Democratize Trustworthy and Efficient Large Language Model Technology for Europe

Informacje na temat projektu

TrustLLM

Identyfikator umowy o grant: 101135671

DOI

10.3030/101135671

Data podpisania przez KE 20 Października 2023

Data rozpoczęcia 1 Listopada 2023

Data zakończenia 31 Października 2026

Finansowanie w ramach

• Digital, Industry and Space

Koszt całkowity

€ 6 929 702,50

Wkład UE

€ 6 929 701,00

Koordynowany przez

LINKOPINGS UNIVERSITET
Sweden

Periodic Reporting for period 1 - TrustLLM (Democratize Trustworthy and Efficient Large Language Model Technology for Europe)

Okres sprawozdawczy: 2023-11-01 do 2025-04-30

Podsumowanie kontekstu i ogólnych celów projektu

The TrustLLM project, funded by the European Union, will develop trustworthy European large language models (LLMs) that will cover a range of underrepresented languages. The main objective for TrustLLM is the development of an open, trustworthy and factual LLM, initially targeting the Germanic languages. This will create the foundation for an advanced open ecosystem for next generation modular and extensible European trustworthy, sustainable, and democratised LLMs. The focus on Germanic languages can serve as a blueprint for future activities in other families of languages. The TrustLLM project and the surrounding ecosystem will enable, support, and improve context-aware human-machine interaction in a wide range of applications.

The consortium has unique expertise and practical experience in building LLMs, combined with leading NLP researchers, as well as organizations, working on transferring the technology to companies and end-users.

Prace wykonane od początku projektu do końca okresu sprawozdawczego oraz najważniejsze dotychczasowe rezultaty

Data Infrastructure & Processing
TrustLLM built a robust infrastructure for Germanic language models, creating multilingual datasets (Danish, Dutch, Faroese, German, Icelandic, Norwegian, Swedish) with strong privacy safeguards. The custom “TrustLLM Trove” framework enhanced open-source tools with quality filters and language ID for Germanic languages. Efficient text filtering and a TDM-compliant HTML pipeline ensured legal, high-quality data extraction.

Model Development & Training
A 7.8B parameter multilingual model was trained on 2.3T tokens across 17 Germanic variants, outperforming models like Llama-2 7B. LoRA-based adapters enabled efficient tuning for under-resourced languages like Icelandic and Faroese.

Evaluation & Applications
The EuroEval platform standardized evaluation across 7 tasks and 8 languages. Bias datasets for Danish and Dutch were introduced. Real-world tools include BijsluiterBot (medical info), BookBot (youth reading), Svarkur (Icelandic Q&A), and apps for automotive and accessibility.

Technical Innovation & Trust
Innovations include RAG-Ex for explainability, retrieval-enhanced transformer insights, tool-augmented reasoning, and dynamic tokenization for morphologically rich languages. These advances support transparent, reliable AI aligned with European values.

Innowacyjność oraz oczekiwany potencjalny wpływ (w tym dotychczasowe znaczenie społeczno-gospodarcze i szersze implikacje społeczne projektu)

Results beyond he state of the art are packages in form of publications in peer reviewed conference papers and journal articles. There are no more results beyond the state of the art at the moments which are produced by the project.