The Air Data Gazette · Air Datasets

Google DeepMind, University of Maryland, Oregon State University, Ellis Institute TübingenPrivacy Backdoors: Enhancing Membership Inference through Poisoning Pre-trained Models In this paper, we introduce a new type of backdoor, the privacy backdoor. Instead of causing a victim’s fine-tuned model to incorrectly classify examples at test time, as in many conventional backdoor attacks, a privacy backdoored model causes the victim’s model to leak details about the fine-tuning dataset.[...] . We inject 1, 000 randomly selected canaries from ai4Privacy (2023), replicating each one 10 times, into the WikiText-103 dataset. [...] Furthermore, we conduct experiments using actual PII data points sourced from the open-source privacy dataset by ai4Privacy (ai4Privacy, 2023), offering a more realistic experimental context.[...] n. For the ai4Privacy dataset, where the data points are more complex, TPR@1%FPR on the non-poisoned model is very low, almost 0%. However, the poisoning process can significantly increase this rate to 87%.uses 200k
Indiana University Bloomington, JD Cloud, Stanford & CISPAThe Janus Interface: How Fine-Tuning in Large Language Models Amplifies the Privacy Risks We further performed experiments on GPT-2 models with three privacy datasets (i.e., Enron [1], ECHR [14], and Ai4Privacy [2]).[...] In particular, in the targeted PII attack, with given PII identifiers, Janus extracted up to 35.19% private emails from Enron, 6.16% geological locations from ECHR, and 2.08% social security numbers from Ai4Privacy[...] Ai4Privacy. It contains over 300k examples covering 6 languages and 27 PII types in 749 discussion subjects. The dataset includes highly sensitive privacy data including driver’s license number (DLN) and social security number (SSN). On the Ai4Privacy dataset, prefix attacks fail to extract any correct social security numbers even with 100-tokens prefixes. [...] In this paper, we introduced a novel privacy attack, Janus, which exploits the fine-tuning interface to recover PIIs from the training data of LLMs. By modeling the privacy leakage problem as recovering PII association tasks, we empirically explained why forgotten PIIs can be recovered from LLMs and how to amplify such a privacy leak. Through experiments on both open-source language models and two latest LLMs, i.e., GPT-3.5-Turbo and LLaMA-2-7b, we demonstrated that Janus effectively amplified the privacy leak in LLMs and significantly outperformed the state-of-the-art privacy extraction attacks including prefix attacks and in-context learning (ICL). Our evaluation also revealed existing commercial fine-tuning APIs including OpenAI and Azure AI Studio failed to apply effective defenses against the Janus attack, allowing an attacker to conduct such a privacy attack at a low cost.uses 300k
Mitsubishi Electric Research Laboratories, Pennsylvania State UniversityForget to Flourish: Leveraging Machine-Unlearning on Pretrained Language Models for Privacy Leakage The latter is an opensource privacy dataset that holds real-life personal identifiable information (PII) data points. We inject 1,000 randomly selected samples from AI4Privacy into the WikiText 103 dataset. Fine-tuning large language models on private data for downstream applications poses significant privacy risks in potentially exposing sensitive information. Several popular community platforms now offer convenient distribution of a large variety of pre-trained models, allowing anyone to publish without rigorous verification. This scenario creates a privacy threat, as pre-trained models can be intentionally crafted to compromise the privacy of fine-tuning datasets. In this study, we introduce a novel poisoning technique that uses modelunlearning as an attack tool. This approach manipulates a pre-trained language model to increase the leakage of private data during the fine-tuning process. Our method enhances both membership inference and data extraction attacks while preserving model utility. Experimental results across different models, datasets, and fine-tuning setups demonstrate that our attacks significantly surpass baseline performanceuses 200k
Bennett University, Georgia Institute of Technology, Cloudemy Technology LabsWhispered Tuning: Data Privacy Preservation in Fine-Tuning LLMs through Differential Privacy In this paper, we proposed a novel multi-faceted approach called Whispered Tuning to address privacy leaks in large language models (LLMs). We integrated a PII redaction model, differential privacy techniques, and an output filter into the LLM fine-tuning process to enhance confidentiality. Additionally, we introduced novel ideas like the Epsilon Dial for adjustable privacy budgeting for differentiated Training Phases per data handler role. [...] We have developed a state-of-the-art PII classification model, specifically for identifying and classifying Personal Identifiable Information (PII) in the “pii-masking-65k” dataset.uses 65k
Northeastern UniversityRescriber: Smaller-LLM-Powered User-Led Data Minimization for Navigating Privacy Trade-offs in LLM-Based Conversational Agent Our evaluation dataset comprises 240 randomly selected samples from Ai4Privacy’s world’s largest open dataset for privacy [2], which encompasses 229 discussion subjects and use cases. These samples were processed by aligning the 54 PII classes from the dataset with the categories in our taxonomy (Table 1). To ensure accuracy, two authors independently reviewed and verified the privacy information labels and the corresponding text.uses 200k
Nanjing UniversityThe Fire Thief Is Also the Keeper: Balancing Usability and Privacy in Prompts To further validate our hypothesis, we randomly sampled 10,000 sentences from the PII-masking43k dataset [59], categorizing words into named entities (NE) and non-named entities (Non-NE), and computed the self-information of each word. Evaluating word importance and privacy before maskinguses 43k
TU WienPrivacy Agents: Utilizing Large Language Models to Safeguard Contextual Integrity in Elderly Care We provided the excerpt from the transcript shown in Section 3.2 to an openly available model that has been fine-tuned on the English subset of the dataset, called distilbert_finetuned_ai4privacy_v2 [19] We introduce a privacy agent that continuously monitors information flows of recorded conversations, and identifies key parameters that are compared to the contextual privacy norms to detect privacy violations.uses 200k, distilbert_finetuned_ai4privacy_v2
Vanderbilt UniversityThe Cleaned Repository of Annotated Personally Identifiable Information This paper introduces a new open-source dataset called the Cleaned Repository of Annotated Personally Identifiable Information (CRAPII). CRAPII is designed to test and evaluate automated deidentification methods for educational data. The dataset comprises over 20,000 student essays that have been annotated for personally identifiable information (PII).uses only cited, Only cited
Johnsnow labsDetect PHI for Deidentificationuses 400k
Brett YoungAI guardrails: Understanding PII detectionuses 400k
Presidio blogpostExperimenting with Automatic PII Detection on the Hub using Presidiouses Presidio, 300k
master(logged citation)
Baltic State Technical University “VOENMEH” named after D.F. UstinovI’m currently working on my diploma thesis, which focuses on privacy protection in fintech and insurance. After learning about your OpenPII and FinPII datasets, I believe they’d be a perfect fit for my research. Thesis Topic (Draft): "Privacy-Preserving NER Models in Fintech and Insurance: Improving PII Masking Techniques." My thesis will focus on developing models to identify and mask sensitive financial and insurance-related data. As fintech and insurance are highly regulated industries, ensuring the privacy of personally identifiable information (PII) is critical. The FinPII dataset, with its specialized classes for finance and insurance, would be invaluable for my research, as would the broader OpenPII dataset. My work involves building models to accurately detect and anonymize PII in text. Your dataset will provide the perfect foundation for training these models in a fintech and insurance context.
Sequoia Capitalused to train
studying Artificial Intelligence in Germany(logged citation)
Machine Learning and Bioinformatics, Song Laboratory School of Computer Science and Engineering Pusan National UniversitMy research is focused on the intersection of machine learning and bioinformatics, specifically on the application of privacy-preserving techniques in sensitive data processing. Research Title (Draft): "Development of Privacy-Preserving Models for Sensitive Data Anonymization in Financial and Insurance Sectors" In my current research, I am developing machine learning models that aim to efficiently identify and anonymize personally identifiable information (PII) in text data. Given the highly sensitive nature of financial and insurance-related data, I am particularly interested in the FinPII and OpenPII datasets, as they align perfectly with the objectives of my work. These datasets would provide a solid foundation for training and evaluating models that can accurately detect and mask PII, thus ensuring compliance with privacy regulations in these sectors.
ElephantLabsdataset releaseuses Inspired, same tags??, 300k
Holistic AI[data can be anonymized] by using entity labels, such as the pii-masking-200k dataset: (example)uses pii-200k, 200k
Analytics VidhyaGuide to Fine-tuning Gemini for Masking PII Data instructions using pii-200k to fine-tune gemini To this load_dataset() function, we pass in the dataset that we wish to download. Here in our example it is “ai4privacy/pii-masking-200k”, which contains 200k rows of masked and unmasked PII data.uses pii-200k, 200k
MAking ScienceProtecting Customer Privacy: How to Remove PII from Call Center Transcripts For this benchmark, we used the pii-masking-200k dataset. This is a public dataset that contains 200,000 pieces of text with PII data labeled. This dataset contains up to 54 different PII categories and the pieces of text are in four different languages (English, French, German, and Italian). It provides a comprehensive source of truth for model comparison since it contains many different PII categories in many languages. Accuracy results for 8 LLMs on 5 PII entities on our dataset. GPT-4 has best results (72-99%).uses pii-200k, 200k
RonathanSanitize Documents in the Browser with AI Tuned for PII ai4privacy/pii-masking-200k · Datasets at Hugging Face. We spent half an hour looking for PII datasets on Huggingface, and there are larger ones. However, I was using a google free tier, so I only trained on 60K.uses pii-200k, 200k
Google ColabNotebook: DSPy PII Masking Demo This notebook uses the dspy library to compile a program that extracts Personally Identifying Information from plain text. It uses the pii-masking-200k dataset from Hugging Face for the training and test data. The Basic program gets 22.5% accuracy using our custom metric. The BootstrapFewShotWithRandomSearch learns which examples to provide as few shot examples in the prompt. After compilation, the same CoT program as above has a performance of 66.9% with our custom metric.uses pii-200k
Towards Data ScienceHow Much Data Do You Need to Fine-Tune Gemini? (only available to Medium members) Loads the pii-masking-200k dataset and splits it into train and test sets.uses 200k
Innodata LabsInnodatalabs/rt2-pii-maskinguses 65k
Hirusha FernandoPII Masking with Python Using ai4privacyuses distilbert_finetuned_ai4privacy_v2
Nitish SinghProtecting PII Data in the AI Era: Next-Gen Sensitive Data Protectionuses distilbert_finetuned_ai4privacy_v2
Muhammet SANTASHow Good Are Open-Source LLM-Based De-identification Tools in a Medical Context?uses 300k
The learning agencyPII Detection | inference pii-masking-200kuses 200k
School of Computer Science and Engineering, Vellore Institute of Technology, Vellore, INDArtificial Intelligence and Privacy Concerns: Balancing Innovation With Securityuses 43k
School of Computer Science, University of Guelph, Guelph, ON N1G 2W1, CanadaPriv-IQ: A Benchmark and Comparative Evaluation of Large Multimodal Models on Privacy Competenciesuses Label set
Casimir Pulaski Radom University, 26-600 Radom, Poland, West Ukrainian National Unversity, Ternopil, 46009, Ukraine, KhmModern strategies for data leak detection and prevention in corporate networksuses 300k
Mälardalen University, School of Innovation, Design and Engineering. VÄSTERÅS, SWEDENHANDLING CONFIDENTIAL DATA IN LLM PROMPTSuses 200k
University of Michigan, Ann Arbor ‡ University of Wisconsin-Madison § Vanderbilt UniversityWhat Really is a Member? Discrediting Membership Inference via Poisoninguses Label set
School of Computer Science, Guangzhou University, Guangzhou, China The Hong Kong Polytechnic University, Hong Kong SAR, Attractive Metadata Attack: Inducing LLM Agents to Invoke Malicious Toolsuses 500k
Zhejiang University, University of VirginiaBe cautious when merging unfamiliar LLMs: A Phishing model capable of stealing privacyuses Label set
University of Michigan, University of Toronto and Vector Institute, University of Wisconsin-Madison, Langroid IncorporatPrϵϵmpt: Sanitizing Sensitive Prompts for LLMsuses 200k
University of Connecticut, CT, USA, Indiana University Bloomington, IN, USA, UC Santa Cruz, CA, USA, Peking University, Selective KV-Cache Sharing to Mitigate Timing Side-Channels in LLM Inferenceuses 200k
Khoury College of Computer Sciences, Northeastern University Boston, Massachusetts, USARescriber: Smaller-LLM-Powered User-Led Data Minimization for LLM-Based Chatbotsuses 200k
AI Tech EthicsUnmasking the Reality of PII Masking Models: Performance Gaps and the Call for Accountabilityuses Label set
Nantes Université, École Centrale Nantes, France.;Explore, Carquefou, FranceAdminSet and AdminBERT: a Dataset and a Pre-trained Language Model to Explore the Unstructured Maze of French Administrative Documentuses 200k
Dept. of C.S., The University of Texas at El Paso; Texas A&M University-Central Texas; Cisco Systems Inc; Amazon Web SerQuantifying the Capabilities and Risks of Large Language Models for PII Redactionuses 300k, 500k
Fine-Tuning Large Language Models: Assessing Memorization and Redaction of Personally Identifiable Informationuses 65k
Vanderbilt UniversityThe Cleaned Repository of Annotated Personally Identifiable Informationuses 200k
Nanjing UniversityThe Fire Thief Is Also the Keeper: Balancing Usability and Privacy in Promptsuses 43k
Hornetsecurity, Univ. Lille, Inria, CNRSτ TAU-EVAL: A Unified Evaluation Framework for Useful and Private Text Anonymizationuses 400k
University of Salento, Lecce, ItalyTitile: Detecting Personally Identifiable Information Through Natural Language Processing: A Step Forwarduses 43k, 200k
Wilmer E. HenaoAI4Privacy PII-Masking Dataset: Safeguarding Personal Data in AI Interactionsuses 200k
University of Groningen and Leiden UniversityEL-MIA: Quantifying Membership Inference Risks of Sensitive Entities in LLMsuses 200k
1 TikTok 2Department of Statistics, Rutgers UniversitySecret-Protected Evolution for Differentially Private Synthetic Text Generationuses No-only cited
† National University of Singapore, ‡ Google Research(Token-Level) InfoRMIA: Stronger Membership Inference and Memorization Assessment for LLMsuses 300k
Centific Global Solutions Inc.An Evaluation Study of Hybrid Methods for Multilingual PII Detectionuses 200k
Bytedance Inc. 2 Tsinghua UniversityPortcullis: A Scalable and Verifiable Privacy Gateway for Third-Party LLM Inferenceuses 43k
Beijing Advanced Innovation Center for Future Blockchain and Privacy Computing 2 Institute of Artificial Intelligence, FedDTRE: Federated Dialogue Generation Models Powered by Trustworthiness Evaluationuses 300k
AMADEUS FranceDiverse And Private Synthetic Datasets Generation for RAG evaluation: A multi-agent frameworkuses 200K, PWI-Masking-200K, and PHI-Masking-200K
Centific, United StatesScalable multilingual PII annotation for responsible AI in LLMsuses 200k
KAIST, South KoreaRefusal Is Not an Option: Unlearning Safety Alignment of Large Language Modelsuses 65k
1Computer Science Department, TU Darmstadt, 2Lab1141, 3Aleph Alpha Research, 4Hessian.AI, 5German Research Center for AMeasuring and Guiding Monosemanticityuses 300k
1Department of Computer Science, University of Manchester, Manchester, United Kingdom 2Mohamed bin Zayed University of GPT, But Backwards: Search-Based Language Model Inversionuses 400k
The Data Provenance InitiativeBRIDGING THE DATA PROVENANCE GAP ACROSS TEXT, SPEECH, AND VIDEOuses 200k
LG AI ResearchDo Not Trust Licenses You See: Dataset Compliance Requires Massive-Scale AI-Powered Lifecycle Tracinguses 200k, 300k
Moscow Institute of Physics and Technology 2.MTS AI 3. ITMO University 4.AIRISPY: Enhancing Privacy with Synthetic PII Detection Datasetuses 300k
Zhejiang University 2Hangzhou Yugu Technology 3Old Dominion University, Norfolk, VA, U.S.A. 4Colorado School of Mines, GUnderstanding PII Leakage in Large Language Models: A Systematic Surveyuses 300k
University of Massachusetts Amherst, USAProxyGPT: Enabling User Anonymity in LLM Chatbots via (Un)Trustworthy Volunteer Proxiesuses 500k
Adam ŁucekPII Masking RL Environmentuses 500k
Hugging facebarflyman/modernbertuses 500k
Prabigya Acharya, Liza ShresthaA COMPARATIVE STUDY OF LIGHT-WEIGHT LANGUAGE MODELS FOR PII MASKING AND THEIR DEPLOYMENT FOR REAL CONVERSATIONAL TEXTSuses 200k
Heba Aljabrine, Md. Mahfuzur RahmanTowards Efficient and Privacy-Preserving Cloud Computinguses 65k
Frista Elsa Haleluyah, Ni Nyoman Fillicia Immanuella, Theresia Seplina, Ika DyahTowards Scalable Privacy-Aware Text Classification: A DistilBERT-Based Approach to PII Identificationuses distilbert_finetuned_ai4privacy_v2
CISPA Helmholtz Center for Information Security, Nokia Bell LabsDefeating Cerberus: Concept-Guided Privacy-Leakage Mitigation in Multimodal Language Modelsuses 200k
ShanghaiTech University, Independent ResearcherCausal-Guided Detoxify Backdoor Attack of Open-Weight LoRA Modelsuses 200k
AgentLeak: A Full-Stack Benchmark for Privacy Leakage in Multi-Agent LLM Systemsuses 200k
BodhiPromptShield: Pre-Inference Prompt Mediation for Suppressing Privacy Propagation in LLM/VLM Agentsuses other
Safer Reasoning Traces: Measuring and Mitigating Chain-of-Thought Leakage in LLMsuses 200k
Contrastive Negative Preference Optimization for Machine Unlearning in LLMsuses 200k
Detecting and Perturbing Privacy-Sensitive Neurons to Defend Embedding Inversion Attacksuses 300k
Concept-Aware Privacy Mechanisms for Defending Embedding Inversion Attacksuses 300k
SafeGen: Benchmarking Inference-Time Methods for Privacy-Preserving Text Generationuses 400k
Adaptive Text Anonymization: Learning Privacy-Utility Trade-offs via Prompt Optimizationuses 500k
WebPII: Benchmarking Visual PII Detection for Computer-Use Agentsuses other
(SKIP)Sripaad/ai4privacy (GitHub)uses 200k
The Last Ones | AI4Privacy [2024] - Neural Wave hackathon projectuses 400k
(SKIP)Ari-S-123/pii-masking
(SKIP)Kostis-S-Z/pii_detector
(SKIP)Ouroborosrex/LLAMA3_Finetuning_PIIuses 300k
How to offline load deberta-ai4privacy model and spacy en-core-web-sm with llm-guard (Issue #215)
dockersamples/labspace-fine-tuninguses 400k
(SKIP)AdamiTitus/pii-masking-300k (third-party HF mirror)uses 300k
AI4Privacy Benchmark Overview
ai4privacy/pii-masking-300k on Oxen.aiuses 300k
Supervised Fine-Tuning (SFT) for PII Masking Using Axolotluses 300k
Supervised Fine-Tuning (SFT) for PII Masking Using Axolotl (Medium)uses 300k
Fine-tuning LLaMa 3.2 3B for PII Masking with Tool Calling: Unsloth + LoRAuses 200k
Best Open Source Models for PII Redactionuses 300k
How to protect PII in LLM solutions
How Good Are Open-Source LLM-Based De-identification Tools in a Medical Context? (John Snow Labs blog)uses 300k
(SKIP)Ai4Privacy - Company Profile
Shadow Unlearning: A Neuro-Semantic Approach to Fidelity-Preserving Faceless Forgetting in LLMsuses 400k (inferred)
Section of Cybersecurity Engineering, Department of Applied Mathematics and Computer Science, Technical University of DePromptShield: Protecting User Privacy in AI Chatbotsuses 400k
ILCC, University of Edinburgh, ILLC, University of AmsterdamControlling What You Share: Assessing Language Model Adherence to Privacy Preferencesuses 200k
Independent ResearcherPIIBench: A Unified Multi-Source Benchmark Corpus for Personally Identifiable Information Detectionuses 400k&300k
1.Industrial System Institute, Athena Research Center, 26504 Patras, Greece 2.Archimedes, Athena Research Center, 15125 On the Applicability of LLMs and SLMs for Privacy-Preserving Named Entity Recognition in Financial Applicationsuses 43k
School of Software Technology, Zhejiang University 2Zhejiang University 3Hefei University of Technology 4 Institute Data-Free Privacy-Preserving for LLMs via Model Inversion and Selective Unlearninguses 200k
Cornelius and Rinaldi Genomics & InformaticsMeasuring the gap: correlating synthetic‑to‑real drift with PHI de‑identification performanceuses 200k
Independent ResearcherAegis DLP: A Unified Endpoint Security Platform with Deep Learning-Based Threat Detectionuses Not specified
Shenzhen Institute of Computing Sciences, China, University of Edinburgh, United Kingdom, and South China University ofShielding PII to Prevent Re-identification and Preserve Utilityuses 500k
Institute of Management Science, TU Wien, 1040 Vienna, AustriaPrivacy at the Core: Toward Automated Detection of Privacy-Sensitive Content in an LLM-Based Care Documentation Support Systemuses 200k
Technische Universitat Berlin ¨Berlin, GermanyAddressing Labelled Data Scarcity: Taxonomy-Agnostic Annotation of PII Values in HTTP Traffic using LLMsuses 200k
Independent ResearcherFine-Tuning Over Architectural Complexity: Broad-Coverage PII Detection on PIIBench with DeBERTauses 300k&400k
ServiceNowREDACT: A Systematically Controlled Multilingual Benchmark for Personal Information Detectionuses 500k&1m
UC Riverside Riverside, California, USA and CSIT, Queen’s University Belfast UKLoss Landscape Poisoning: Targeted Extraction of Unseen Training Data from LLMsuses 200k
Independent ResearcherSecurity Document Classification with a Fine-Tuned Local Large Language Model: Benchmark Data and an Open-Source Systemuses Not specified
1Centre for Cyber Security, Faculty of Information Science and Technology, Universiti Kebangsaan Malaysia (UKM), Bangi, Predicting IoT Privacy Concerns From Tweets Through BERTweet-Based Fusion of Psycholinguistic Big Five and Risk Features: Toward Integration With IoT Security Frameworksuses 500k
School of Computer Science and Engineering, Ningxia Institute of Science and Technology, Shizuishan 753000, ChinaOptimized MS-SIDM: A Context-Aware Sensitive Information Detection Framework via LLM Reasoning and Non-Linear Risk Quantificationuses 200k
Independent ResearcherSecureOutbox: A Local-First, Context-Aware Browser-Extension Framework for Preventing Sensitive Data Exfiltration via Quantized Small Language Modelsuses 200k
Independent Researchers &University Of South Dakota, Vermillion, USAPrivacy Sentinel: A Multilingual Context-Aware Framework for Real-Time PII Detection and Risk-Based Redactionuses Not specified
Purdue University & University of ExeterPAAC: Privacy-Aware Agentic Device-Cloud Collaborationuses 400k
University of Waterloo & Vector InstituteCAPID: Context-Aware PII Detection for Question-Answering Systemsuses 300k
HKUST, NUS, Tongyi Lab, & Alibaba GroupCDR-Bench: Evaluating Faithful Execution of Compositional, Order-Sensitive Data Refinement Recipesuses 1M,400k&200k
A10 Networks, Inc.REDACTIONBENCH
Public University of Navarre Pamplona, Spain 2.Vicomtech San Sebastian, SpainIdentification and Anonymization of Named Entities in Unstructured Information Sources for Use in Social Engineering Detectionuses 400k
Stanford UniversityWEBPII: BENCHMARKING VISUAL PII DETECTION FOR COMPUTER-USE AGENTSuses 43k
Northeastern UniversityAPWA: A Distributed Architecture for Parallelizable Agentic Workflowsuses 300k
Department of Computer Science and Engineering University of California, San DiegoCyberMaskQA: A Privacy-Aware Benchmark for Evaluating Large Language Models in Cybersecurity Question Answeringuses 300k
Lenovo Research, Beijing 100094, ChinaTrustFusion AI: A Deep Fusion Architecture for Trusted AI Applicationsuses 200k
Trustworthy Human Language Technologies, Technical University of Darmstadt, Germany, and Research Center for TrustworthyThe Conundrum of Trustworthy Research on Attacking Personally Identifiable Information Removal Techniquesuses 500k
Northeastern UniversityReconstruction of Personally Identifiable Information from Supervised Finetuned Modelsuses Not specified
University of HelsinkiLabel-Space-Aware Evaluation of Sparse Autoencoder Feature Monosemanticityuses 300k
King Mongkut's University of Technology North BangkokEnhancing Personally Identifiable Information Recognition Using Compact Transformer With Hybrid Sequential Architectureuses 43k
IOE, Pulchowk CampusA COMPARATIVE STUDY OF LIGHT-WEIGHT LANGUAGE MODELS FOR PII MASKING AND THEIR DEPLOYMENT FOR REAL CONVERSATIONAL TEXTSuses 200k
Leibniz University HannoverA Reproducible Web-Scraping Pipeline for Collecting Geolocated German Online Job Advertisements at Scaleuses 500k
Independent Researcher. Zürich, Switzerlandgheim-ch-pii-212k and gheim-ch-560m: An Open Swiss-Language PII Dataset and NER Model. Technical report.uses 1M&500k

The Air Data Gazette.

Built something on our data?