Ai2’s new solution allows researchers and developers to trace model-generated text back to the original documents in the model's training data—unlocking new levels of trust and transparency

SEATTLE: As AI adoption accelerates—even in high-stakes industries like healthcare, life science, finance, and security—one pressing challenge remains: how can we trust and fully harness its potential if we can’t understand how it makes decisions? The lack of transparency in LLMs limits not only our ability to build trusted AI solutions that can be validated, audited, and regulated - it also fundamentally limits our ability to build better, scientifically grounded AI.

To help address that gap, today the Allen Institute for AI (Ai2) announced OLMoTrace, a one-of-a-kind feature that enables you to trace the outputs of language models back to their multi-trillion-token training data in real time. This level of traceability helps you understand hallucinations and fact check model responses, supports more effective model debugging, and provides the data traces required for AI governance, auditing, and regulation.

OLMoTrace is a unique, first-of-its-kind addition to the Ai2 Playground—a platform interface that enables users to interact with Ai2’s advanced, fully open language models. Ai2’s models are uniquely capable of enabling a feature like OLMoTrace because all data behind them is fully open.

“OLMoTrace marks a pivotal step forward for the future of AI development, laying the foundation for more, transparent AI systems that researchers and developers can better understand,” said Jiacheng Liu, lead researcher for OLMoTrace. “By offering greater insight into how AI models generate their responses, anyone using our models can ensure that the data supporting their outputs is trustworthy and verifiable. This data traceability is essential not only for researchers and developers learning more about how these systems work, but for anyone who wants to build solutions with a verifiable AI model they can trust.”

A New Era of AI Transparency

The OLMoTrace feature introduces an unprecedented level of transparency, allowing users to inspect the relationship between a model’s output and its training data. By analyzing long, unique text spans within the model’s responses, this tool provides researchers, developers, and the public with a new opportunity to understand how AI systems "learn" and utilize specific information. Users can now verify the sources behind key model outputs, gaining new insight into both factual and creative content generated by the AI.

Currently available with Ai2’s flagship open model, OLMo 2 32B, which is trained on a vast and diverse dataset of over 3.2 billion documents, OLMoTrace offers users visibility into these documents, the ability to explore their relevance to specific model responses, and to understand how the AI integrates a wide array of data sources to generate its outputs.

Ai2 Playground Now Enables Everyone to Look Inside LLMs

Activating OLMoTrace is simple yet powerful. After generating a response from OLMo in the Ai2 Playground, users can activate the tool using the “Show OLMoTrace” button below the output. OLMoTrace quickly scans all 3.2 billion documents in the model’s training corpus and highlights text spans in the response that have a match in the training data. Each highlighted span is linked to a set of documents, enabling users to explore the original sources and see where and in what context these phrases appeared.

The tool’s design ensures that the most unique and relevant spans are prioritized, with matches ranked by a retrieval relevance score.

Advancing Open Science and Public Understanding

OLMoTrace is a testament to Ai2’s commitment to transparency and advancing an open scientific approach to AI. This pioneering feature showcases how complex AI models can be demystified and made accessible to a broader audience. By unlocking access to the data behind AI outputs, Ai2 is not only enhancing the public's understanding of machine learning, but also driving collaboration, accountability, and responsible AI development.

Availability

OLMoTrace is available now on the Ai2 Playground for the OLMo 2 32B Instruct and OLMo 2 13B Instruct and OLMoE 1B 7B Instruct models. OLMoTrace is hosted by Google Cloud.

Learn More

For more information about OLMoTrace, please visit:

• Try it now on the Ai2 Playground
• Learn more on our blog
• Watch the demo

About Ai2

Ai2 is a Seattle-based non-profit AI research institute with the mission of building breakthrough AI to solve the world’s biggest problems. Founded in 2014 by the late Paul G. Allen, Ai2 develops foundational AI research and innovative new applications that deliver real-world impact through large-scale open models, open data, robotics, conservation platforms, and more. Ai2 champions true openness through initiatives like OLMo, the world’s first truly open language model framework, Molmo, a family of open state-of-the-art multimodal AI models, and Tulu, the first application of fully open post-training recipes to the largest open-weight models. These solutions empower researchers, engineers, and tech leaders to participate in the creation of state-of-the-art AI and to directly benefit from the many ways it can advance critical fields like medicine, scientific research, climate science, and conservation efforts. For more information, visit allenai.org.

Fonte: Business Wire