{ "rows": 555, "unique_documents": 127, "cache_sha256": "66ed25a1d7a030a7a803eeb6b4e5b48e220f9784e48cf5357d2a3d8a65a753bd", "small_model_version": "3.7.1", "ner": [ { "model": "spacy_small", "gold_set": "masked", "tp": 4442, "fp": 12516, "fn": 1642, "precision": 0.26194126665880413, "recall": 0.7301117685733071, "f1": 0.38555680930474784 }, { "model": "spacy_small", "gold_set": "all", "tp": 8817, "fp": 8141, "fn": 3212, "precision": 0.519931595707041, "recall": 0.7329786349655001, "f1": 0.6083416704039742 }, { "model": "roberta_names_only", "gold_set": "masked", "tp": 5120, "fp": 1250, "fn": 964, "precision": 0.8037676609105181, "recall": 0.84155161078238, "f1": 0.8222257909105508 }, { "model": "roberta_names_only", "gold_set": "all", "tp": 5766, "fp": 604, "fn": 6263, "precision": 0.9051805337519623, "recall": 0.4793415911547095, "f1": 0.6267731941953367 } ], "policies": [ { "policy": "roberta_spans", "gold_set": "masked", "mentions": 6084, "fully_removed": 4824, "rows": 555, "clean_rows": 277, "removed_chars": 250359, "total_chars": 2646610, "non_target_removed": 177993, "non_target_chars": 2554933, "person_n": 4138, "person_full": 3681, "org_n": 1946, "org_full": 1143, "full_removal": 0.7928994082840237, "clean_rate": 0.4990990990990991, "text_removed": 0.09459610596196644, "non_target_removal": 0.06966640612493556 }, { "policy": "roberta_spans", "gold_set": "all", "mentions": 12029, "fully_removed": 5347, "rows": 555, "clean_rows": 0, "removed_chars": 250359, "total_chars": 2646610, "non_target_removed": 168533, "non_target_chars": 2431595, "person_n": 4685, "person_full": 4037, "org_n": 7344, "org_full": 1310, "full_removal": 0.4445091030010807, "clean_rate": 0.0, "text_removed": 0.09459610596196644, "non_target_removal": 0.06930965066139715 }, { "policy": "majority_spans", "gold_set": "masked", "mentions": 6084, "fully_removed": 5071, "rows": 555, "clean_rows": 311, "removed_chars": 254946, "total_chars": 2646610, "non_target_removed": 180872, "non_target_chars": 2554933, "person_n": 4138, "person_full": 3870, "org_n": 1946, "org_full": 1201, "full_removal": 0.8334976988823143, "clean_rate": 0.5603603603603604, "text_removed": 0.09632926649563026, "non_target_removal": 0.07079324585028257 }, { "policy": "majority_spans", "gold_set": "all", "mentions": 12029, "fully_removed": 5700, "rows": 555, "clean_rows": 0, "removed_chars": 254946, "total_chars": 2646610, "non_target_removed": 170075, "non_target_chars": 2431595, "person_n": 4685, "person_full": 4274, "org_n": 7344, "org_full": 1426, "full_removal": 0.47385485077728823, "clean_rate": 0.0, "text_removed": 0.09632926649563026, "non_target_removal": 0.06994380231905395 }, { "policy": "union_spans", "gold_set": "masked", "mentions": 6084, "fully_removed": 5187, "rows": 555, "clean_rows": 348, "removed_chars": 268128, "total_chars": 2646610, "non_target_removed": 192127, "non_target_chars": 2554933, "person_n": 4138, "person_full": 3934, "org_n": 1946, "org_full": 1253, "full_removal": 0.8525641025641025, "clean_rate": 0.6270270270270271, "text_removed": 0.10130997766954708, "non_target_removal": 0.07519844943096356 }, { "policy": "union_spans", "gold_set": "all", "mentions": 12029, "fully_removed": 5931, "rows": 555, "clean_rows": 0, "removed_chars": 268128, "total_chars": 2646610, "non_target_removed": 178203, "non_target_chars": 2431595, "person_n": 4685, "person_full": 4367, "org_n": 7344, "org_full": 1564, "full_removal": 0.49305844209826255, "clean_rate": 0.0, "text_removed": 0.10130997766954708, "non_target_removal": 0.07328646423438114 }, { "policy": "union_sentences", "gold_set": "masked", "mentions": 6084, "fully_removed": 5854, "rows": 555, "clean_rows": 485, "removed_chars": 1689869, "total_chars": 2646610, "non_target_removed": 1601657, "non_target_chars": 2554933, "person_n": 4138, "person_full": 4050, "org_n": 1946, "org_full": 1804, "full_removal": 0.9621959237343852, "clean_rate": 0.8738738738738738, "text_removed": 0.6385032173232928, "non_target_removal": 0.6268880632094853 }, { "policy": "union_sentences", "gold_set": "all", "mentions": 12029, "fully_removed": 10581, "rows": 555, "clean_rows": 131, "removed_chars": 1689869, "total_chars": 2646610, "non_target_removed": 1498090, "non_target_chars": 2431595, "person_n": 4685, "person_full": 4541, "org_n": 7344, "org_full": 6040, "full_removal": 0.8796242414165766, "clean_rate": 0.23603603603603604, "text_removed": 0.6385032173232928, "non_target_removal": 0.6160935517633488 }, { "policy": "suppress_all", "gold_set": "masked", "mentions": 6084, "fully_removed": 6084, "rows": 555, "clean_rows": 555, "removed_chars": 2646610, "total_chars": 2646610, "non_target_removed": 2554933, "non_target_chars": 2554933, "person_n": 4138, "person_full": 4138, "org_n": 1946, "org_full": 1946, "full_removal": 1.0, "clean_rate": 1.0, "text_removed": 1.0, "non_target_removal": 1.0 }, { "policy": "suppress_all", "gold_set": "all", "mentions": 12029, "fully_removed": 12029, "rows": 555, "clean_rows": 555, "removed_chars": 2646610, "total_chars": 2646610, "non_target_removed": 2431595, "non_target_chars": 2431595, "person_n": 4685, "person_full": 4685, "org_n": 7344, "org_full": 7344, "full_removal": 1.0, "clean_rate": 1.0, "text_removed": 1.0, "non_target_removal": 1.0 } ], "notes": [ "NER: PERSON/ORG only, DIRECT+QUASI gold, typed partial matching; small model run live, RoBERTa replayed at 512 tokens.", "Redaction uses predictions of ALL entity types; union combines character coverage without dropping overlaps.", "Full removal means every gold mention character removed; clean rate is over annotation rows, not unique judgments.", "Text removed measures all characters, including whitespace; non-target means outside the selected PERSON/ORG gold, not necessarily harmless text.", "Sentence policy removes punctuation/newline-delimited segments containing any prediction. Suppress-all is a zero-utility bound.", "Exploratory held-out replay, no new external validation, speed benchmarks or confidence-threshold sweep." ] }