Member of Technical StaffFastino Labs
Correcting Stochastic Update Bias in Preconditioned Language Model Optimizers
arXiv preprint, 2026
NLP Researcher · Member of Technical Staff at Fastino Labs
Urchade Zaratiana
I work on small language models and structured prediction, with a particular focus on information extraction. My research explores how compact encoder models can produce reliable structured outputs for entity recognition, relation extraction, classification, safety, and privacy.
At Fastino Labs, I am responsible for developing small language model architectures, including the GLiNER family of models. I completed my PhD at the Laboratoire Informatique de Paris Nord (LIPN), supervised by Nadi Tomeh and Thierry Charnois.
I am French and Malagasy. I grew up in Madagascar and am now based in Réunion Island, after spending several years in Paris for my studies and research.
News
Released GLiNER2-PII, a multilingual model for personally identifiable information extraction.
Released GLiGuard, a schema-conditioned encoder for efficient LLM safeguarding.
Released Pioneer Agent on continual improvement of small language models in production.
Received the ATALA Best Thesis Award for doctoral research in natural language processing.
Selected publications
Complete list on Google Scholar-
2026
-
2026
Pioneer Agent: Continual Improvement of Small Language Models in Production
arXiv preprint, 2026Senior author
-
2026
GLiNER2-PII: A Multilingual Model for Personally Identifiable Information Extraction
arXiv preprint, 2026
-
2026
GLiGuard: Schema-Conditioned Classification for LLM Safeguard
COLM, 2026
-
2025
GLiNER2: An Efficient Multi-Task Information Extraction System with Schema-Driven Interface
EMNLP System Demonstrations, 2025
-
2024
GLiNER: Generalist Model for Named Entity Recognition using Bidirectional Transformer
NAACL, 2024
-
2024
An Autoregressive Text-to-Graph Framework for Joint Entity and Relation Extraction
AAAI, 2024
-
2024
GraphER: A Structure-aware Text-to-Graph Model for Entity and Relation Extraction
arXiv preprint, 2024
-
2023
Filtered Semi-Markov CRF
Findings of EMNLP, 2023
Research software
All repositoriesGLiNER
Generalist and lightweight model for named entity recognition with arbitrary entity types.
GitHub repositoryGLiNER2
Schema-driven multi-task system for NER, classification, and structured extraction.
GitHub repositoryGraphER
Structure-aware text-to-graph model for joint entity and relation extraction.
GitHub repositoryATG
Autoregressive span-based text-to-graph framework for information extraction.
GitHub repositoryGLiNER adoption
Updated 29 July 2026The GLiNER family has become a widely used open-source foundation for efficient information extraction, bridging research and production. Its compact encoder-based models make zero-shot entity recognition, relation extraction, classification, and structured extraction practical without relying on large generative models. The growing ecosystem now spans hundreds of community models and millions of monthly downloads across PyPI and Hugging Face.
937k
PyPI downloads · last 30 days
4.01M
Hugging Face downloads · last 30 days
PyPI ecosystem
Monthly downloads across all 16 GLiNER-related packages
Hugging Face ecosystem
Downloads across every public model carrying either tag
Background
Appointments
Research ScientistFI Group
Education
PhD in Computer ScienceUniversité Sorbonne Paris Nord
Software Engineering DegreeISEP
2025
ATALA Best Thesis Award. Awarded for outstanding doctoral research in natural language processing in France.