# dialogpii-multi - multi / pii (500 rows, 10345 spans, 0 negatives)

Main metric: **missed** - gold spans no predicted character touched. Score threshold 0.5; spans without a score always count. `hidden` - gold spans every character of which is covered: touching one character counts as detection, not as hiding. `hidden` is counted over normalized predictions - a prediction is stretched to whole words before it is compared, so it is what a masker that repeats the same normalization would hide; a masker that masks the raw offsets of the model hides no more than this. Char P / R / F1 are reference only. Trivial baseline, mask everything: 0 missed, 100% hidden, char P 0.055 (94.5% of the text over-masked). `rows touched` / `chars masked` count rows the source left without annotations where a mask touched something, and the characters masked in them; that absence of annotations is not evidence that those rows hold nothing sensitive, so this is not a false-alarm rate. The exact CSVs keep the technical column names `fp_rows` and `fp_chars`.

| model | missed | missed % [95% CI] | hidden | char F1 [95% CI] | P | R | entity exact | entity overlap | typed F1 | rows touched | chars masked | dropped spans | ms/row | params | train |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| pii-shield-onnx | **1837** | 17.8% [16.0%, 19.8%] | 74.3% | 0.119 [0.113, 0.126] | 0.064 | 0.827 | 0.041 | 0.239 | 0.005 | 0/0 | 0 | 0 | 2120 | - |  |
| gliner2-fastino | **2286** | 22.1% [19.8%, 24.6%] | 71.7% | 0.538 [0.521, 0.554] | 0.431 | 0.714 | 0.441 | 0.563 | 0.430 | 0/0 | 0 | 0 | 56 | 307M |  |
| bardsai-eu | **2373** | 22.9% [21.3%, 24.6%] | 64.7% | 0.718 [0.703, 0.733] | 0.794 | 0.656 | 0.593 | 0.810 | 0.168 | 0/0 | 0 | 0 | 1132 | - |  |
| gliner-urchade | **2579** | 24.9% [22.7%, 27.2%] | 70.5% | 0.459 [0.446, 0.474] | 0.343 | 0.694 | 0.365 | 0.474 | 0.377 | 0/0 | 0 | 0 | 38 | 289M |  |
| gliner2-hivetrace-omni | **2694** | 26.0% [23.9%, 28.1%] | 66.3% | 0.604 [0.586, 0.621] | 0.558 | 0.658 | 0.520 | 0.656 | 0.474 | 0/0 | 0 | 0 | 59 | 307M |  |
| nym-base | **2859** | 27.6% [25.8%, 29.4%] | 55.5% | 0.696 [0.678, 0.712] | 0.902 | 0.566 | 0.566 | 0.802 | 0.210 | 0/0 | 0 | 0 | 32 | 308M |  |
| pplx | **3017** | 29.2% [27.6%, 30.8%] | 65.2% | 0.679 [0.662, 0.695] | 0.789 | 0.595 | 0.580 | 0.775 | 0.611 | 0/0 | 0 | 0 | 534 | 596M |  |
| gliner-multi-v21 | **3094** | 29.9% [27.6%, 32.3%] | 64.1% | 0.458 [0.443, 0.474] | 0.368 | 0.605 | 0.387 | 0.498 | 0.367 | 0/0 | 0 | 0 | 39 | 289M |  |
| davlan-xlmr | **3163** | 30.6% [28.8%, 32.2%] | 60.6% | 0.680 [0.661, 0.700] | 0.920 | 0.539 | 0.646 | 0.799 | 0.474 | 0/0 | 0 | 0 | 38 | 277M |  |
| mmbert32k | **3267** | 31.6% [30.1%, 33.3%] | 49.7% | 0.485 [0.470, 0.499] | 0.451 | 0.524 | 0.341 | 0.555 | 0.314 | 0/0 | 0 | 0 | 46 | 308M |  |
| nuner-zero | **3290** | 31.8% [29.2%, 34.6%] | 62.7% | 0.443 [0.427, 0.459] | 0.343 | 0.626 | 0.359 | 0.459 | 0.391 | 0/0 | 0 | 0 | 113 | 449M |  |
| nym-small | **3315** | 32.0% [30.2%, 34.0%] | 51.4% | 0.667 [0.649, 0.684] | 0.902 | 0.529 | 0.530 | 0.768 | 0.175 | 0/0 | 0 | 0 | 944 | - |  |
| gliner-nvidia | **3357** | 32.5% [29.9%, 35.1%] | 57.3% | 0.479 [0.460, 0.499] | 0.418 | 0.561 | 0.425 | 0.542 | 0.408 | 0/0 | 0 | 0 | 112 | 445M |  |
| ner-ru-yqelz | **3592** | 34.7% [33.1%, 36.4%] | 56.3% | 0.644 [0.630, 0.659] | 0.699 | 0.597 | 0.513 | 0.685 | 0.436 | 0/0 | 0 | 0 | 46 | 559M |  |
| davlan-mbert | **3662** | 35.4% [33.5%, 37.3%] | 55.0% | 0.635 [0.616, 0.654] | 0.830 | 0.514 | 0.552 | 0.728 | 0.415 | 0/0 | 0 | 0 | 43 | 177M |  |
| gliner25-fastino | **3722** | 36.0% [33.7%, 38.5%] | 59.1% | 0.560 [0.541, 0.579] | 0.545 | 0.576 | 0.473 | 0.602 | 0.480 | 0/0 | 0 | 0 | 26 | 287M |  |
| gliner2-large | **4068** | 39.3% [36.5%, 42.1%] | 54.9% | 0.433 [0.416, 0.450] | 0.359 | 0.546 | 0.345 | 0.452 | 0.337 | 0/0 | 0 | 0 | 164 | 486M |  |
| opf-ru | **4181** | 40.4% [38.6%, 42.3%] | 44.6% | 0.359 [0.346, 0.373] | 0.287 | 0.480 | 0.234 | 0.403 | 0.236 | 0/0 | 0 | 0 | 285 | 1.4B |  |
| gliner-pii-base | **4306** | 41.6% [38.7%, 44.5%] | 53.3% | 0.532 [0.514, 0.552] | 0.551 | 0.515 | 0.444 | 0.566 | 0.432 | 0/0 | 0 | 0 | 36 | 166M |  |
| gliner2-hivetrace-uni | **4429** | 42.8% [40.5%, 45.1%] | 49.2% | 0.511 [0.492, 0.529] | 0.572 | 0.462 | 0.424 | 0.589 | 0.351 | 0/0 | 0 | 0 | 86 | 147M |  |
| openmed-multilingual | **4457** | 43.1% [40.8%, 45.5%] | 40.6% | 0.527 [0.507, 0.546] | 0.664 | 0.437 | 0.387 | 0.597 | 0.435 | 0/0 | 0 | 0 | 230 | 1.4B |  |
| gliner-pii-edge | **4460** | 43.1% [40.2%, 46.1%] | 51.1% | 0.537 [0.515, 0.560] | 0.602 | 0.485 | 0.425 | 0.586 | 0.378 | 0/0 | 0 | 0 | 659 | 45M |  |
| ner-ru-gherman | **4526** | 43.8% [41.7%, 45.6%] | 40.0% | 0.520 [0.501, 0.539] | 0.900 | 0.365 | 0.461 | 0.688 | 0.090 | 0/0 | 0 | 0 | 43 | 177M |  |
| ru-legal-ner | **4573** | 44.2% [41.6%, 46.9%] | 45.4% | 0.111 [0.105, 0.118] | 0.062 | 0.538 | 0.033 | 0.214 | 0.021 | 0/0 | 0 | 0 | 47 | 29M |  |
| stanza-ru | **4659** | 45.0% [42.2%, 47.9%] | 49.8% | 0.164 [0.157, 0.171] | 0.100 | 0.456 | 0.088 | 0.300 | 0.018 | 0/0 | 0 | 0 | 384 | - |  |
| gliner2-vladlinv | **4894** | 47.3% [45.1%, 49.6%] | 47.1% | 0.550 [0.529, 0.570] | 0.871 | 0.402 | 0.519 | 0.673 | 0.409 | 0/0 | 0 | 0 | 32 | 287M |  |
| openmed-nemotron | **5038** | 48.7% [47.2%, 50.4%] | 34.2% | 0.308 [0.294, 0.321] | 0.250 | 0.400 | 0.178 | 0.321 | 0.216 | 0/0 | 0 | 0 | 237 | 1.4B |  |
| gliner-stream-pii | **5460** | 52.8% [50.5%, 54.9%] | 39.8% | 0.518 [0.495, 0.540] | 0.796 | 0.383 | 0.452 | 0.591 | 0.372 | 0/0 | 0 | 0 | 247 | 677M |  |
| gravitee-small | **5585** | 54.0% [51.8%, 56.1%] | 39.4% | 0.249 [0.233, 0.265] | 0.183 | 0.388 | 0.146 | 0.326 | 0.202 | 0/0 | 0 | 0 | 46 | 29M |  |
| traciora | **5608** | 54.2% [52.6%, 55.9%] | 36.5% | 0.464 [0.447, 0.481] | 0.512 | 0.424 | 0.250 | 0.519 | 0.320 | 0/0 | 0 | 0 | 342 | 1.4B |  |
| kalyan-ettin | **5717** | 55.3% [52.6%, 57.8%] | 32.2% | 0.459 [0.432, 0.484] | 0.642 | 0.357 | 0.336 | 0.526 | 0.391 | 0/0 | 0 | 0 | 64 | 68M |  |
| ru-pii-ner | **6254** | 60.5% [57.9%, 63.0%] | 35.7% | 0.392 [0.373, 0.409] | 0.455 | 0.344 | 0.319 | 0.442 | 0.219 | 0/0 | 0 | 0 | 830 | 358M |  |
| apararti | **6390** | 61.8% [59.8%, 63.7%] | 30.8% | 0.414 [0.396, 0.433] | 0.461 | 0.375 | 0.244 | 0.423 | 0.310 | 0/0 | 0 | 0 | 434 | 1.4B |  |
| openai-base | **6634** | 64.1% [62.0%, 66.2%] | 31.4% | 0.363 [0.346, 0.381] | 0.369 | 0.357 | 0.211 | 0.370 | 0.288 | 0/0 | 0 | 0 | 498 | 1.4B |  |
| opf-kz-ru | **7288** | 70.4% [68.6%, 72.3%] | 24.6% | 0.367 [0.346, 0.389] | 0.463 | 0.304 | 0.210 | 0.359 | 0.290 | 0/0 | 0 | 0 | 398 | 1.4B |  |
| opf-ru-v2 | **7351** | 71.1% [69.4%, 72.7%] | 22.6% | 0.409 [0.389, 0.428] | 0.699 | 0.289 | 0.189 | 0.422 | 0.316 | 0/0 | 0 | 0 | 92 | 1.4B |  |
| spacy-ru-lg | **8270** | 79.9% [78.8%, 81.0%] | 18.4% | 0.123 [0.114, 0.130] | 0.103 | 0.152 | 0.094 | 0.215 | 0.020 | 0/0 | 0 | 0 | 59 | - |  |
| natasha | **9416** | 91.0% [90.1%, 91.9%] | 8.4% | 0.096 [0.087, 0.106] | 0.327 | 0.057 | 0.115 | 0.149 | 0.015 | 0/0 | 0 | 0 | 20 | - |  |
| rules-ru | **10054** | 97.2% [96.7%, 97.7%] | 2.1% | 0.100 [0.083, 0.116] | 0.980 | 0.053 | 0.041 | 0.055 | 0.005 | 0/0 | 0 | 0 | 4 | - |  |
| fef2-secret-ru | **10243** | 99.0% [98.8%, 99.2%] | 0.8% | 0.018 [0.014, 0.022] | 0.613 | 0.009 | 0.010 | 0.019 | 0.016 | 0/0 | 0 | 0 | 52 | 177M |  |
| spacy-alrosait | **10331** | 99.9% [99.8%, 99.9%] | 0.1% | 0.002 [0.001, 0.004] | 0.188 | 0.001 | 0.001 | 0.003 | 0.000 | 0/0 | 0 | 0 | 57 | - |  |

Neighbours by rank a paired bootstrap cannot tell apart (95%, missed %) - pairs, not a transitive chain, so `a ≈ b` and `b ≈ c` do not make `a ≈ c`: gliner2-fastino ≈ bardsai-eu; gliner-urchade ≈ gliner2-hivetrace-omni; gliner2-hivetrace-omni ≈ nym-base; pplx ≈ gliner-multi-v21; gliner-multi-v21 ≈ davlan-xlmr; davlan-xlmr ≈ mmbert32k; mmbert32k ≈ nuner-zero; nuner-zero ≈ nym-small; nym-small ≈ gliner-nvidia; gliner-nvidia ≈ ner-ru-yqelz; ner-ru-yqelz ≈ davlan-mbert; davlan-mbert ≈ gliner25-fastino; gliner2-large ≈ opf-ru; opf-ru ≈ gliner-pii-base; gliner-pii-base ≈ gliner2-hivetrace-uni; gliner2-hivetrace-uni ≈ openmed-multilingual; openmed-multilingual ≈ gliner-pii-edge; gliner-pii-edge ≈ ner-ru-gherman; ner-ru-gherman ≈ ru-legal-ner; ru-legal-ner ≈ stanza-ru; gliner2-vladlinv ≈ openmed-nemotron; gliner-stream-pii ≈ gravitee-small; gravitee-small ≈ traciora; traciora ≈ kalyan-ettin; ru-pii-ner ≈ apararti; opf-kz-ru ≈ opf-ru-v2

## Missed by group

| model | PERSON | ADDRESS | CONTACT | ID | NET | ORG |
|---|---|---|---|---|---|---|
| spans in gold | 5083 | 1868 | 750 | 818 | 20 | 1806 |
| pii-shield-onnx | 1025 (20.2%) | 329 (17.6%) | 66 (8.8%) | 119 (14.5%) | 2 (10.0%) | 296 (16.4%) |
| gliner2-fastino | 867 (17.1%) | 384 (20.6%) | 181 (24.1%) | 308 (37.7%) | 1 (5.0%) | 545 (30.2%) |
| bardsai-eu | 606 (11.9%) | 396 (21.2%) | 262 (34.9%) | 347 (42.4%) | 4 (20.0%) | 758 (42.0%) |
| gliner-urchade | 836 (16.4%) | 447 (23.9%) | 191 (25.5%) | 484 (59.2%) | 2 (10.0%) | 619 (34.3%) |
| gliner2-hivetrace-omni | 1208 (23.8%) | 439 (23.5%) | 187 (24.9%) | 215 (26.3%) | 3 (15.0%) | 642 (35.5%) |
| nym-base | 633 (12.5%) | 422 (22.6%) | 277 (36.9%) | 342 (41.8%) | 6 (30.0%) | 1179 (65.3%) |
| pplx | 306 (6.0%) | 761 (40.7%) | 174 (23.2%) | 196 (24.0%) | 9 (45.0%) | 1571 (87.0%) |
| gliner-multi-v21 | 968 (19.0%) | 597 (32.0%) | 347 (46.3%) | 480 (58.7%) | 13 (65.0%) | 689 (38.2%) |
| davlan-xlmr | 375 (7.4%) | 549 (29.4%) | 672 (89.6%) | 794 (97.1%) | 17 (85.0%) | 756 (41.9%) |
| mmbert32k | 1054 (20.7%) | 731 (39.1%) | 220 (29.3%) | 257 (31.4%) | 2 (10.0%) | 1003 (55.5%) |
| nuner-zero | 1067 (21.0%) | 941 (50.4%) | 227 (30.3%) | 333 (40.7%) | 2 (10.0%) | 720 (39.9%) |
| nym-small | 913 (18.0%) | 519 (27.8%) | 302 (40.3%) | 307 (37.5%) | 3 (15.0%) | 1271 (70.4%) |
| gliner-nvidia | 1211 (23.8%) | 576 (30.8%) | 269 (35.9%) | 403 (49.3%) | 10 (50.0%) | 888 (49.2%) |
| ner-ru-yqelz | 1746 (34.3%) | 445 (23.8%) | 546 (72.8%) | 630 (77.0%) | 3 (15.0%) | 222 (12.3%) |
| davlan-mbert | 849 (16.7%) | 605 (32.4%) | 654 (87.2%) | 777 (95.0%) | 15 (75.0%) | 762 (42.2%) |
| gliner25-fastino | 1168 (23.0%) | 891 (47.7%) | 209 (27.9%) | 322 (39.4%) | 2 (10.0%) | 1130 (62.6%) |
| gliner2-large | 1731 (34.1%) | 773 (41.4%) | 273 (36.4%) | 425 (52.0%) | 8 (40.0%) | 858 (47.5%) |
| opf-ru | 1423 (28.0%) | 1019 (54.6%) | 204 (27.2%) | 331 (40.5%) | 4 (20.0%) | 1200 (66.4%) |
| gliner-pii-base | 1692 (33.3%) | 896 (48.0%) | 349 (46.5%) | 422 (51.6%) | 13 (65.0%) | 934 (51.7%) |
| gliner2-hivetrace-uni | 1810 (35.6%) | 789 (42.2%) | 302 (40.3%) | 406 (49.6%) | 7 (35.0%) | 1115 (61.7%) |
| openmed-multilingual | 1548 (30.5%) | 856 (45.8%) | 338 (45.1%) | 349 (42.7%) | 14 (70.0%) | 1352 (74.9%) |
| gliner-pii-edge | 1710 (33.6%) | 1014 (54.3%) | 407 (54.3%) | 427 (52.2%) | 14 (70.0%) | 888 (49.2%) |
| ner-ru-gherman | 987 (19.4%) | 696 (37.3%) | 639 (85.2%) | 770 (94.1%) | 16 (80.0%) | 1418 (78.5%) |
| ru-legal-ner | 2625 (51.6%) | 708 (37.9%) | 177 (23.6%) | 285 (34.8%) | 4 (20.0%) | 774 (42.9%) |
| stanza-ru | 1610 (31.7%) | 945 (50.6%) | 657 (87.6%) | 518 (63.3%) | 17 (85.0%) | 912 (50.5%) |
| gliner2-vladlinv | 966 (19.0%) | 1447 (77.5%) | 307 (40.9%) | 389 (47.6%) | 17 (85.0%) | 1768 (97.9%) |
| openmed-nemotron | 1783 (35.1%) | 1032 (55.2%) | 338 (45.1%) | 589 (72.0%) | 15 (75.0%) | 1281 (70.9%) |
| gliner-stream-pii | 2611 (51.4%) | 785 (42.0%) | 512 (68.3%) | 404 (49.4%) | 17 (85.0%) | 1131 (62.6%) |
| gravitee-small | 2102 (41.4%) | 1112 (59.5%) | 404 (53.9%) | 654 (80.0%) | 15 (75.0%) | 1298 (71.9%) |
| traciora | 2422 (47.6%) | 1066 (57.1%) | 282 (37.6%) | 486 (59.4%) | 14 (70.0%) | 1338 (74.1%) |
| kalyan-ettin | 2250 (44.3%) | 1050 (56.2%) | 375 (50.0%) | 532 (65.0%) | 17 (85.0%) | 1493 (82.7%) |
| ru-pii-ner | 2402 (47.3%) | 1381 (73.9%) | 404 (53.9%) | 374 (45.7%) | 19 (95.0%) | 1674 (92.7%) |
| apararti | 3104 (61.1%) | 1202 (64.3%) | 283 (37.7%) | 264 (32.3%) | 15 (75.0%) | 1522 (84.3%) |
| openai-base | 3012 (59.3%) | 1307 (70.0%) | 295 (39.3%) | 364 (44.5%) | 16 (80.0%) | 1640 (90.8%) |
| opf-kz-ru | 3715 (73.1%) | 1311 (70.2%) | 298 (39.7%) | 283 (34.6%) | 15 (75.0%) | 1666 (92.2%) |
| opf-ru-v2 | 3557 (70.0%) | 1428 (76.4%) | 363 (48.4%) | 403 (49.3%) | 17 (85.0%) | 1583 (87.7%) |
| spacy-ru-lg | 3600 (70.8%) | 1613 (86.3%) | 685 (91.3%) | 764 (93.4%) | 19 (95.0%) | 1589 (88.0%) |
| natasha | 4274 (84.1%) | 1823 (97.6%) | 746 (99.5%) | 802 (98.0%) | 20 (100.0%) | 1751 (97.0%) |
| rules-ru | 5082 (100.0%) | 1868 (100.0%) | 479 (63.9%) | 813 (99.4%) | 6 (30.0%) | 1806 (100.0%) |
| fef2-secret-ru | 5008 (98.5%) | 1850 (99.0%) | 748 (99.7%) | 814 (99.5%) | 20 (100.0%) | 1803 (99.8%) |
| spacy-alrosait | 5079 (99.9%) | 1864 (99.8%) | 748 (99.7%) | 815 (99.6%) | 20 (100.0%) | 1805 (99.9%) |

## Char recall by gold type

| type | group | pii-shield-onnx | gliner2-fastino | bardsai-eu | gliner-urchade | gliner2-hivetrace-omni | nym-base | pplx | gliner-multi-v21 | davlan-xlmr | mmbert32k | nuner-zero | nym-small | gliner-nvidia | ner-ru-yqelz | davlan-mbert | gliner25-fastino | gliner2-large | opf-ru | gliner-pii-base | gliner2-hivetrace-uni | openmed-multilingual | gliner-pii-edge | ner-ru-gherman | ru-legal-ner | stanza-ru | gliner2-vladlinv | openmed-nemotron | gliner-stream-pii | gravitee-small | traciora | kalyan-ettin | ru-pii-ner | apararti | openai-base | opf-kz-ru | opf-ru-v2 | spacy-ru-lg | natasha | rules-ru | fef2-secret-ru | spacy-alrosait |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| CODE | ID | 0.854 | 0.447 | 0.441 | 0.300 | 0.515 | 0.371 | 0.639 | 0.269 | 0.018 | 0.428 | 0.416 | 0.403 | 0.356 | 0.234 | 0.029 | 0.447 | 0.318 | 0.351 | 0.339 | 0.342 | 0.358 | 0.342 | 0.023 | 0.490 | 0.207 | 0.379 | 0.176 | 0.238 | 0.124 | 0.296 | 0.204 | 0.414 | 0.541 | 0.450 | 0.509 | 0.352 | 0.039 | 0.011 | 0.004 | 0.003 | 0.004 |
| CODE_PHONE | CONTACT | 0.898 | 0.590 | 0.621 | 0.545 | 0.584 | 0.591 | 0.660 | 0.260 | 0.003 | 0.590 | 0.577 | 0.579 | 0.563 | 0.108 | 0.000 | 0.603 | 0.552 | 0.520 | 0.455 | 0.536 | 0.512 | 0.561 | 0.000 | 0.594 | 0.043 | 0.507 | 0.361 | 0.301 | 0.410 | 0.495 | 0.267 | 0.487 | 0.679 | 0.618 | 0.600 | 0.504 | 0.039 | 0.000 | 0.014 | 0.000 | 0.000 |
| CODE_URL | NET | 0.930 | 0.456 | 0.257 | 0.524 | 0.431 | 0.274 | 0.407 | 0.113 | 0.045 | 0.335 | 0.496 | 0.305 | 0.311 | 0.372 | 0.119 | 0.414 | 0.266 | 0.375 | 0.178 | 0.249 | 0.109 | 0.099 | 0.045 | 0.536 | 0.058 | 0.033 | 0.079 | 0.044 | 0.069 | 0.142 | 0.032 | 0.025 | 0.171 | 0.165 | 0.193 | 0.048 | 0.025 | 0.000 | 0.267 | 0.000 | 0.000 |
| LOC_CITY | ADDRESS | 0.828 | 0.846 | 0.894 | 0.796 | 0.790 | 0.830 | 0.624 | 0.774 | 0.893 | 0.536 | 0.446 | 0.777 | 0.755 | 0.933 | 0.870 | 0.571 | 0.614 | 0.387 | 0.623 | 0.687 | 0.607 | 0.569 | 0.758 | 0.616 | 0.583 | 0.228 | 0.378 | 0.663 | 0.446 | 0.373 | 0.449 | 0.278 | 0.299 | 0.242 | 0.212 | 0.178 | 0.176 | 0.048 | 0.000 | 0.018 | 0.001 |
| LOC_COUNTRY | ADDRESS | 0.797 | 0.515 | 0.477 | 0.533 | 0.482 | 0.421 | 0.170 | 0.536 | 0.406 | 0.328 | 0.334 | 0.405 | 0.553 | 0.720 | 0.413 | 0.401 | 0.474 | 0.238 | 0.352 | 0.301 | 0.140 | 0.303 | 0.440 | 0.605 | 0.492 | 0.006 | 0.344 | 0.292 | 0.205 | 0.194 | 0.274 | 0.071 | 0.076 | 0.046 | 0.041 | 0.057 | 0.105 | 0.010 | 0.000 | 0.000 | 0.000 |
| LOC_HOUSENUMBER | ADDRESS | 0.790 | 0.592 | 0.732 | 0.625 | 0.706 | 0.786 | 0.963 | 0.478 | 0.266 | 0.809 | 0.534 | 0.718 | 0.615 | 0.040 | 0.124 | 0.378 | 0.534 | 0.473 | 0.338 | 0.469 | 0.508 | 0.219 | 0.427 | 0.583 | 0.203 | 0.466 | 0.420 | 0.420 | 0.417 | 0.711 | 0.497 | 0.462 | 0.648 | 0.559 | 0.590 | 0.455 | 0.051 | 0.000 | 0.000 | 0.000 | 0.000 |
| LOC_OTHER | ADDRESS | 0.863 | 0.776 | 0.551 | 0.756 | 0.674 | 0.418 | 0.168 | 0.536 | 0.477 | 0.269 | 0.179 | 0.324 | 0.338 | 0.813 | 0.475 | 0.099 | 0.476 | 0.316 | 0.355 | 0.196 | 0.210 | 0.359 | 0.131 | 0.614 | 0.398 | 0.031 | 0.176 | 0.402 | 0.210 | 0.252 | 0.139 | 0.109 | 0.133 | 0.108 | 0.107 | 0.091 | 0.120 | 0.000 | 0.000 | 0.000 | 0.000 |
| LOC_STREET | ADDRESS | 0.795 | 0.854 | 0.850 | 0.827 | 0.823 | 0.823 | 0.752 | 0.777 | 0.826 | 0.645 | 0.650 | 0.750 | 0.699 | 0.668 | 0.818 | 0.712 | 0.607 | 0.567 | 0.412 | 0.623 | 0.545 | 0.350 | 0.398 | 0.630 | 0.540 | 0.325 | 0.496 | 0.564 | 0.441 | 0.634 | 0.448 | 0.344 | 0.545 | 0.497 | 0.459 | 0.404 | 0.141 | 0.007 | 0.000 | 0.003 | 0.003 |
| LOC_ZIP | ADDRESS | 0.811 | 0.672 | 0.535 | 0.606 | 0.662 | 0.644 | 0.810 | 0.213 | 0.063 | 0.664 | 0.597 | 0.612 | 0.545 | 0.176 | 0.015 | 0.208 | 0.432 | 0.535 | 0.465 | 0.251 | 0.530 | 0.427 | 0.114 | 0.468 | 0.110 | 0.282 | 0.446 | 0.382 | 0.396 | 0.455 | 0.476 | 0.342 | 0.560 | 0.503 | 0.618 | 0.376 | 0.007 | 0.000 | 0.000 | 0.000 | 0.007 |
| ORG | ORG | 0.822 | 0.735 | 0.545 | 0.695 | 0.635 | 0.295 | 0.096 | 0.621 | 0.563 | 0.280 | 0.607 | 0.230 | 0.497 | 0.813 | 0.536 | 0.401 | 0.521 | 0.256 | 0.442 | 0.271 | 0.193 | 0.443 | 0.112 | 0.556 | 0.460 | 0.012 | 0.231 | 0.322 | 0.280 | 0.237 | 0.135 | 0.068 | 0.146 | 0.094 | 0.066 | 0.123 | 0.111 | 0.021 | 0.000 | 0.001 | 0.000 |
| PERSON | PERSON | 0.793 | 0.821 | 0.868 | 0.830 | 0.740 | 0.805 | 0.907 | 0.806 | 0.851 | 0.761 | 0.794 | 0.765 | 0.667 | 0.672 | 0.790 | 0.773 | 0.676 | 0.700 | 0.682 | 0.637 | 0.657 | 0.669 | 0.746 | 0.496 | 0.691 | 0.755 | 0.624 | 0.481 | 0.611 | 0.545 | 0.541 | 0.547 | 0.432 | 0.458 | 0.317 | 0.366 | 0.270 | 0.140 | 0.000 | 0.021 | 0.001 |
| PERSON_EMAIL | CONTACT | 0.892 | 0.602 | 0.399 | 0.674 | 0.549 | 0.456 | 0.727 | 0.366 | 0.091 | 0.452 | 0.639 | 0.427 | 0.487 | 0.238 | 0.104 | 0.599 | 0.452 | 0.526 | 0.451 | 0.371 | 0.342 | 0.283 | 0.075 | 0.553 | 0.113 | 0.479 | 0.388 | 0.221 | 0.276 | 0.597 | 0.436 | 0.373 | 0.504 | 0.553 | 0.519 | 0.392 | 0.076 | 0.006 | 0.407 | 0.002 | 0.001 |

## char F1 by domain

| domain | n | pii-shield-onnx | gliner2-fastino | bardsai-eu | gliner-urchade | gliner2-hivetrace-omni | nym-base | pplx | gliner-multi-v21 | davlan-xlmr | mmbert32k | nuner-zero | nym-small | gliner-nvidia | ner-ru-yqelz | davlan-mbert | gliner25-fastino | gliner2-large | opf-ru | gliner-pii-base | gliner2-hivetrace-uni | openmed-multilingual | gliner-pii-edge | ner-ru-gherman | ru-legal-ner | stanza-ru | gliner2-vladlinv | openmed-nemotron | gliner-stream-pii | gravitee-small | traciora | kalyan-ettin | ru-pii-ner | apararti | openai-base | opf-kz-ru | opf-ru-v2 | spacy-ru-lg | natasha | rules-ru | fef2-secret-ru | spacy-alrosait |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| aidashboard | 60 | 0.068 | 0.510 | 0.797 | 0.464 | 0.606 | 0.742 | 0.693 | 0.502 | 0.770 | 0.471 | 0.445 | 0.695 | 0.463 | 0.715 | 0.726 | 0.639 | 0.479 | 0.318 | 0.585 | 0.511 | 0.547 | 0.575 | 0.676 | 0.071 | 0.132 | 0.618 | 0.249 | 0.563 | 0.232 | 0.360 | 0.457 | 0.381 | 0.289 | 0.269 | 0.172 | 0.270 | 0.081 | 0.070 | 0.000 | 0.010 | 0.000 |
| anamnesis | 60 | 0.104 | 0.490 | 0.684 | 0.420 | 0.565 | 0.736 | 0.654 | 0.414 | 0.761 | 0.467 | 0.408 | 0.709 | 0.412 | 0.541 | 0.710 | 0.527 | 0.422 | 0.333 | 0.546 | 0.508 | 0.562 | 0.527 | 0.626 | 0.101 | 0.137 | 0.597 | 0.328 | 0.574 | 0.263 | 0.503 | 0.445 | 0.396 | 0.425 | 0.367 | 0.387 | 0.511 | 0.113 | 0.065 | 0.004 | 0.030 | 0.002 |
| billing | 60 | 0.129 | 0.594 | 0.734 | 0.501 | 0.644 | 0.621 | 0.575 | 0.506 | 0.715 | 0.472 | 0.488 | 0.597 | 0.512 | 0.697 | 0.655 | 0.603 | 0.471 | 0.382 | 0.551 | 0.479 | 0.520 | 0.555 | 0.470 | 0.121 | 0.200 | 0.492 | 0.335 | 0.529 | 0.266 | 0.487 | 0.428 | 0.359 | 0.420 | 0.387 | 0.352 | 0.431 | 0.125 | 0.093 | 0.114 | 0.020 | 0.002 |
| carinsurance | 100 | 0.245 | 0.573 | 0.632 | 0.503 | 0.639 | 0.615 | 0.739 | 0.449 | 0.442 | 0.535 | 0.515 | 0.602 | 0.515 | 0.541 | 0.413 | 0.590 | 0.459 | 0.434 | 0.521 | 0.522 | 0.507 | 0.478 | 0.317 | 0.200 | 0.196 | 0.566 | 0.353 | 0.394 | 0.263 | 0.526 | 0.482 | 0.436 | 0.518 | 0.469 | 0.501 | 0.442 | 0.150 | 0.099 | 0.220 | 0.008 | 0.003 |
| grouptherapy | 30 | 0.037 | 0.443 | 0.790 | 0.380 | 0.611 | 0.884 | 0.818 | 0.430 | 0.916 | 0.396 | 0.348 | 0.853 | 0.447 | 0.706 | 0.801 | 0.643 | 0.364 | 0.255 | 0.504 | 0.584 | 0.585 | 0.635 | 0.816 | 0.029 | 0.119 | 0.750 | 0.201 | 0.618 | 0.150 | 0.256 | 0.530 | 0.490 | 0.197 | 0.170 | 0.144 | 0.131 | 0.088 | 0.181 | 0.000 | 0.015 | 0.003 |
| police | 70 | 0.120 | 0.478 | 0.746 | 0.405 | 0.504 | 0.719 | 0.636 | 0.428 | 0.702 | 0.465 | 0.369 | 0.680 | 0.467 | 0.694 | 0.665 | 0.471 | 0.345 | 0.312 | 0.478 | 0.480 | 0.460 | 0.550 | 0.557 | 0.121 | 0.157 | 0.447 | 0.278 | 0.535 | 0.264 | 0.417 | 0.452 | 0.341 | 0.380 | 0.314 | 0.344 | 0.416 | 0.125 | 0.097 | 0.035 | 0.044 | 0.004 |
| therapy | 60 | 0.098 | 0.604 | 0.810 | 0.523 | 0.695 | 0.802 | 0.673 | 0.514 | 0.826 | 0.491 | 0.421 | 0.748 | 0.507 | 0.823 | 0.789 | 0.538 | 0.503 | 0.356 | 0.557 | 0.523 | 0.548 | 0.610 | 0.650 | 0.099 | 0.173 | 0.512 | 0.289 | 0.617 | 0.245 | 0.407 | 0.465 | 0.366 | 0.302 | 0.238 | 0.182 | 0.250 | 0.144 | 0.102 | 0.000 | 0.009 | 0.002 |
| triage | 60 | 0.129 | 0.467 | 0.796 | 0.334 | 0.503 | 0.810 | 0.758 | 0.356 | 0.752 | 0.518 | 0.376 | 0.779 | 0.416 | 0.608 | 0.735 | 0.442 | 0.343 | 0.316 | 0.524 | 0.566 | 0.589 | 0.467 | 0.592 | 0.140 | 0.153 | 0.574 | 0.339 | 0.554 | 0.285 | 0.520 | 0.431 | 0.369 | 0.445 | 0.344 | 0.438 | 0.484 | 0.126 | 0.113 | 0.004 | 0.002 | 0.001 |

## Missed at thresholds 0.5 / 0.3 / 0.2 / 0.1

Only models that return a score. A lower threshold keeps more spans: fewer misses, more masking.

| model | 0.5 | 0.3 | 0.2 | 0.1 |
|---|---|---|---|---|
| pii-shield-onnx | 1837 (17.8%) | 1423 (13.8%) | 1410 (13.6%) | 1410 (13.6%) |
| gliner2-fastino | 2286 (22.1%) | 1833 (17.7%) | 1558 (15.1%) | 1208 (11.7%) |
| bardsai-eu | 2373 (22.9%) | 2270 (21.9%) | 2265 (21.9%) | 2264 (21.9%) |
| gliner-urchade | 2579 (24.9%) | 2026 (19.6%) | 1735 (16.8%) | 1393 (13.5%) |
| gliner2-hivetrace-omni | 2694 (26.0%) | 1982 (19.2%) | 1647 (15.9%) | 1278 (12.4%) |
| nym-base | 2859 (27.6%) | 2798 (27.0%) | 2796 (27.0%) | 2796 (27.0%) |
| gliner-multi-v21 | 3094 (29.9%) | 2131 (20.6%) | 1731 (16.7%) | 1293 (12.5%) |
| davlan-xlmr | 3163 (30.6%) | 3139 (30.3%) | 3138 (30.3%) | 3138 (30.3%) |
| mmbert32k | 3267 (31.6%) | 2009 (19.4%) | 1794 (17.3%) | 1777 (17.2%) |
| nuner-zero | 3290 (31.8%) | 2633 (25.5%) | 2236 (21.6%) | 1747 (16.9%) |
| nym-small | 3315 (32.0%) | 3255 (31.5%) | 3255 (31.5%) | 3255 (31.5%) |
| gliner-nvidia | 3357 (32.5%) | 2810 (27.2%) | 2530 (24.5%) | 2177 (21.0%) |
| ner-ru-yqelz | 3592 (34.7%) | 3198 (30.9%) | 3184 (30.8%) | 3184 (30.8%) |
| davlan-mbert | 3662 (35.4%) | 3629 (35.1%) | 3629 (35.1%) | 3629 (35.1%) |
| gliner25-fastino | 3722 (36.0%) | 3330 (32.2%) | 3116 (30.1%) | 2772 (26.8%) |
| gliner2-large | 4068 (39.3%) | 3069 (29.7%) | 2547 (24.6%) | 1985 (19.2%) |
| opf-ru | 4181 (40.4%) | 4051 (39.2%) | 4049 (39.1%) | 4049 (39.1%) |
| gliner-pii-base | 4306 (41.6%) | 2113 (20.4%) | 1421 (13.7%) | 613 (5.9%) |
| gliner2-hivetrace-uni | 4429 (42.8%) | 3052 (29.5%) | 2458 (23.8%) | 1808 (17.5%) |
| openmed-multilingual | 4457 (43.1%) | 4128 (39.9%) | 4116 (39.8%) | 4116 (39.8%) |
| gliner-pii-edge | 4460 (43.1%) | 2262 (21.9%) | 1354 (13.1%) | 207 (2.0%) |
| ner-ru-gherman | 4526 (43.8%) | 4323 (41.8%) | 4267 (41.2%) | 4254 (41.1%) |
| ru-legal-ner | 4573 (44.2%) | 3209 (31.0%) | 3107 (30.0%) | 3105 (30.0%) |
| gliner2-vladlinv | 4894 (47.3%) | 4760 (46.0%) | 4701 (45.4%) | 4610 (44.6%) |
| openmed-nemotron | 5038 (48.7%) | 4228 (40.9%) | 4161 (40.2%) | 4157 (40.2%) |
| gliner-stream-pii | 5460 (52.8%) | 4133 (40.0%) | 3469 (33.5%) | 2649 (25.6%) |
| gravitee-small | 5585 (54.0%) | 5230 (50.6%) | 5229 (50.5%) | 5229 (50.5%) |
| traciora | 5608 (54.2%) | 5559 (53.7%) | 5557 (53.7%) | 5557 (53.7%) |
| kalyan-ettin | 5717 (55.3%) | 5126 (49.6%) | 5073 (49.0%) | 5069 (49.0%) |
| apararti | 6390 (61.8%) | 6285 (60.8%) | 6283 (60.7%) | 6283 (60.7%) |
| openai-base | 6634 (64.1%) | 6614 (63.9%) | 6614 (63.9%) | 6614 (63.9%) |
| opf-kz-ru | 7288 (70.4%) | 7252 (70.1%) | 7252 (70.1%) | 7252 (70.1%) |
| opf-ru-v2 | 7351 (71.1%) | 7340 (71.0%) | 7338 (70.9%) | 7338 (70.9%) |
| fef2-secret-ru | 10243 (99.0%) | 10238 (99.0%) | 10238 (99.0%) | 10238 (99.0%) |

Notes: the unit of counting is one gold span after boundary normalization - the same unit in `missed`, in the group table, in the threshold table and in the ensembles; two annotations that share boundaries but not type are two spans, so an empty answer misses 100%. Gold and predicted boundaries are normalized the same way (whitespace, word boundaries, merge of adjacent same-label pieces). `dropped spans` - predicted intervals outside the text: rejected, not clipped to fit. A run made before `run.py` started clipping span ends to the piece it fed the model can carry them; `meta.clipped` counts what the clipping fixes in newer runs. Zero-shot models get the taxonomy of the set as labels (the `labels` field of `meta.json`). `train` - the model was trained on the source of this set (its slices and corrupted copies included): its row is not comparable with the others and stays out of the pooled numbers; an empty `contaminated` list in `benchmark/models.toml` means no evidence of overlap was found, not proof of none. `hidden` - gold spans every character of which is covered; a touched span counts as detected, not as hidden. missed % and char F1 carry 95% bootstrap intervals over rows (1000 resamples, the same rows for every model of the set). The group of a gold type comes from `BENCH/<set>/meta.json`; a model label with no group falls into OTHER. ms/row is a median and includes chunking of long texts; hardware and batch are in the meta line of the prediction file.