Open source PII detection benchmark

Start with what you need to protect

Compare personal data, secrets, and the datasets where detectors struggle.

Performance from easy to extreme

Dataset average · see how much sensitive data each detector hides completely at every difficulty level. Higher is better.

Detector	Easy	Medium	Hard	Extreme
Fastino	86.7%	79.6%	75.1%	36.3%
PPLX	94.7%	81.6%	39.9%	26.0%
NuNER Zero	75.0%	72.6%	61.4%	26.4%
bardsai-eu	79.9%	62.5%	70.6%	22.2%
gliner2-large	69.6%	66.3%	68.1%	27.7%
gliner-urchade	69.4%	66.3%	56.8%	33.8%
gliner25-fastino	68.7%	62.7%	58.6%	31.9%
apararti	79.5%	65.2%	37.7%	10.9%
openai-base	80.0%	62.3%	29.4%	17.1%
gliner-pii-edge	59.9%	61.1%	70.0%	23.9%
opf-kz-ru	72.7%	58.8%	30.9%	9.4%
pii-shield-onnx	54.4%	56.9%	35.6%	13.0%
ru-pii-ner	77.2%	58.8%	39.5%	19.9%
gliner-multi-v21	59.9%	50.0%	61.4%	27.3%
gliner-stream-pii	62.3%	54.4%	47.3%	19.6%
mmbert32k	65.3%	54.7%	35.9%	18.7%
traciora	69.9%	52.8%	32.9%	15.6%
opf-ru-v2	68.1%	50.8%	28.9%	10.6%
gliner2-vladlinv	63.0%	51.9%	31.3%	21.7%
gliner-pii-base	46.5%	46.8%	50.3%	18.9%
davlan-xlmr	49.5%	40.7%	60.9%	18.7%
ru-legal-ner	53.7%	44.7%	39.1%	19.3%
stanza-ru	50.8%	39.4%	71.0%	11.3%
gravitee-small	66.8%	52.7%	32.5%	18.2%
davlan-mbert	46.8%	37.7%	60.1%	16.4%
ner-ru-yqelz	45.6%	38.1%	55.6%	41.4%
fef2-secret-ru	34.4%	30.0%	47.3%	17.5%
ner-ru-gherman	34.1%	23.8%	37.4%	5.6%
spacy-ru-lg	28.1%	20.2%	47.0%	25.6%
natasha	24.8%	18.3%	42.5%	8.8%
rules-ru	12.8%	24.8%	3.4%	1.7%
pplx+cpu-int8Not all datasets	98.8%	94.3%	47.9%	—
pplx+homoglyphNot all datasets	—	93.8%	78.7%	—
gliner2-fastino-ruNot all datasets	78.6%	80.0%	78.6%	35.8%
HiveTrace OmniNot all datasets	84.5%	74.6%	64.5%	36.3%
gliner-nvidia+sent300Not all datasets	65.9%	72.7%	73.2%	—
gliner25-fastino+sent300Not all datasets	66.8%	72.5%	72.1%	—
nym-base+homoglyphNot all datasets	—	71.3%	72.0%	—
gliner2-hivetrace-omni-ruNot all datasets	79.0%	73.3%	62.8%	35.9%
openai-base-onnxNot all datasets	76.7%	71.7%	22.3%	—
gliner25-fastino+ov100Not all datasets	69.2%	71.9%	71.8%	—
nym-base+sent300Not all datasets	61.4%	66.3%	62.2%	—
pplx+sent300Not all datasets	84.3%	70.3%	27.9%	—
gliner25-fastino-ruNot all datasets	75.4%	65.4%	62.5%	31.4%
nym-baseNot all datasets	73.9%	62.7%	67.4%	16.5%
gliner-nvidiaNot all datasets	82.8%	68.9%	63.6%	25.9%
nym-base+ov100Not all datasets	60.8%	63.6%	67.1%	—
ru-legal-ner+cpuNot all datasets	—	62.9%	—	—
nym-smallNot all datasets	73.8%	61.0%	63.8%	11.8%
gliner-nvidia+homoglyphNot all datasets	—	68.7%	60.1%	—
pplx+ov100Not all datasets	87.2%	68.2%	23.3%	—
ru-legal-ner+homoglyphNot all datasets	—	68.6%	57.3%	—
nym-base+cpu-int8Not all datasets	94.4%	63.4%	41.5%	—
gliner-nvidia+ov100Not all datasets	63.5%	60.0%	65.4%	—
openmed-multilingualNot all datasets	67.7%	53.7%	30.7%	6.1%
gliner-nvidia-ruNot all datasets	67.5%	60.9%	58.9%	21.4%
gliner2-hivetrace-uniNot all datasets	65.2%	53.3%	28.2%	23.2%
opf-ru-v2+homoglyphNot all datasets	—	68.7%	44.7%	—
opf-ruNot all datasets	62.1%	54.9%	32.7%	11.7%
openmed-nemotronNot all datasets	61.4%	48.1%	29.4%	8.3%
presidio-ruNot all datasets	—	54.5%	—	—
gliner25-fastino+nochunkNot all datasets	64.8%	58.1%	49.9%	—
gliner2-vladlinv+homoglyphNot all datasets	—	60.8%	42.4%	—
gliner25-fastino-ru+nochunkNot all datasets	59.3%	53.8%	50.7%	—
gliner-urchade-ruNot all datasets	50.7%	51.0%	45.7%	20.3%
gravitee-small+cpu-int8Not all datasets	71.1%	51.9%	40.2%	—
ru-legal-ner+sent300Not all datasets	61.4%	40.6%	32.5%	—
gliner-multi-v21-ruNot all datasets	51.8%	42.6%	50.3%	25.0%
kalyan-ettinNot all datasets	60.2%	45.0%	27.8%	8.8%
gliner2-vladlinv-ruNot all datasets	72.6%	53.3%	39.7%	21.7%
opf-ru-v2+sent300Not all datasets	64.4%	41.8%	22.7%	—
ru-legal-ner+ov100Not all datasets	59.5%	34.4%	29.7%	—
opf-ru-v2+ov100Not all datasets	64.1%	39.4%	22.0%	—
ru-legal-ner+cpu-int8Not all datasets	43.4%	44.4%	6.1%	—
fef2-secret-ru+cpu-int8Not all datasets	3.6%	31.9%	2.1%	—
davlan-mbert+cpu-int8Not all datasets	55.3%	25.0%	54.5%	—
ner-ru-gherman+homoglyphNot all datasets	—	1.4%	31.2%	—
mmbert32k+cpu-int8Not all datasets	50.9%	29.1%	5.9%	—
gliner2-hivetrace-uni-ruNot all datasets	26.3%	22.5%	31.3%	16.1%
kalyan-ettin+cpu-int8Not all datasets	41.6%	21.0%	7.8%	—
mmbert32k+nochunkNot all datasets	24.9%	31.4%	16.6%	—
ner-ru-gherman-onnxNot all datasets	55.0%	18.0%	14.4%	—
davlan-xlmr+cpu-int8Not all datasets	51.4%	18.7%	36.2%	—
ner-ru-gherman+cpu-int8Not all datasets	52.1%	16.9%	13.8%	—
spacy-alrosaitNot all datasets	15.7%	11.8%	20.3%	7.0%
gliner-pii-edge+cpu-int8Not all datasets	17.5%	15.8%	3.3%	—
detect-secretsNot all datasets	—	1.6%	—	—
kingfisherNot all datasets	—	1.0%	—	—
noseyparkerNot all datasets	—	1.0%	—	—
titusNot all datasets	—	1.0%	—	—
credsweeper-nomlNot all datasets	—	0.7%	—	—
gliner-pii-base+cpu-int8Not all datasets	<0.1%	1.2%	2.3%	—
credsweeperNot all datasets	—	0.5%	—	—
deepsecretsNot all datasets	—	0.1%	—	—
trufflehogNot all datasets	—	0.1%	—	—
betterleaksNot all datasets	0.0%	0.2%	—	—
gitleaksNot all datasets	0.0%	0.2%	—	—
gliner-multi-v21-ru+cpu-int8Not all datasets	—	0.0%	—	—
gliner-multi-v21+cpu-int8Not all datasets	0.0%	0.0%	0.0%	—
gliner-nvidia-ru+cpu-int8Not all datasets	—	0.0%	—	—
gliner-nvidia+cpu-int8Not all datasets	0.0%	0.0%	0.0%	—
gliner-urchade-ru+cpu-int8Not all datasets	—	0.0%	—	—
gliner-urchade+cpu-int8Not all datasets	0.0%	0.0%	0.0%	—

PII datasets vs secrets datasets

Pooled annotations · compare how well each detector hides labeled items in datasets assigned to each task.

DetectorPII datasetsSecrets datasets

Fastino

80.1%

—

PPLX

74.6%

—

NuNER Zero

71.9%

—

bardsai-eu

66.5%

—

gliner2-large

66.4%

—

gliner-urchade

65.9%

—

gliner25-fastino

62.7%

—

apararti

62.0%

—

openai-base

59.2%

—

gliner-pii-edge

56.8%

—

opf-kz-ru

55.9%

—

pii-shield-onnx

55.7%

—

ru-pii-ner

53.7%

—

gliner-multi-v21

51.9%

—

gliner-stream-pii

51.6%

—

mmbert32k

50.3%

—

traciora

48.0%

—

opf-ru-v2

46.9%

—

gliner2-vladlinv

46.3%

—

gliner-pii-base

45.0%

—

davlan-xlmr

44.8%

—

ru-legal-ner

44.0%

—

stanza-ru

43.9%

—

gravitee-small

43.1%

—

davlan-mbert

42.8%

—

ner-ru-yqelz

40.2%

—

fef2-secret-ru

35.3%

—

ner-ru-gherman

28.1%

—

spacy-ru-lg

27.2%

—

natasha

24.6%

—

rules-ru

23.8%

—

pplx+cpu-int8Not all datasets

93.2%

—

pplx+homoglyphNot all datasets

82.2%

—

gliner2-fastino-ruNot all datasets

77.8%

—

HiveTrace OmniNot all datasets

75.4%

—

gliner-nvidia+sent300Not all datasets

75.2%

—

gliner25-fastino+sent300Not all datasets

72.1%

—

nym-base+homoglyphNot all datasets

71.9%

—

gliner2-hivetrace-omni-ruNot all datasets

71.8%

—

openai-base-onnxNot all datasets

70.9%

—

gliner25-fastino+ov100Not all datasets

70.5%

—

nym-base+sent300Not all datasets

67.2%

—

pplx+sent300Not all datasets

65.2%

—

gliner25-fastino-ruNot all datasets

64.6%

—

nym-baseNot all datasets

64.3%

—

gliner-nvidiaNot all datasets

64.0%

—

nym-base+ov100Not all datasets

63.1%

—

ru-legal-ner+cpuNot all datasets

63.0%

—

nym-smallNot all datasets

62.7%

—

gliner-nvidia+homoglyphNot all datasets

62.1%

—

pplx+ov100Not all datasets

61.5%

—

ru-legal-ner+homoglyphNot all datasets

60.0%

—

nym-base+cpu-int8Not all datasets

59.7%

—

gliner-nvidia+ov100Not all datasets

56.7%

—

openmed-multilingualNot all datasets

51.3%

—

gliner-nvidia-ruNot all datasets

51.0%

—

gliner2-hivetrace-uniNot all datasets

50.7%

—

opf-ru-v2+homoglyphNot all datasets

50.3%

—

opf-ruNot all datasets

48.9%

—

openmed-nemotronNot all datasets

48.3%

—

presidio-ruNot all datasets

47.6%

—

gliner25-fastino+nochunkNot all datasets

46.7%

—

gliner2-vladlinv+homoglyphNot all datasets

46.6%

—

gliner25-fastino-ru+nochunkNot all datasets

45.5%

—

gliner-urchade-ruNot all datasets

45.5%

—

gravitee-small+cpu-int8Not all datasets

43.7%

—

ru-legal-ner+sent300Not all datasets

43.4%

—

gliner-multi-v21-ruNot all datasets

43.1%

—

kalyan-ettinNot all datasets

43.1%

—

gliner2-vladlinv-ruNot all datasets

42.5%

—

opf-ru-v2+sent300Not all datasets

41.4%

—

ru-legal-ner+ov100Not all datasets

38.2%

—

opf-ru-v2+ov100Not all datasets

37.4%

—

ru-legal-ner+cpu-int8Not all datasets

36.0%

—

fef2-secret-ru+cpu-int8Not all datasets

27.1%

—

davlan-mbert+cpu-int8Not all datasets

24.7%

—

ner-ru-gherman+homoglyphNot all datasets

24.3%

—

mmbert32k+cpu-int8Not all datasets

21.8%

—

gliner2-hivetrace-uni-ruNot all datasets

21.6%

—

kalyan-ettin+cpu-int8Not all datasets

20.7%

—

mmbert32k+nochunkNot all datasets

20.7%

—

ner-ru-gherman-onnxNot all datasets

19.2%

—

davlan-xlmr+cpu-int8Not all datasets

18.7%

—

ner-ru-gherman+cpu-int8Not all datasets

17.6%

—

spacy-alrosaitNot all datasets

14.6%

—

gliner-pii-edge+cpu-int8Not all datasets

9.6%

—

detect-secretsNot all datasets

1.6%

—

kingfisherNot all datasets

1.0%

—

noseyparkerNot all datasets

1.0%

—

titusNot all datasets

1.0%

—

credsweeper-nomlNot all datasets

0.7%

—

gliner-pii-base+cpu-int8Not all datasets

0.6%

—

credsweeperNot all datasets

0.5%

—

deepsecretsNot all datasets

0.1%

—

trufflehogNot all datasets

0.1%

—

betterleaksNot all datasets

0.1%

—

gitleaksNot all datasets

0.1%

—

gliner-multi-v21-ru+cpu-int8Not all datasets

0.0%

—

gliner-multi-v21+cpu-int8Not all datasets

0.0%

—

gliner-nvidia-ru+cpu-int8Not all datasets

0.0%

—

gliner-nvidia+cpu-int8Not all datasets

0.0%

—

gliner-urchade-ru+cpu-int8Not all datasets

0.0%

—

gliner-urchade+cpu-int8Not all datasets

0.0%

—

Use one detector or combine several?

See what combining detectors adds and how much extra text it masks.

One detector vs several combined

See whether combining detectors hides more sensitive data and how much extra text it masks.

See combinations

Detector setup	Completely hidden, % ↑ 0%50%100%	Extra text masked, % ↓	Not fully hidden, items ↓
Four detectorsFastino + PPLX + BardsAI EU + mmBERT	94.39%	7.93%	12,533
CPU + GPUFastino + PPLX + BardsAI EU	94.06%	6.63%	13,268
Three detectorsFastino + PPLX + mmBERT	92.94%	7.18%	15,777
Two detectorsFastino + PPLX	91.90%	5.72%	18,094
Single detectorGLiNER2 Fastino	80.11%	3.03%	44,412
Single detectorPPLX	74.62%	3.56%	56,689

Single detectorRuns on CUDACPU + GPU

Find the best detector for each data type

Focus on the sensitive information your application needs to protect.

Hardest datasets

Higher scores mean the dataset was harder for the tested detectors.

Best detector by data type

Pooled annotations by annotation type · choose what sensitive data you need to hide.

Passwords, keys & tokens96.8%PPLX2,742 labeled items

Logins & usernames98.9%PPLX11,382 labeled items

Bank accounts & cards76.2%PPLX5,253 labeled items

Documents & identifiers90.7%PPLX15,034 labeled items

People's names91.2%davlan-xlmr65,736 labeled items

Phone numbers & email94.9%PPLX21,981 labeled items

Addresses & locations78.2%Fastino39,602 labeled items

Dates & times96.3%PPLX5,502 labeled items

Organizations78.6%Fastino19,274 labeled items

Network identifiers96.2%NuNER Zero25,631 labeled items

Customer & employee IDs86.9%PPLX7,994 labeled items

Other sensitive attributes53.3%PPLX3,207 labeled items

Rank		Family
01	gliner2-fastino	GLINER2	80.11%	87.02%	3.03%	0.797	14.93	0.27
02	pplx	PPLX	74.62%	78.72%	3.56%	0.771	9.94	2.20
03	nuner-zero	GLINER	71.88%	78.97%	2.89%	0.788	20.12	0.38
04	bardsai-eu	ONNX	66.54%	81.26%	2.51%	0.740	—	—
05	gliner2-large	GLINER2	66.40%	72.77%	3.56%	0.723	24.69	0.61
06	gliner-urchade	GLINER	65.91%	72.94%	3.17%	0.717	9.68	0.20
07	gliner25-fastino	GLINER2	62.74%	69.12%	2.72%	0.747	12.47	0.15
08	apararti	OPF	62.00%	69.73%	4.13%	0.696	4.09	1.69
09	openai-base	OPF	59.22%	64.88%	2.84%	0.687	4.21	1.80
10	gliner-pii-edge	GLINER	56.77%	67.94%	4.01%	0.660	6.88	5.80
11	opf-kz-ru	OPF	55.88%	64.56%	3.64%	0.670	3.39	1.59
12	pii-shield-onnx	ONNX	55.74%	69.11%	21.26%	0.466	—	—
13	ru-pii-ner	RUPII	53.74%	57.94%	2.84%	0.631	1.42	—
14	gliner-multi-v21	GLINER	51.87%	60.74%	2.82%	0.570	9.89	0.19
15	gliner-stream-pii	GLINER	51.58%	63.76%	2.48%	0.650	8.20	0.21
16	mmbert32k	HF	50.32%	78.44%	3.39%	0.654	2.32	0.18
17	traciora	OPF	48.00%	58.87%	3.12%	0.658	3.29	1.36
18	opf-ru-v2	OPF	46.86%	56.44%	1.97%	0.648	3.47	—
19	gliner2-vladlinv	GLINER2	46.27%	49.81%	0.57%	0.596	14.49	0.17
20	gliner-pii-base	GLINER	45.01%	50.96%	1.89%	0.576	5.24	0.16
21	davlan-xlmr	HF	44.82%	52.27%	0.99%	0.533	2.26	0.14
22	ru-legal-ner	HF	44.03%	65.04%	9.58%	0.462	0.20	0.13
23	stanza-ru	STANZA	43.87%	51.43%	10.72%	0.346	8.35	—
24	gravitee-small	HF	43.05%	50.39%	4.36%	0.518	0.66	—
25	davlan-mbert	HF	42.76%	49.78%	1.14%	0.513	2.51	0.15
26	ner-ru-yqelz	HF	40.16%	50.20%	7.89%	0.485	7.08	0.18
27	fef2-secret-ru	HF	35.31%	42.97%	1.11%	0.532	3.05	0.15
28	ner-ru-gherman	HF	28.13%	49.68%	0.44%	0.483	2.48	0.15
29	spacy-ru-lg	SPACY	27.17%	31.85%	3.88%	0.334	0.19	—
30	natasha	NATASHA	24.59%	28.54%	1.59%	0.336	0.08	—
31	rules-ru	RULES	23.79%	24.63%	0.69%	0.521	—	—

Open source PII detection benchmark

Start with what you need to protect

Performance from easy to extreme

PII datasets vs secrets datasets

Use one detector or combine several?

One detector vs several combined

Find the best detector for each data type

Hardest datasets

Best detector by data type

Detector leaderboard

What was tested