Antonio Gelameris

We just published 24 Models 🌏 🌎 🌍 The languages are: Armenian, Basque, Brazilian Portuguese, Bulgarian, Catalan, Croatian, Czech, Galician, Georgian, German, Irish Gaelic, Kyrgyz, Naija, Polish, Scottish Gaelic, Sindhi, Slovenian, Spanish, Tamil, Turkish, Ukrainian, Uyghur, Welsh and Wolof. Our models are dependency parsers, similar to UDPipe and Stanza. They are built using biLSTM architectures and designed for two morphologically complex language types: Inflected and Agglutinative. Each model card includes information on the corpus/corpora used in training and benchmarks (including comparisons to Stanza where applicable). We record details on the energy used in training and the pollution incurred. For example, German is a flagship model. The training ran for 31.5 hours consuming 47 KWH of power and emitting the same pollution as driving an average car 30 miles or consuming 15 cheeseburgers. It has stunning performance with a LAS of 91.50% (Stanza was 93.56 %). Our Wolof model consumed 0.24 KWH of power, emitting the same pollution as driving 3 miles in an average car or consuming 1/2 a cheeseburger. It reached a credible LAS of 59.10% (Stanza was 77.32%). You can run these models on windows or ubuntu by downloading the files and running a python script which has been designed using the openVINO library - enabling maximum x86 compatibility for inference, including older Intel and AMD CPUs . Our models are rubbish for Isolating / Analytic languages, such as English and Mandarin. And that is the whole point. Our models are designed to capture the symbolic information in conllu corpora to instruct our models how to identify morphological complexity. Detailed write up to follow, but if you are a trained linguist and interested in any of the published models please get in touch. We are planning the next iteration for more state-of-the-art performance (German is ready for dialectal adaptation to Bavarian and Spanish is close behind). Huge thanks to Denvr for inclusion in their AI Ascend program and Intel Corporation for all the support on the Liftoff program. #IndigenousLanguages #LowResourceLanguages #NoLanguageLeftBehind https://lnkd.in/eUiiQuK9

5

2 commenti

𝗪𝗛𝗘𝗡 𝗬𝗢𝗨𝗥 𝗗𝗢𝗠𝗔𝗜𝗡 𝗥𝗨𝗟𝗘𝗦 𝗕𝗥𝗘𝗔𝗞 https://lnkd.in/dy8zDQyU Real-world systems rarely follow clean architectural boundaries. Business rules cross aggregates. Decisions span multiple concepts. Consistency becomes difficult. At CodeCrafts 2026, Sara Pellegrini introduces 𝗗𝘆𝗻𝗮𝗺𝗶𝗰 𝗖𝗼𝗻𝘀𝗶𝘀𝘁𝗲𝗻𝗰𝘆 𝗕𝗼𝘂𝗻𝗱𝗮𝗿𝗶𝗲𝘀 — a new way to model complex constraints in event-sourced systems. Instead of forcing rules into rigid structures, boundaries can evolve as 𝘁𝗵𝗲 𝗱𝗼𝗺𝗮𝗶𝗻 𝗯𝗲𝗰𝗼𝗺𝗲𝘀 𝗰𝗹𝗲𝗮𝗿𝗲𝗿. A talk about 𝗺𝗼𝗱𝗲𝗹𝗹𝗶𝗻𝗴 𝗰𝗼𝗺𝗽𝗹𝗲𝘅 𝘀𝘆𝘀𝘁𝗲𝗺𝘀 without locking yourself into the wrong architecture too early.

21

2 commenti

We have now published Matteo's tutorial video from The Arathon - A Hackathon on vector processors and Ara. Find it here: https://lnkd.in/d7AiaksM

30

⚙️ SKILLRL AND THE FIGHT AGAINST TOKEN BLOAT Many agent designs are getting expensive for a simple reason: they burn tokens to “think”. This paper proposes evolving agent skills recursively, aiming to reduce token footprint while improving reasoning utility. If approaches like this work, they could shift the economics of agentic AI dramatically, making longer-horizon tasks more feasible. A few implications for builders: • Efficiency becomes a first-class design goal, not a cost afterthought • Skills can be optimised and reused, reducing repeated reasoning • Evaluation must include cost and latency alongside success rate The next wave of AI products will compete on outcomes per pound, not only on cleverness. What matters more in your use case: the best possible answer, or the best answer at a predictable cost? Sources: - https://lnkd.in/ehrKqB4w #ArtificialIntelligence #Agents #Efficiency #AIResearch

1

How can domain understanding lead to more meaningful refactoring decisions? In Domain-Driven Refactoring, Alessandro Colla and Alberto Acerbis offer a practical approach to evolving software systems by aligning technical changes with business context. Now recognized as the #1 New Release in Systems Architecture on Amazon, this book is a valuable resource for software engineers and architects working with complex or aging codebases. 📖 Explore the book:https://lnkd.in/dK4u-bya #SoftwareArchitecture #DomainDrivenDesign #Refactoring #TechnicalLeadership

35

1 commento

🚀 Exciting news! Our paper is now published in the Journal of Cloud Computing: “ArtifactOps and ArtifactDL: a methodology and a language for conceptualizing and operationalising different types of pipelines” Raul Miñon PhD, Josu Díaz de Arcaya, Ana Isabel Torre, Juan Lopez de Armentia Mendizabal, Gorka Zárate Martinez, Lander Bonilla Viana, Asier García Pérez, and Jon Aguirre Usandizaga This work presents: 🔹 ArtifactOps: Unifying DevOps, DataOps & MLOps with artifacts as first-class citizens. 🔹 ArtifactDL: A declarative language to model pipelines — CI/CD, IaC, data, training, inference — consistently and expressively. 🔹 Infrastructure & Governance: Modeling, rules, and deployment strategies all in one framework. Conducted within the PLIADES project, this research bridges the gap between academic insight and real-world deployment, supporting monitoring, packaging, and multi-output pipelines. 📖 Read more here: https://lnkd.in/dHZdB-Nu

19