The purpose of the trained models is to remove personally identifiable information (PII) from text, especially in the context of AI assistants and LLMs.

The example texts have 54 PII classes (types of sensitive data), targeting 229 discussion subjects / use cases split across business, education, psychology and legal fields, and 5 interactions styles (e.g. casual conversation, formal document, emails etc...).

Take a look at the Github implementation for specific reasearch.

Intended uses & limitations

More information needed

Training and evaluation data

More information needed

Training hyperparameters

The following hyperparameters were used during training:

learning_rate: 6e-04
train_batch_size: 32
eval_batch_size: 32
seed: 412
optimizer: Adam with betas=(0.96,0.996) and epsilon=1e-08
lr_scheduler_type: cosine_with_restarts
lr_scheduler_warmup_ratio: 0.22
num_epochs: 7
mixed_precision_training: N/A

Class wise metrics

It achieves the following results on the evaluation set:

Loss: 0.0211
Overall Precision: 0.9722
Overall Recall: 0.9792
Overall F1: 0.9757
Overall Accuracy: 0.9915
Accountname F1: 0.9993
Accountnumber F1: 0.9986
Age F1: 0.9884
Amount F1: 0.9984
Bic F1: 0.9942
Bitcoinaddress F1: 0.9974
Buildingnumber F1: 0.9898
City F1: 1.0
Companyname F1: 1.0
County F1: 0.9976
Creditcardcvv F1: 0.9541
Creditcardissuer F1: 0.9970
Creditcardnumber F1: 0.9754
Currency F1: 0.8966
Currencycode F1: 0.9946
Currencyname F1: 0.7697
Currencysymbol F1: 0.9958
Date F1: 0.9778
Dob F1: 0.9546
Email F1: 1.0
Ethereumaddress F1: 1.0
Eyecolor F1: 0.9925
Firstname F1: 0.9947
Gender F1: 1.0
Height F1: 1.0
Iban F1: 0.9978
Ip F1: 0.5404
Ipv4 F1: 0.8455
Ipv6 F1: 0.8855
Jobarea F1: 0.9091
Jobtitle F1: 1.0
Jobtype F1: 0.9672
Lastname F1: 0.9855
Litecoinaddress F1: 0.9949
Mac F1: 0.9965
Maskednumber F1: 0.9836
Middlename F1: 0.7385
Nearbygpscoordinate F1: 1.0
Ordinaldirection F1: 1.0
Password F1: 1.0
Phoneimei F1: 0.9978
Phonenumber F1: 0.9975
Pin F1: 0.9820
Prefix F1: 0.9872
Secondaryaddress F1: 1.0
Sex F1: 0.9916
Ssn F1: 0.9960
State F1: 0.9967
Street F1: 0.9991
Time F1: 1.0
Url F1: 1.0
Useragent F1: 0.9981
Username F1: 1.0
Vehiclevin F1: 0.9950
Vehiclevrm F1: 0.9870
Zipcode F1: 0.9966

Training results

Training Loss	Epoch	Step	Validation Loss	Overall Precision	Overall Recall	Overall F1	Overall Accuracy	Accountname F1	Accountnumber F1	Age F1	Amount F1	Bic F1	Bitcoinaddress F1	Buildingnumber F1	City F1	Companyname F1	County F1	Creditcardcvv F1	Creditcardissuer F1	Creditcardnumber F1	Currency F1	Currencycode F1	Currencyname F1	Currencysymbol F1	Date F1	Dob F1	Email F1	Ethereumaddress F1	Eyecolor F1	Firstname F1	Gender F1	Height F1	Iban F1	Ip F1	Ipv4 F1	Ipv6 F1	Jobarea F1	Jobtitle F1	Jobtype F1	Lastname F1	Litecoinaddress F1	Mac F1	Maskednumber F1	Middlename F1	Nearbygpscoordinate F1	Ordinaldirection F1	Password F1	Phoneimei F1	Phonenumber F1	Pin F1	Prefix F1	Secondaryaddress F1	Sex F1	Ssn F1	State F1	Street F1	Time F1	Url F1	Useragent F1	Username F1	Vehiclevin F1	Vehiclevrm F1	Zipcode F1
0.3984	1.0	2393	0.5120	0.7268	0.7819	0.7533	0.8741	0.9265	0.9819	0.8237	0.5053	0.2315	0.8197	0.7840	0.4886	0.8657	0.6338	0.8775	0.8575	0.7152	0.4533	0.0959	0.0	0.6480	0.7621	0.1884	0.9840	1.0	0.6194	0.8740	0.6610	0.9642	0.9039	0.0	0.8500	0.0220	0.6325	0.7840	0.6899	0.7667	0.0	0.2966	0.0	0.3682	0.9986	0.9387	0.8558	0.9879	0.9687	0.7455	0.9252	0.9661	0.9110	0.9771	0.5282	0.7988	0.8453	0.9648	0.9804	0.9356	0.7741	0.6780	0.7915
0.2097	2.0	4786	0.1406	0.8392	0.8913	0.8645	0.9509	0.9760	0.9114	0.9227	0.7647	0.9190	0.9554	0.8975	0.8881	0.9535	0.8414	0.9114	0.9820	0.8503	0.7525	0.6171	0.0077	0.8787	0.3161	0.2847	0.9924	0.9918	0.9495	0.9076	0.9625	0.9890	0.9870	0.0	0.8484	0.8007	0.8651	0.9660	0.9164	0.8695	0.8756	0.9685	0.7768	0.6697	0.9956	0.9754	0.9652	0.9976	0.9849	0.7977	0.9373	0.9923	0.9815	0.9828	0.8093	0.9445	0.9735	0.9933	0.9651	0.9854	0.9843	0.975	0.8123
0.1271	3.0	7179	0.1049	0.9218	0.9312	0.9265	0.9618	0.9950	0.9880	0.9172	0.9309	0.9652	0.8222	0.9160	0.9364	0.9749	0.9556	0.9211	0.9856	0.8939	0.8237	0.76	0.0080	0.9360	0.8735	0.5567	0.9993	0.9973	0.9872	0.9547	0.9773	0.9574	0.9694	0.0	0.8510	0.8032	0.9404	0.9844	0.9522	0.9294	0.8584	1.0	0.8603	0.8908	1.0	0.9829	0.9513	1.0	0.9792	0.8579	0.9413	0.9968	0.9513	0.9929	0.9278	0.9484	0.9862	0.9940	0.8884	0.9943	0.9616	0.9648	0.9395
0.1345	4.0	9572	0.0941	0.9463	0.9580	0.9521	0.9659	0.9975	0.9979	0.9356	0.9597	0.9084	0.9569	0.9827	0.9734	0.9835	0.9780	0.9634	0.9904	0.9393	0.8542	0.8915	0.4069	0.9636	0.8873	0.6572	0.9993	1.0	0.9923	0.9796	0.9983	0.9917	0.9972	0.0	0.8515	0.8027	0.9689	0.9943	0.9685	0.9668	0.8162	0.9912	0.9110	0.9364	1.0	0.9848	0.9734	0.9976	0.9949	0.9739	0.9609	0.9968	0.9906	0.9899	0.9772	0.9875	0.9855	0.9978	1.0	0.9972	0.9867	0.9817	0.9780
0.1067	5.0	11965	0.0724	0.9556	0.9659	0.9607	0.9699	0.9967	0.9965	0.9705	0.9742	0.9892	0.9736	0.9891	0.9794	0.9951	0.9860	0.9897	0.9892	0.9517	0.8386	0.9770	0.4186	0.9822	0.8869	0.7016	1.0	1.0	0.9949	0.9859	0.9983	1.0	0.9954	0.0075	0.8569	0.8012	0.9819	0.9979	0.9856	0.9843	0.9383	1.0	0.9318	0.9461	1.0	0.9905	1.0	1.0	0.9978	0.9906	0.9646	0.9981	0.9924	0.9970	0.9862	0.9966	0.9951	0.9970	1.0	0.9981	0.9933	1.0	0.9913
0.0808	6.0	14358	0.0693	0.9664	0.9732	0.9698	0.9728	1.0	1.0	0.9760	0.9897	0.9978	0.9907	0.9906	0.9930	0.9994	0.9939	1.0	0.9891	0.9590	0.9052	0.9875	0.7022	0.9892	0.9126	0.7438	1.0	1.0	1.0	0.9934	0.9991	1.0	1.0	0.1551	0.8393	0.8034	0.9942	0.9993	0.9928	0.9877	0.9770	1.0	0.9451	0.9773	1.0	0.9924	1.0	1.0	1.0	0.9929	0.9722	0.9974	0.9949	0.9970	0.9941	0.9972	0.9967	1.0	1.0	0.9991	1.0	1.0	0.9890
0.0779	7.0	16751	0.0697	0.9698	0.9756	0.9727	0.9739	0.9983	1.0	0.9815	0.9904	1.0	0.9938	0.9935	0.9930	0.9994	0.9935	1.0	0.9903	0.9584	0.9206	0.9917	0.7753	0.9914	0.9315	0.8305	1.0	1.0	1.0	0.9939	1.0	1.0	1.0	0.1404	0.8382	0.8029	0.9958	1.0	0.9944	0.9910	0.9875	1.0	0.9480	0.9788	1.0	0.9924	1.0	1.0	1.0	0.9929	0.9747	0.9961	0.9949	0.9970	0.9925	0.9983	0.9967	1.0	1.0	0.9991	1.0	1.0	0.9953

Framework versions

Transformers 4.35.2
Pytorch 2.1.0+cu118
Datasets 2.15.0
Tokenizers 0.15.0

Isotonic/deberta-v3-base_finetuned_ai4privacy_v2

deberta-v3-base_finetuned_ai4privacy_v2

Useage

Model description