Skip to content

Welcome to GenSIE 2026

GenSIE (General-purpose Schema-guided Information Extraction) is a shared task at IberLEF 2026 focusing on the ability of systems to extract nested, structured information (JSON) from general-domain Spanish texts.

Read the full task description, including score metrics and detailed constraints.

The task challenges participants to use Small Language Models (SLMs) and inference-time techniques to handle Zero-Shot Schemas—where the extraction target is defined dynamically at runtime.

  • Zero-Shot Schema

    Extract data using schemas seen only at inference time. No fixed ontology.

  • General Domain

    From legal contracts to medical reports and news.

  • Inference-Time Focus

    Focus on prompting, RAG, and constrained decoding. No massive fine-tuning.

  • Structured Output

    Strict adherence to JSON Schema and complex semantic constraints.

Resources

Explore our technical documentation to get started:

Schedule

Date Event
✅ March 06, 2026 🚀 Starter Kit Released (View Guide)
✅ April 15, 2026 📂 Dev Set Released (150 silver instances, for development)
May 22, 2026 🛑 Submission Deadline (extended from May 15)
May 22, 2026 🔓 Test Set Release (100 instances, immediately after submissions close)
May 22–31, 2026 ⚙️ Evaluation Period (fast turnaround)
June 01, 2026 🏆 Results Announcement
June 07, 2026 📝 First Round of Papers (for those who want early feedback)
June 12, 2026 📝 Paper Feedback
June 19, 2026 📝 Camera-Ready Paper Deadline
Sept 22, 2026 🎤 IberLEF Workshop (León, Spain)

News & Updates

  • May 12, 2026: Submission deadline extended to May 22, 2026 (one extra week, by participant request). The test set (100 instances) will be released the same day, immediately after submissions close, and evaluation will run with a fast turnaround toward the June 1 results date. We have also finalized the evaluation protocol: a new primary leaderboard based on the fraction of the baseline-to-perfect gap each system closes, averaged over several evaluation models (some published/recommended, some held out); and token/time limits are now soft averages over the test set rather than hard per-instance caps. See the updated Task Description and Submission Guidelines.
  • April 15, 2026: Dev set released! 150 silver instances across 8 domains available in data/dev/dev.jsonl. Generated with Claude Opus 4.6, audit-passed, for development use. Baseline scores coming soon. See Starter Kit.
  • April 15, 2026: Schedule updated - submission blind (May 8), test set private (May 15), evaluation May 16-31, results June 1, camera-ready June 15.
  • April 15, 2026: Evaluation metrics updated with null hallucination penalty and list matching documentation. See Task Description.
  • April 15, 2026: We apologize for the delay in releasing the dev set. Thank you for your patience!
  • March 01, 2026: We've had some delays with the preparation of the starter-kit which forced to push the date back to March 09 at the latest.
  • Jan 26, 2026: Website launched.

Motivation

The rise of Agentic Workflows has created a massive demand for systems that can communicate via structured protocols. To identify user intent, invoke external tools, or exchange information, an AI must output rigid, error-free structured data.

While massive proprietary models (like GPT-5) solve this through scale, GenSIE targets the innovation gap in Small Language Models (<14B). We aim to prove that with clever engineering (Chain-of-Thought, ReAct, Constrained Decoding), commodity hardware can perform complex structured extraction reliably.

Furthermore, we aim to prioritize efficiency and sustainability to ensure that high-performance extraction pipelines remain deployable in real-world scenarios. By focusing on models that run on consumer-grade hardware, we promote sustainable AI and cost-effective solutions that are accessible to smaller research groups and industry practitioners.

Organizing Committee

The GenSIE task is organized by a consortium between the Research Group on Artificial Intelligence and Data Science (GIA-UH) at the University of Havana and the Research Group in Natural Language Processing and Information Systems (GPLSI) at the University of Alicante.

This team brings together expertise in both Computer Science (Generative AI, Large Language Models) and Linguistics (Corpus Annotation, Semantic Evaluation).

Members

Name Affiliation Role
Yudivian Almeida Cruz University of Havana PhD, Professor
Suilan Estévez Velarde University of Havana PhD, Professor
Alejandro Piad Morffis University of Havana PhD, Professor
Isabel Espinosa Zaragoza University of Alicante PhD, Assistant Professor
María Miró Maestre University of Alicante PhD, Postdoc Researcher
Lucía Sevilla Requena University of Alicante PhD Student, Assoc. Prof.
Alba Pérez Montero University of Alicante PhD Student
Ernesto Estevanell Valladares University of Havana PhD Student

Contact

For questions regarding the task, dataset, or evaluation, please contact the corresponding author, Alejandro Piad Morffis.