ClimateGPT is the first open source foundational AI-Platform, created by Endowment for Climate Intelligence (ECI), dedicated to addressing the impact of climate change.
The Endowment for Climate Intelligence (ECI) has unveiled ClimateGPT – the first open source ensemble of AI models dedicated to addressing the fast-moving impact of climate change. ClimateGPT seeks to drive resilient climate action for researchers, policymakers, and business leaders, to make informed decisions in this climate of uncertainty.
Available on Hugging Face, users can download the model, its research paper, and use a new AI lineage explorer to get visibility into its ClimateGPT training lifecycle. The breakthrough underscores the ECI’s commitment to the open science and open source AI communities, heralding a significant step toward responsible AI development.
The model benchmarks scores show a 10x the efficiency on climate-specific tasks and novel cascading machine translation that recovers nearly 94% of fluency performance compared to native multilingual language models. The result archives an equitable and audited AI model that is extensively fine-tuned by humans with diverse forms of expert and local stakeholder perspectives.
After over four years of research, testing, building, and fine-tuning more than 100 Large Language Models, Erasmus.AI developed the corpus of ClimateGPT from its planetary scale corpora —one of the world’s largest web and academic collections, with research and insights on climate, extreme weather, the Club of Rome’s Earth4All, and UN Sustainable Development Goals (SDGs).
The Erasmus corpus is drawn from over 10 billion web pages and millions of open-access academic articles. ClimateGPT is trained to synthesize interdisciplinary research and break silos to form a holistic understanding of the impacts of climate change across the natural, social and economic sciences.
In collaboration with AppTek’s AI and language research scientists, the model was trained with a new climate-specific instruction fine-tuning (IFT) dataset and benchmark that allows users to access knowledge across scientific disciplines in over 20 languages. EQTY Lab worked closely with Further Ventures to architect the ClimateGPT platform to leverage a new advanced cryptographic framework that authenticates, secures, and governs responsible AI models.
AI Powered by the Sun: The ECI launched the initial node of ClimateGPT during COP28 at Abu Dhabi’s Al Dhafra Solar PV, a 2-GW facility and the world’s largest single-site solar plant. Access to the model expanded today to Microsoft’s green energy data centers worldwide.
The ECI trained ClimateGPT on an array of 256 Nvidia H100s, the most-energy efficient cloud GPUs, and powered by hydroelectric energy in Puyallup, Washington.
Designed for Enterprise: Using a Retrieval-Augmented Generation (RAG) AI platform, enterprises can harness the open source model and apply it to real-time and proprietary datasets. Integrations into Salesforce and Databricks provide rapid deployment and fine-tuning of the model. The renewable energy company Masdar is also among the first adopters of the model.
Responsible AI: To ensure proper transparency and governance, the ClimateGPT leverages a new, advanced, trusted AI solution from EQTY Lab that registers the entire AI lifecycle on the Hedera enterprise-grade blockchain and preserves the model data on Protocol Lab’s IPFS and the Filecoin protocols. Responsible AI pilots have been initiated with experts from the open source, trust and safety, climate mis/disinformation communities to establish proper guardrails for the model’s deployment.
Daniel Erasmus, CEO Erasmus.AI, said, “This is more than an AI technical achievement; it is designed to accelerate our social intelligence together for the transition ahead. Policymakers, business leaders, and researchers can hopefully benefit from the decision support that this platform provides, to move us a little bit closer to a sustainable future.”
Christian Dugast, from AppTek, held, “Our IFT training set has been designed to support both completion and citation mechanisms, to teach the system to provide well-summarized answers and retrieve only the documents that are relevant to the prompt.”