d3716918", "validation_source_rows.npy": "20ed44c2d8342befac05ea70d2bae83cc1b17d005886d992390bf710933321a2" } }, "train": { "rejected_quality": 1666, "rejected_duplicate": 0, "rejected_alignment": 0, "windows": 512533, "candidate_spans": 4100264, "files": { "train.npy": "22dafe745270f2214b94df0c0d92f8ab20ea7d7e068249b703e4b9f558613ff2", "train_spans.npy": "720e8d1e6845715c276f6813de084bd31a375b78a7523bcdb89374edf75aea40", "train_source_rows.npy": "81a587e95a81cfd053f8a6f5d9c9f5d6705dd1e3156aa4ea6bad34be39f46605" } } }, "supervision": "Self-supervised reconstruction; original spans are not verified corrections", "deduplication": "Normalized exact window hashes across splits; no general near-deduplication claim" }, "procedural_manifest": { "version": 2, "seed": 2841, "provenance": "Procedural rule-checked synthetic contrasts; no model teacher; no human labels.", "license": "Apache-2.0", "splits": { "train": { "contexts": 12000, "templates": [ 0, 1, 2, 3, 4, 5, 6, 7 ], "sha256": "05fea23d1bf65a4225d7b0c0e6d3fb288336248c5edc095b7e9410db3d9d26c2" }, "validation": { "contexts": 1200, "templates": [ 8, 9 ], "sha256": "93070cd4700665b65cbc6756012e99cf381e1aa3086bd973548c0e034a8487ec" }, "test": { "contexts": 300, "templates": [ 10, 11 ], "sha256": "aeb45357e4a7a451a46662ab43c566aecac15628b859855bd1db57d5f77449d6" } }, "mining": { "contexts": 1200, "added_negatives": 2699, "source_corpus": { "version": 2, "seed": 2841, "provenance": "Procedural rule-checked synthetic contrasts; no model teacher; no human labels.", "license": "Apache-2.0", "splits": { "train": { "contexts": 12000, "templates": [ 0, 1, 2, 3, 4, 5, 6, 7 ], "sha256": "ed50146f5c2e8fc76cfe515ff121930800ad55269478476152289b3667504a31" }, "validation": { "contexts": 1200, "templates": [ 8, 9 ], "sha256": "93070cd4700665b65cbc6756012e99cf381e1aa3086bd973548c0e034a8487ec" }, "test": { "contexts": 300, "templates": [ 10, 11 ], "sha256": "aeb45357e4a7a451a46662ab43c566aecac15628b859855bd1db57d5f77449d6" } } }, "generator_sha256": "ecfb91fa37a292def9eccb730bdcad1367cde9276ef1288e071f7ef7df3d039a", "seed": 701, "label_rule": "Closed-world exact accepted list; not human judged. Training split only." } }, "provenance": "Mixed procedural contrasts and synthetic TinyStories excerpts; no human preference labels.", "license": "Corpus excerpts CDLA-Sharing-1.0; procedural material Apache-2.0.", "deduplication": "Exact normalized natural excerpts across splits; original source document splits retained. No near-duplicate claim.", "splits": { "test": { "contexts": 600, "sha256": "56bde01c131ab9300a4d448306d835c655f2d88676a3ef4843751fe857656c11" }, "validation": { "contexts": 1200, "sha256": "37bb6f84175073e705a818c7ae0fabcd630f53bd8d261a76ac37cc3c5c78f186" }, "train": { "contexts": 24000, "sha256": "7ff87556c13ffe59bd650fd65cc16e940a0fa5e7f2f9e5175553cb525439348c" } } }, "validation_contexts": 240, "precision": "bf16 generator / fp32 ranker", "device": "cuda", "seed": 301 }, "selection": { "blending": { "weight": 0.05, "formula": "-local_nll + weight * clip(learned_score, -8, 8)", "selection": "Validation top1 improvement with no top1 loss in either corpus or procedural domain; ties prefer lower weight.", "baseline": { "corpus": 0.39166666666666666, "procedural": 0.4166666666666667 }, "trials": [ { "weight": 0.0, "top1_by_domain": { "corpus": 0.39166666666666666, "procedural": 0.4166666666666667 }, "passes": true }, { "weight": 0.05, "top1_by_domain": { "corpus": 0.4083333333333333, "procedural": 0.5583333333333333 }, "passes": true }, { "weight": 0.1, "top1_by_domain": { "corpus": 0.4, "procedural": 0.5583333333333333 }, "passes": true }, { "weight": 0.2, "top1_by_domain": { "corpus": 0.39166666666666666, "procedural": 0.5583333333333333 }, "passes": true }, { "weight": 0.4, "top1_by_domain": { "corpus": 0.38333333333333336, "procedural": 0.5583333333333333 }, "passes": false }, { "weight": 0.8, "top1_by_domain": { "corpus": 0.35, "procedural": 0.5583333333333333 }, "passes": false }, { "weight": 1.6, "top1_by_domain": { "corpus": 0.31666666666666665, "procedural": 0.5583333333333333 }, "passes": false } ] }, "calibration": { "margin_threshold": 1.5, "score_floor": -2, "validation": { "error_cases": 240, "clean_cases": 240, "error_corrected": 0.11666666666666667, "clean_preserved": 0.9958333333333333, "balanced_success": 0.55625, "wrong_replacement_on_errors": 0.09583333333333334, "clean_modified": 0.004166666666666652 }, "method": "Grid search on validation only: maximize balanced success subject to <=5% clean edits and <=10% wrong edits on errors.", "margin_grid": [ 0, 0.25, 0.5, 1, 1.5, 2, 3, 4, 6, 8, 12, 1000000.0 ], "score_grid": [ -1000000.0, -4, -2, 0, 1, 2, 3, 4, 6, 8, 1000000.0 ], "contexts": 240, "limitations": "Finite synthetic validation; thresholds are not guarantees on user text." } }, "test": { "all": { "nll_keep_or_replace": { "error_cases": 600, "clean_cases": 600, "error_corrected": 0.33166666666666667, "clean_preserved": 0.4633333333333333, "balanced_success": 0.39749999999999996, "wrong_replacement_on_errors": 0.6383333333333333, "clean_modified": 0.5366666666666666 }, "learned_keep_or_replace": { "error_cases": 600, "clean_cases": 600, "error_corrected": 0.085, "clean_preserved": 0.9983333333333333, "balanced_success": 0.5416666666666666, "wrong_replacement_on_errors": 0.10333333333333333, "clean_modified": 0.0016666666666667052 }, "same_pool_error_top1_nll": 0.34, "same_pool_error_top1_ranker": 0.435, "error_candidate_coverage": 0.47333333333333333, "context_preserved": 1.0 }, "corpus": { "nll_keep_or_replace": { "error_cases": 300, "clean_cases": 300, "error_corrected": 0.3466666666666667, "clean_preserved": 0.6333333333333333, "balanced_success": 0.49, "wrong_replacement_on_errors": 0.64, "clean_modified": 0.3666666666666667 }, "learned_keep_or_replace": { "error_cases": 300, "clean_cases": 300, "error_corrected": 0.15, "clean_preserved": 0.9966666666666667, "balanced_success": 0.5733333333333334, "wrong_replacement_on_errors": 0.19333333333333333, "clean_modified": 0.0033333333333332993 }, "same_pool_error_top1_nll": 0.3466666666666667, "same_pool_error_top1_ranker": 0.36333333333333334, "error_candidate_coverage": 0.43666666666666665, "context_preserved": 1.0 }, "procedural": { "nll_keep_or_replace": { "error_cases": 300, "clean_cases": 300, "error_corrected": 0.31666666666666665, "clean_preserved": 0.29333333333333333, "balanced_success": 0.305, "wrong_replacement_on_errors": 0.6366666666666667, "clean_modified": 0.7066666666666667 }, "learned_keep_or_replace": { "error_cases": 300, "clean_cases": 300, "error_corrected": 0.02, "clean_preserved": 1.0, "balanced_success": 0.51, "wrong_replacement_on_errors": 0.013333333333333334, "clean_modified": 0.0 }, "same_pool_error_top1_nll": 0.3333333333333333, "same_pool_error_top1_ranker": 0.5066666666666667, "error_candidate_coverage": 0.51, "context_preserved": 1.0 } }, "bootstrap": { "metric": "learned minus NLL balanced success", "paired_contexts": 600, "resamples": 1000, "percentile_95": [ 0.11249999999999993, 0.17500000000000004 ], "limitations": "Within this procedural test; does not measure real-world validity." }, "legacy": { "summary": { "nll": 0.75, "blended": 0.75 }, "rows": [ { "case": { "id": "challenge-000", "category": "emotion", "text": "Lily lost her favorite toy. She felt happy. Tears ran down her face.", "start": 37, "end": 42, "accepted": [ "sad", "very sad", "upset" ] }, "nll": "very sad", "blended": "very sad", "recommendation": { "action": "keep", "replacement": "happy", "reason": "insufficient_margin", "margin_threshold": 1.5, "score_floor": -2 }, "candidates": [ { "replacement": "very sad", "local_nll": 2.5403027534484863, "ranker_score": -2.1787380456924437, "raw_ranker_score": 7.23129415512085 }, { "replacement": "sad", "local_nll": 2.969944953918457, "ranker_score": -2.6120747327804565, "raw_ranker_score": 7.15740442276001 } ], "previous_ranker_top1": "sad" }, { "case": { "id": "challenge-001", "category": "emotion", "text": "Ben won the game. He felt sad. He smiled and cheered.", "start": 26, "end": 29, "accepted": [ "happy", "very happy", "excited", "proud" ] }, "nll": "happy", "blended": "happy", "recommendation": { "action": "replace", "replacement": "happy", "reason": "learned_preference", "margin_threshold": 1.5, "score_floor": -2 }, "candidates": [ { "replacement": "happy", "local_nll": 2.0426506996154785, "ranker_score": -1.6426506996154786, "raw_ranker_score": 8.441244125366211 }, { "replacement": "proud", "local_nll": 2.215261936187744, "ranker_score": -1.9322461366653443, "raw_ranker_score": 5.660315990447998 }, { "replacement": "a happy", "local_nll": 3.402787685394287, "ranker_score": -3.309392035007477, "raw_ranker_score": 1.867913007736206 } ], "previous_ranker_top1": "happy" }, { "case": { "id": "challenge-002", "category": "emotion", "text": "A loud crash woke Mia. She felt calm. Her hands shook with fear.", "start": 32, "end": 36, "accepted": [ "scared", "afraid", "frightened" ] }, "nll": "scared", "blended": "scared", "recommendation": { "action": "keep", "replacement": "calm", "reason": "insufficient_margin", "margin_threshold": 1.5, "score_floor": -2 }, "candidates": [ { "replacement": "scared", "local_nll": 2.9203615188598633, "ranker_score": -2.5235899448394776, "raw_ranker_score": 7.935431480407715 }, { "replacement": "helpless", "local_nll": 2.9384212493896484, "ranker_score": -2.5549498081207274, "raw_ranker_score": 7.669428825378418 }, { "replacement": "sad", "local_nll": 2.9780538082122803, "ranker_score": -2.782265305519104, "raw_ranker_score": 3.9157700538635254 }, { "replacement": "pain", "local_nll": 3.5786001682281494, "ranker_score": -3.343713164329529, "raw_ranker_score": 4.697740077972412 }, { "replacement": "hurt", "local_nll": 3.4479033946990967, "ranker_score": -3.642672824859619, "raw_ranker_score": -3.895388603210449 }, { "replacement": "sad,", "local_nll": 3.83996844291687, "ranker_score": -4.098251700401306, "raw_ranker_score": -5.165665149688721 } ], "previous_ranker_top1": "scared" }, { "case": { "id": "challenge-003", "category": "emotion", "text": "Sam had not eaten all day. He felt full. His stomach growled.", "start": 35, "end": 39, "accepted": [ "hungry", "very hungry" ] }, "nll": "so sad", "blended": "so sad", "recommendation": { "action": "keep", "replacement": "full", "reason": "insufficient_margin", "margin_threshold": 1.5, "score_floor": -2 }, "candidates": [ { "replacement": "so sad", "local_nll": 3.548433542251587, "ranker_score": -3.6275506556034087, "raw_ranker_score": -1.582342267036438 }, { "replacement": "sick", "local_nll": 4.002574443817139, "ranker_score": -4.327047228813171, "raw_ranker_score": -6.489455699920654 }, { "replacement": "scared", "local_nll": 4.584311485290527, "ranker_score": -4.541442203521728, "raw_ranker_score": 0.8573856353759766 }, { "replacement": "sad", "local_nll": 4.279909133911133, "ranker_score": -4.575915956497193, "raw_ranker_score": -5.920136451721191 }, { "replacement": "cold", "local_nll": 4.486077785491943, "ranker_score": -4.886077785491944, "raw_ranker_score": -9.27464485168457 }, { "replacement": "hot", "local_nll": 4.683385372161865, "ranker_score": -5.083385372161866, "raw_ranker_score": -8.594403266906738 } ], "previous_ranker_top1": "scared" }, { "case": { "id": "challenge-004", "category": "emotion", "text": "The baby had played for hours. She felt awake. She yawned and closed her eyes.", "start": 40, "end": 45, "accepted": [ "tired", "sleepy", "very tired" ] }, "nll": "very tired", "blended": "very tired", "recommendation": { "action": "replace", "replacement": "very tired", "reason": "learned_preference", "margin_threshold": 1.5, "score_floor": -2 }, "candidates": [ { "replacement": "very tired", "local_nll": 1.4291845560073853, "ranker_score": -1.0291845560073853, "raw_ranker_score": 8.47041130065918 }, { "replacement": "sleepy", "local_nll": 1.5798200368881226, "ranker_score": -1.19981529712677, "raw_ranker_score": 7.600094795227051 }, { "replacement": "tired", "local_nll": 1.6626222133636475, "ranker_score": -1.265163516998291, "raw_ranker_score": 7.949173927307129 }, { "replacement": "warm", "local_nll": 2.351592779159546, "ranker_score": -2.2616754472255707, "raw_ranker_score": 1.7983466386795044 } ], "previous_ranker_top1": "warm" }, { "case": { "id": "challenge-005", "category": "emotion", "text": "Dad came home with a puppy. Ana was angry. She smiled and hugged her new pet.", "start": 36, "end": 41, "accepted": [ "happy", "excited", "very happy" ] }, "nll": "so happy", "blended": "so happy", "recommendation": { "action": "replace", "replacement": "so happy", "reason": "learned_preference", "margin_threshold": 1.5, "score_floor": -2 }, "candidates": [ { "replacement": "so happy", "local_nll": 1.6790379285812378, "ranker_score": -1.6527889788150787, "raw_ranker_score": 0.5249789953231812 }, { "replacement": "happy", "local_nll": 2.0893747806549072, "ranker_score": -2.465674567222595, "raw_ranker_score": -7.52599573135376 } ], "previous_ranker_top1": "so happy" }, { "case": { "id": "challenge-006", "category": "object", "text": "Mia had a blue cup. She drank from her red cup. It was the same blue cup.", "start": 39, "end": 42, "accepted": [ "blue" ] }, "nll": "blue", "blended": "blue", "recommendation": { "action": "keep", "replacement": "red", "reason": "insufficient_margin", "margin_threshold": 1.5, "score_floor": -2 }, "candidates": [ { "replacement": "blue", "local_nll": 1.880785346031189, "ranker_score": -1.480785346031189, "raw_ranker_score": 9.894739151000977 }, { "replacement": "blue blue", "local_nll": 2.437495231628418, "ranker_score": -2.037495231628418, "raw_ranker_score": 8.045941352844238 }, { "replacement": "yellow", "local_nll": 2.061751127243042, "ranker_score": -2.461751127243042, "raw_ranker_score": -9.14158821105957 } ], "previous_ranker_top1": "blue" }, { "case": { "id": "challenge-007", "category": "object", "text": "Ben put on his red hat. His green hat kept his head warm. He loved its red color.", "start": 28, "end": 33, "accepted": [ "red" ] }, "nll": "red", "blended": "red", "recommendation": { "action": "keep", "replacement": "green", "reason": "insufficient_margin", "margin_threshold": 1.5, "score_floor": -2 }, "candidates": [ { "replacement": "red", "local_nll": 3.2721590995788574, "ranker_score": -2.8721590995788575, "raw_ranker_score": 10.475955963134766 }, { "replacement": "blue", "local_nll": 3.2733154296875, "ranker_score": -3.6733154296875, "raw_ranker_score": -9.486680030822754 }, { "replacement": "red hat", "local_nll": 3.6767120361328125, "ranker_score": -3.8299659729003905, "raw_ranker_score": -3.0650787353515625 } ], "previous_ranker_top1": "red" }, { "case": { "id": "challenge-008", "category": "object", "text": "The dog chased a ball. It picked up the book in its mouth and brought the ball back.", "start": 40, "end": 44, "accepted": [ "ball" ] }, "nll": "ball", "blended": "ball", "recommendation": { "action": "keep", "replacement": "book", "reason": "insufficient_margin", "margin_threshold": 1.5, "score_floor": -2 }, "candidates": [ { "replacement": "ball", "local_nll": 1.810939073562622, "ranker_score": -1.4262932300567628, "raw_ranker_score": 7.6929168701171875 }, { "replacement": "ball and", "local_nll": 2.643805503845215, "ranker_score": -2.5677929878234864, "raw_ranker_score": 1.5202503204345703 } ], "previous_ranker_top1": "ball and" }, { "case": { "id": "challenge-009", "category": "object", "text": "Lucy planted a seed. She watered the stone every day until the seed began to grow.", "start": 37, "end": 42, "accepted": [ "seed" ] }, "nll": "seed", "blended": "seed", "recommendation": { "action": "keep", "replacement": "stone", "reason": "insufficient_margin", "margin_threshold": 1.5, "score_floor": -2 }, "candidates": [ { "replacement": "seed", "local_nll": 1.1965181827545166, "ranker_score": -0.7965181827545166, "raw_ranker_score": 9.1006498336792 }, { "replacement": "seed every", "local_nll": 2.174877882003784, "ranker_score": -1.9485410928726197, "raw_ranker_score": 4.526735782623291 } ], "previous_ranker_top1": "seed every" }, { "case": { "id": "challenge-010", "category": "object", "text": "Tom opened his book. He read the shoe until he reached the last page.", "start": 33, "end": 37, "accepted": [ "book", "story" ] }, "nll": "book", "blended": "book", "recommendation": { "action": "keep", "replacement": "shoe", "reason": "insufficient_margin", "margin_threshold": 1.5, "score_floor": -2 }, "candidates": [ { "replacement": "book", "local_nll": 2.601956844329834, "ranker_score": -2.201956844329834, "raw_ranker_score": 9.65537166595459 }, { "replacement": "same book", "local_nll": 2.8530356884002686, "ranker_score": -2.458784651756287, "raw_ranker_score": 7.885020732879639 } ], "previous_ranker_top1": "same book" }, { "case": { "id": "challenge-011", "category": "object", "text": "Mom baked a cake. She cut the chair into slices and served the cake.", "start": 30, "end": 35, "accepted": [ "cake" ] }, "nll": "cake", "blended": "cake", "recommendation": { "action": "keep", "replacement": "chair", "reason": "insufficient_margin", "margin_threshold": 1.5, "score_floor": -2 }, "candidates": [ { "replacement": "cake", "local_nll": 2.4397940635681152, "ranker_score": -2.0397940635681153, "raw_ranker_score": 10.100774765014648 }, { "replacement": "candles", "local_nll": 3.092719078063965, "ranker_score": -2.692719078063965, "raw_ranker_score": 8.752116203308105 }, { "replacement": "dough on", "local_nll": 4.4553961753845215, "ranker_score": -4.646898531913758, "raw_ranker_score": -3.830047130584717 } ], "previous_ranker_top1": "cake" }, { "case": { "id": "challenge-012", "category": "state", "text": "The box was empty. There was something inside. It contained nothing at all.", "start": 29, "end": 38, "accepted": [ "nothing" ] }, "nll": "no surprise", "blended": "no surprise", "recommendation": { "action": "keep", "replacement": "something", "reason": "insufficient_margin", "margin_threshold": 1.5, "score_floor": -2 }, "candidates": [ { "replacement": "no surprise", "local_nll": 3.628756046295166, "ranker_score": -3.526484155654907, "raw_ranker_score": 2.045437812805176 }, { "replacement": "no", "local_nll": 3.7752432823181152, "ranker_score": -3.8383737444877624, "raw_ranker_score": -1.2626092433929443 } ], "previous_ranker_top1": "no surprise" }, { "case": { "id": "challenge-013", "category": "state", "text": "The rain soaked his coat. His coat was dry. Water dripped from it.", "start": 39, "end": 42, "accepted": [ "wet", "very wet" ] }, "nll": "wet", "blended": "wet", "recommendation": { "action": "keep", "replacement": "dry", "reason": "insufficient_margin", "margin_threshold": 1.5, "score_floor": -2 }, "candidates": [ { "replacement": "wet", "local_nll": 3.9919676780700684, "ranker_score": -3.5919676780700684, "raw_ranker_score": 9.105957984924316 }, { "replacement": "so cold", "local_nll": 3.992619037628174, "ranker_score": -3.844361925125122, "raw_ranker_score": 2.965142250061035 }, { "replacement": "cold", "local_nll": 4.217764854431152, "ranker_score": -4.617764854431153, "raw_ranker_score": -8.541912078857422 }, { "replacement": "hot", "local_nll": 4.500383377075195, "ranker_score": -4.900383377075196, "raw_ranker_score": -10.205334663391113 } ], "previous_ranker_top1": "wet" }, { "case": { "id": "challenge-014", "category": "state", "text": "The fire went out. The room grew hot. Everyone needed a blanket.", "start": 33, "end": 36, "accepted": [ "cold", "colder" ] }, "nll": "so big", "blended": "so big", "recommendation": { "action": "keep", "replacement": "hot", "reason": "insufficient_margin", "margin_threshold": 1.5, "score_floor": -2 }, "candidates": [ { "replacement": "so big", "local_nll": 3.5452218055725098, "ranker_score": -3.14522180557251, "raw_ranker_score": 8.143821716308594 }, { "replacement": "bigger", "local_nll": 3.871809720993042, "ranker_score": -4.264934539794922, "raw_ranker_score": -7.862496376037598 } ], "previous_ranker_top1": "so big" }, { "case": { "id": "challenge-015", "category": "state", "text": "Dad turned off the only lamp. The room became bright. Nobody could see in the dark.", "start": 46, "end": 52, "accepted": [ "dark" ] }, "nll": "dark", "blended": "dark", "recommendation": { "action": "keep", "replacement": "bright", "reason": "insufficient_margin", "margin_threshold": 1.5, "score_floor": -2 }, "candidates": [ { "replacement": "dark", "local_nll": 2.7137789726257324, "ranker_score": -2.3137789726257325, "raw_ranker_score": 11.245848655700684 }, { "replacement": "dark inside", "local_nll": 2.8494725227355957, "ranker_score": -2.8570137202739714, "raw_ranker_score": -0.1508239507675171 } ], "previous_ranker_top1": "dark" }, { "case": { "id": "challenge-016", "category": "state", "text": "The door was locked. Mia could open it. She needed the key first.", "start": 31, "end": 35, "accepted": [ "not open", "not unlock" ] }, "nll": "not open", "blended": "not open", "recommendation": { "action": "keep", "replacement": "open", "reason": "insufficient_margin", "margin_threshold": 1.5, "score_floor": -2 }, "candidates": [ { "replacement": "not open", "local_nll": 2.0344557762145996, "ranker_score": -1.9840773224830628, "raw_ranker_score": 1.0075690746307373 }, { "replacement": "unlock", "local_nll": 2.935807466506958, "ranker_score": -2.582952857017517, "raw_ranker_score": 7.057092189788818 }, { "replacement": "do", "local_nll": 2.998377799987793, "ranker_score": -3.316229248046875, "raw_ranker_score": -6.357028961181641 } ], "previous_ranker_top1": "not open" }, { "case": { "id": "challenge-017", "category": "state", "text": "Lucy gave away all her cookies. She had many cookies left. Her plate was empty.", "start": 40, "end": 44, "accepted": [ "no", "zero" ] }, "nll": "no", "blended": "no", "recommendation": { "action": "keep", "replacement": "many", "reason": "insufficient_margin", "margin_threshold": 1.5, "score_floor": -2 }, "candidates": [ { "replacement": "no", "local_nll": 2.2782347202301025, "ranker_score": -2.671682596206665, "raw_ranker_score": -7.86895751953125 }, { "replacement": "a yummy", "local_nll": 3.1334025859832764, "ranker_score": -3.0714222967624663, "raw_ranker_score": 1.2396057844161987 }, { "replacement": "all", "local_nll": 3.5266520977020264, "ranker_score": -3.546010452508926, "raw_ranker_score": -0.3871670961380005 } ], "previous_ranker_top1": "a yummy" }, { "case": { "id": "challenge-018", "category": "location", "text": "The children stayed inside the house. They played outside all afternoon without leaving the house.", "start": 50, "end": 57, "accepted": [ "inside", "indoors" ] }, "nll": "and played", "blended": "and played", "recommendation": { "action": "keep", "replacement": "outside", "reason": "insufficient_margin", "margin_threshold": 1.5, "score_floor": -2 }, "candidates": [ { "replacement": "and played", "local_nll": 2.7578656673431396, "ranker_score": -2.544952630996704, "raw_ranker_score": 4.258260726928711 }, { "replacement": "happily", "local_nll": 3.2870261669158936, "ranker_score": -3.3314939022064207, "raw_ranker_score": -0.8893547058105469 }, { "replacement": "and", "local_nll": 3.984219551086426, "ranker_score": -4.0055818289518355, "raw_ranker_score": -0.427245557308197 }, { "replacement": "until", "local_nll": 4.426279067993164, "ranker_score": -4.675834584236145, "raw_ranker_score": -4.991110324859619 } ], "previous_ranker_top1": "until" }, { "case": { "id": "challenge-019", "category": "location", "text": "Tom put his toy under the bed. Later he found it on the bed, exactly where he had left it.", "start": 49, "end": 51, "accepted": [ "under", "underneath", "beneath" ] }, "nll": "under", "blended": "under", "recommendation": { "action": "keep", "replacement": "on", "reason": "insufficient_margin", "margin_threshold": 1.5, "score_floor": -2 }, "candidates": [ { "replacement": "under", "local_nll": 2.5232717990875244, "ranker_score": -2.1232717990875245, "raw_ranker_score": 10.067623138427734 }, { "replacement": "was in", "local_nll": 3.0939877033233643, "ranker_score": -3.266255259513855, "raw_ranker_score": -3.4453511238098145 }, { "replacement": "in", "local_nll": 2.9847395420074463, "ranker_score": -3.384739542007446, "raw_ranker_score": -8.055986404418945 } ], "previous_ranker_top1": "under" }, { "case": { "id": "challenge-020", "category": "location", "text": "The bird sat in its nest. It stayed in the pond all night and left the nest at dawn.", "start": 43, "end": 47, "accepted": [ "nest" ] }, "nll": "nest", "blended": "nest", "recommendation": { "action": "keep", "replacement": "pond", "reason": "insufficient_margin", "margin_threshold": 1.5, "score_floor": -2 }, "candidates": [ { "replacement": "nest", "local_nll": 4.082644462585449, "ranker_score": -3.6826444625854493, "raw_ranker_score": 8.184283256530762 }, { "replacement": "nest for", "local_nll": 4.2003254890441895, "ranker_score": -3.807837176322937, "raw_ranker_score": 7.849766254425049 } ], "previous_ranker_top1": "nest for" }, { "case": { "id": "challenge-021", "category": "location", "text": "The fish lived in a pond. It swam in the sky while the other fish watched.", "start": 41, "end": 44, "accepted": [ "pond", "water" ] }, "nll": "pond,", "blended": "pond,", "recommendation": { "action": "keep", "replacement": "sky", "reason": "insufficient_margin", "margin_threshold": 1.5, "score_floor": -2 }, "candidates": [ { "replacement": "pond,", "local_nll": 2.124826192855835, "ranker_score": -1.7782915353775024, "raw_ranker_score": 6.93069314956665 }, { "replacement": "pond", "local_nll": 2.255293369293213, "ranker_score": -1.855293369293213, "raw_ranker_score": 9.653685569763184 } ], "previous_ranker_top1": "pond" }, { "case": { "id": "challenge-022", "category": "location", "text": "Mia hid the key in her pocket. She reached into her shoe and took the key out of her pocket.", "start": 52, "end": 56, "accepted": [ "pocket" ] }, "nll": "pocket", "blended": "pocket", "recommendation": { "action": "keep", "replacement": "shoe", "reason": "insufficient_margin", "margin_threshold": 1.5, "score_floor": -2 }, "candidates": [ { "replacement": "pocket", "local_nll": 1.2805202007293701, "ranker_score": -0.8805202007293701, "raw_ranker_score": 8.95186996459961 }, { "replacement": "pocket and", "local_nll": 2.039170265197754, "ranker_score": -2.3012430667877197, "raw_ranker_score": -5.241456031799316 } ], "previous_ranker_top1": "pocket" }, { "case": { "id": "challenge-023", "category": "location", "text": "The children sat at a table. Their plates were on the floor. They ate from the table.", "start": 54, "end": 59, "accepted": [ "table" ] }, "nll": "table", "blended": "table", "recommendation": { "action": "keep", "replacement": "floor", "reason": "insufficient_margin", "margin_threshold": 1.5, "score_floor": -2 }, "candidates": [ { "replacement": "table", "local_nll": 2.022895097732544, "ranker_score": -1.6937190294265747, "raw_ranker_score": 6.583521366119385 }, { "replacement": "table outside", "local_nll": 2.8226521015167236, "ranker_score": -2.968308663368225, "raw_ranker_score": -2.9131312370300293 } ], "previous_ranker_top1": "table outside" }, { "case": { "id": "challenge-024", "category": "identity", "text": "Lily was a little girl. He put on her coat and went out.", "start": 24, "end": 26, "accepted": [ "She" ] }, "nll": "She", "blended": "She", "recommendation": { "action": "keep", "replacement": "He", "reason": "insufficient_margin", "margin_threshold": 1.5, "score_floor": -2 }, "candidates": [ { "replacement": "She", "local_nll": 2.7060937881469727, "ranker_score": -2.3060937881469727, "raw_ranker_score": 9.46899127960205 }, { "replacement": "She always", "local_nll": 2.741779327392578, "ranker_score": -2.8433783531188963, "raw_ranker_score": -2.031980514526367 } ], "previous_ranker_top1": "She" }, { "case": { "id": "challenge-025", "category": "identity", "text": "Tom was a little boy. She picked up his toy and smiled.", "start": 22, "end": 25, "accepted": [ "He" ] }, "nll": "He", "blended": "He", "recommendation": { "action": "keep", "replacement": "She", "reason": "insufficient_margin", "margin_threshold": 1.5, "score_floor": -2 }, "candidates": [ { "replacement": "He", "local_nll": 3.0154550075531006, "ranker_score": -2.6154550075531007, "raw_ranker_score": 9.00454044342041 }, { "replacement": "He had", "local_nll": 3.18315052986145, "ranker_score": -2.937783694267273, "raw_ranker_score": 4.907336711883545 } ], "previous_ranker_top1": "He" }, { "case": { "id": "challenge-026", "category": "identity", "text": "Two girls went to the park. She played together until it was time to go home.", "start": 28, "end": 31, "accepted": [ "They" ] }, "nll": "They", "blended": "They", "recommendation": { "action": "keep", "replacement": "She", "reason": "insufficient_margin", "margin_threshold": 1.5, "score_floor": -2 }, "candidates": [ { "replacement": "They", "local_nll": 1.0522031784057617, "ranker_score": -0.6564175605773925, "raw_ranker_score": 7.915712356567383 }, { "replacement": "They both", "local_nll": 1.547157883644104, "ranker_score": -1.7885940790176391, "raw_ranker_score": -4.828723907470703 } ], "previous_ranker_top1": "They" }, { "case": { "id": "challenge-027", "category": "identity", "text": "The puppy was a small dog. The cat wagged its tail and barked.", "start": 31, "end": 34, "accepted": [ "dog", "puppy" ] }, "nll": "puppy", "blended": "puppy", "recommendation": { "action": "keep", "replacement": "cat", "reason": "insufficient_margin", "margin_threshold": 1.5, "score_floor": -2 }, "candidates": [ { "replacement": "puppy", "local_nll": 1.750527024269104, "ranker_score": -1.350527024269104, "raw_ranker_score": 8.539040565490723 }, { "replacement": "dog was", "local_nll": 2.219879627227783, "ranker_score": -1.884831166267395, "raw_ranker_score": 6.700969219207764 } ], "previous_ranker_top1": "puppy" }, { "case": { "id": "challenge-028", "category": "identity", "text": "Mia and Ben were friends. He both liked the same game.", "start": 26, "end": 28, "accepted": [ "They" ] }, "nll": "They", "blended": "They", "recommendation": { "action": "keep", "replacement": "He", "reason": "insufficient_margin", "margin_threshold": 1.5, "score_floor": -2 }, "candidates": [ { "replacement": "They", "local_nll": 3.124868392944336, "ranker_score": -2.724868392944336, "raw_ranker_score": 8.637487411499023 }, { "replacement": "They were", "local_nll": 4.150475978851318, "ranker_score": -3.7802287578582763, "raw_ranker_score": 7.40494441986084 } ], "previous_ranker_top1": "They" }, { "case": { "id": "challenge-029", "category": "identity", "text": "The bird built a nest. The fish laid an egg in its nest.", "start": 27, "end": 31, "accepted": [ "bird" ] }, "nll": "bird", "blended": "bird", "recommendation": { "action": "keep", "replacement": "fish", "reason": "insufficient_margin", "margin_threshold": 1.5, "score_floor": -2 }, "candidates": [ { "replacement": "bird", "local_nll": 2.5359787940979004, "ranker_score": -2.149574708938599, "raw_ranker_score": 7.728081703186035 }, { "replacement": "bird had", "local_nll": 2.8026721477508545, "ranker_score": -2.6933255076408384, "raw_ranker_score": 2.1869328022003174 }, { "replacement": "nest", "local_nll": 3.3174939155578613, "ranker_score": -3.6228899717330934, "raw_ranker_score": -6.107921123504639 }, { "replacement": "egg", "local_nll": 3.9023170471191406, "ranker_score": -4.286371040344238, "raw_ranker_score": -7.681079864501953 } ], "previous_ranker_top1": "nest" }, { "case": { "id": "challenge-030", "category": "agreement", "text": "Yesterday, Tom walk to the park. He stayed there all afternoon.", "start": 15, "end": 19, "accepted": [ "walked" ] }, "nll": "went", "blended": "went", "recommendation": { "action": "replace", "replacement": "went", "reason": "learned_preference", "margin_threshold": 1.5, "score_floor": -2 }, "candidates": [ { "replacement": "went", "local_nll": 2.0387723445892334, "ranker_score": -1.852819013595581, "raw_ranker_score": 3.719066619873047 }, { "replacement": "went outside", "local_nll": 2.3577680587768555, "ranker_score": -2.2333672881126403, "raw_ranker_score": 2.4880154132843018 } ], "previous_ranker_top1": "went" }, { "case": { "id": "challenge-031", "category": "agreement", "text": "Every morning, she eat her breakfast before school.", "start": 19, "end": 22, "accepted": [ "eats" ] }, "nll": "ate", "blended": "ate", "recommendation": { "action": "keep", "replacement": "eat", "reason": "insufficient_margin", "margin_threshold": 1.5, "score_floor": -2 }, "candidates": [ { "replacement": "ate", "local_nll": 2.83488392829895, "ranker_score": -2.4494802713394166, "raw_ranker_score": 7.708073139190674 }, { "replacement": "waited for", "local_nll": 3.1999478340148926, "ranker_score": -3.1705774128437043, "raw_ranker_score": 0.5874084234237671 }, { "replacement": "had", "local_nll": 3.6242282390594482, "ranker_score": -3.3885207653045653, "raw_ranker_score": 4.714149475097656 }, { "replacement": "put", "local_nll": 4.452193737030029, "ranker_score": -4.581978976726532, "raw_ranker_score": -2.5957047939300537 } ], "previous_ranker_top1": "ate" }, { "case": { "id": "challenge-032", "category": "agreement", "text": "There was three apples on the table.", "start": 6, "end": 9, "accepted": [ "were", "are" ] }, "nll": "were", "blended": "were", "recommendation": { "action": "keep", "replacement": "was", "reason": "insufficient_margin", "margin_threshold": 1.5, "score_floor": -2 }, "candidates": [ { "replacement": "were", "local_nll": 3.312285900115967, "ranker_score": -2.9635926723480224, "raw_ranker_score": 6.973864555358887 }, { "replacement": "was a", "local_nll": 3.4742205142974854, "ranker_score": -3.8742205142974853, "raw_ranker_score": -8.374324798583984 } ], "previous_ranker_top1": "were" }, { "case": { "id": "challenge-033", "category": "agreement", "text": "The two dogs was playing together in the garden.", "start": 13, "end": 16, "accepted": [ "were" ] }, "nll": "were", "blended": "were", "recommendation": { "action": "replace", "replacement": "were", "reason": "learned_preference", "margin_threshold": 1.5, "score_floor": -2 }, "candidates": [ { "replacement": "were", "local_nll": 0.8926716446876526, "ranker_score": -0.5096674084663391, "raw_ranker_score": 7.6600847244262695 }, { "replacement": "were always", "local_nll": 1.613624095916748, "ranker_score": -1.8262638807296754, "raw_ranker_score": -4.252795696258545 } ], "previous_ranker_top1": "were" }, { "case": { "id": "challenge-034", "category": "agreement", "text": "Yesterday, Mia go home after school.", "start": 15, "end": 17, "accepted": [ "went" ] }, "nll": "went", "blended": "went", "recommendation": { "action": "keep", "replacement": "go", "reason": "insufficient_margin", "margin_threshold": 1.5, "score_floor": -2 }, "candidates": [ { "replacement": "went", "local_nll": 3.7416648864746094, "ranker_score": -3.5823410749435425, "raw_ranker_score": 3.186476230621338 }, { "replacement": "and her", "local_nll": 4.225393772125244, "ranker_score": -4.365866041183471, "raw_ranker_score": -2.809445381164551 } ], "previous_ranker_top1": "went" }, { "case": { "id": "challenge-035", "category": "agreement", "text": "He gave me a apple. I ate the apple.", "start": 11, "end": 12, "accepted": [ "an" ] }, "nll": "the", "blended": "the", "recommendation": { "action": "keep", "replacement": "a", "reason": "insufficient_margin", "margin_threshold": 1.5, "score_floor": -2 }, "candidates": [ { "replacement": "the", "local_nll": 2.6240718364715576, "ranker_score": -2.760662925243378, "raw_ranker_score": -2.7318217754364014 }, { "replacement": "the banana", "local_nll": 3.644852876663208, "ranker_score": -3.489479100704193, "raw_ranker_score": 3.107475519180298 } ], "previous_ranker_top1": "the banana" } ], "previous_ranker_matched_pool": { "top1": 0.5555555555555556, "device": "cpu", "precision": "fp32", "weights_sha256": "922edda7af28191f4f0f735ea20cb1d833a77a566fce3fe8e184bca2a16f0deb", "candidate_pool": "exact stored candidate pool shared with NLL and v3; no regeneration" } }, "reference_pool_test": { "pool_top1": 0.9916666666666667, "bce": 0.04809597507119179, "positive_mean": 8.719537399663887, "negative_mean": -7.907711292239344, "contexts": 600 }, "limits": [ "TinyStories is synthetic. No human preference labels.", "Corpus metric is original-span reconstruction, not semantic correctness.", "Procedural cases are the previous benchmark reused for regression tracking.", "Legacy probes are repeated diagnostics, not a new unseen test." ] }