Beitrag von Helmholtz Information

🌟 Now available: OpenGPT-X model Teuken-7B 🌟 We’re excited to announce the public release of Teuken-7B, a multilingual, open-source large language model #LLM developed from the ground up and tailored to Europe’s linguistic diversity 🇪🇺 . What sets Teuken-7B apart? ✨ Trained in all 24 official EU languages, with over 50% non-English data ⚡ Includes a custom multilingual tokenizer, enhancing multilingual performance and efficiency across European languages 🔬 Backed by a comprehensive technology stack from data pre-processing to evaluation (see our 🏆 European LLM Leaderboard: https://lnkd.in/exzUBvxf) Two instruction-tuned versions to choose from: 1️⃣ Open source version for research and commercial use under the Apache 2.0 license 2️⃣ Research-only version Teuken-7B was developed by the OpenGPT-X consortium with the computational resources of two of Germany's leading #HPC centers, the JUWELS system operated by Jülich Supercomputing Centre (JSC) at Forschungszentrum Jülich and the HPC systems operated by ZIH at Technische Universität Dresden. ⬇️ Download Teuken-7B and model cards on Hugging Face: https://lnkd.in/ep5zFWaP 🤝 Connect with the developers on Discord: https://lnkd.in/espgZ2Pb 🔍 Learn more on our website: https://lnkd.in/egcCJcQV Thanks to all consortium partners for making this possible Fraunhofer IAIS Fraunhofer IIS Deutsches Forschungszentrum für Künstliche Intelligenz (DFKI) Forschungszentrum Jülich Technische Universität Dresden IONOS Aleph Alpha ControlExpert GmbH WDR KI Bundesverband The OpenGPT-X research project is funded by the German Federal Ministry of Economic Affairs and Climate Action (BMWK) and will run until 31 March 2025.

🌟 Now available: OpenGPT-X model Teuken-7B 🌟 We’re excited to announce the public release of Teuken-7B, a multilingual, open-source large language model #LLM developed from the ground up and tailored to Europe’s linguistic diversity 🇪🇺 . What sets Teuken-7B apart? ✨ Trained in all 24 official EU languages, with over 50% non-English data ⚡ Includes a custom multilingual tokenizer, enhancing multilingual performance and efficiency across European languages 🔬 Backed by a comprehensive technology stack from data pre-processing to evaluation (see our 🏆 European LLM Leaderboard: https://lnkd.in/exzUBvxf) Two instruction-tuned versions to choose from: 1️⃣ Open source version for research and commercial use under the Apache 2.0 license 2️⃣ Research-only version Teuken-7B was developed by the OpenGPT-X consortium with the computational resources of two of Germany's leading #HPC centers, the JUWELS system operated by Jülich Supercomputing Centre (JSC) at Forschungszentrum Jülich and the HPC systems operated by ZIH at Technische Universität Dresden. ⬇️ Download Teuken-7B and model cards on Hugging Face: https://lnkd.in/ep5zFWaP 🤝 Connect with the developers on Discord: https://lnkd.in/espgZ2Pb 🔍 Learn more on our website: https://lnkd.in/egcCJcQV Thanks to all consortium partners for making this possible Fraunhofer IAIS Fraunhofer IIS Deutsches Forschungszentrum für Künstliche Intelligenz (DFKI) Forschungszentrum Jülich Technische Universität Dresden IONOS Aleph Alpha ControlExpert GmbH WDR KI Bundesverband The OpenGPT-X research project is funded by the German Federal Ministry of Economic Affairs and Climate Action (BMWK) and will run until 31 March 2025.

Combining large and small language models for decoding strikes a balance between quality and efficiency. Learn how to overcome memory and compute challenges in Richa Gadgil's newest article.

UChicago Genie is now open source! How we built a zero-hallucination AI chatbot that answered over 10000 questions of students at the University of Chicago: Ingredients 🫘 (Techstack): Next.js + Vercel AI SDK + Neo4j Graph RAG This is how we did it:👇 1) Scraped the Course Catalogue Website of UChicago with an LLM to build an intelligent Knowledge Graph that connects all the courses, departments, instructors, etc.. together  2) Configured GPT-4o to understand the intent of the question of students 3) Gave the chatbot access to the Knowledge Graph through Agentic Graph RAG Inner workings: 🧪 When a user asks a question such as 'What are the prerequisites for MATH 207?', we use a GPT-4o Agent to understand the query's intent. The agent then calls an LLM tool to retrieve the relevant information from the user query and then searches for the information in a Neo4j knowledge graph database. Once the information is fetched UChicago Genie outputs a response using a predefined template, such as 'The prerequisite for MATH 207 is x,'. Here is your Hallucination-free answer. What’s amazing about graph RAG is how good it can get at answering hard questions. Planning double major track,  making morning schedules, navigating complicated prerequisites for classes… These are all matters that the UChicago Genie will eventually be able to address. We just need to build it. Check out UChicago Genie’s source code on Github! (see our website) Please reach out to us via our websiteAre interested in getting access to the LLM Knowledge Graph that powers the Graph RAG of UChicago Genie. #OpenSource #GPT4o #GraphRAG #NextJS #Neo4j

How does user language affect the information provided by large-language models? Our research provides novel empirical evidence on this question focusing on the context of the Israeli–Palestinian and Turkish–Kurdish conflicts. We employed an automated query procedure in ChatGPT to repeatedely inquire about casualties in specific airstrikes. We iteratively asked the same questions in Hebrew and Arabic for the Israeli-Palestinian conflict, and in Turkish and Kurdish for the Turkish-Kurdish conflict. Our analysis reveals that ChatGPT provides 34±11% lower fatality estimates when queried in the language of the attacker than in the language of the targeted group. Our evidence further shows that the large-language model describes the airstrikes as more violent and bloodier when queried on the language of the targeted group. For instance, when queried about Israeli airstrikes in Gaza, ChatGPT is on average twice as likely to mention civilian casualties in Arabic compared to Hebrew, three times more likely to mention female victims in Arabic, and six times more likely to mention children killed in Arabic. Where does this bias come from? First, large-language models reproduce broader biases that are present in their training data. Hence, to the extent that online media reports on these airstrikes differ across languages, this will shape the responses provided by the ChatGPT. On top of this media bias, we show a novel source of bias that has no equivalent in classical search engines. ChatGPT retrieves casualty numbers from the language-specific training data but fails to correctly match them to the corresponding events. Due to its reliance on co-occurring words, ChatGPT may report casualty counts from different attacks with greater news impact, or provide cumulative casualty counts that are prevalent in the news media. Because such high-profile events are systematically different from low-profile events – for example, involving more deaths and more civilian casualties – the reported death counts tend to be inflated. However, what constitutes a high-profile event is likely to be language-specific, leading to systematic differences across languages. What are the broader consequences? Our findings suggest that citizens of states that have conducted airstrikes may underestimate the human toll of these attacks based on the information obtained through LLMs. In contrast, citizens of attacked groups may perceive these airstrikes as especially brutal and indiscriminate based on the information available in their language. These antithetical perceptions may contribute to radicalized identities. To the extent that LLMs are increasingly used for information purposes, potentially through their integration in search engines such as Microsoft Bing or Google Gemini, this bias could promote information bubbles in the future. Check out the full paper with the following link: https://lnkd.in/eBusAKNn #AI #SocialAI #LLMs

New agentic powerhouse: Anthropic vastly upgrades Claude 3.5 Sonnet, leaving GPT-4o in the dust! 🚀 And on top of this, they also release a brand new Haiku in version 3.5.(Haiku previously only existed in version 3) ⚡️ Claude 3.5 Sonnet: a powerhouse tailored for agentic reasoning ▸ On coding, it improves performance on SWE-bench Verified from 33.4% to 49.0%, scoring higher than all publicly available models—including reasoning models like OpenAI o1-preview and specialized systems designed for agentic coding. ▸ Yet on hard problems like Math olympiad problems (AIME ‘24), even though it’s in front of GPT-4o (16% vs 5%), it seems far behind the score of 74% announced by OpenAI o1. But this is expected as o1 expands many more tokens! ▸ Available for the same price and speed as the previous Claude 3.5 Sonnet 🏎️ New Claude 3.5 Haiku: ▸ As blazing fast as previous Haiku but matches Claude 3 Opus performance ▸ Amazing 40.6% on SWE-bench (better than original 3.5 Sonnet!) 🖥️ On top of this, they release a revolutionary "computer use" feature: ▸ Claude can now control computers like humans: by clicking, scrolling and typing! ▸ Evaluated on OSWorld, an environment made fo revaluating end-to-end agents that use an OS with mouse and keyboard, Claude reached 22% score, leaving the previous best of 8% in the dust The new Claude-3.5-Sonnet sounds like a game-changer for agentic applications. I really look forward to trying it in agentic workflows, especially on GAIA. Game is on - now if OpenAI doesn’t react, they’ll start having a really hard time selling GPT-4o! 👀