First-ever education-specific language models open door to trustworthy generative AI for teachers and students

Today we are releasing three large language models that open the door for trustworthy generative AI in education.

Built for the unique workflows and safety needs of education, our open-source LLMs are components of a broader generative AI platform for education that our team is building. The platform will enable teachers and students to have a generative AI experience that retrieves content from curriculum chosen by the user, not from the entirety of the internet. The result is an engagement that is curriculum-aligned, hallucination-resistant, and age-appropriate.

Our LLMs prioritize local content, low hallucinations, efficacy, privacy and safety, and efficiency as core tenets of their existence:

1. Local content: Education institutions, school leaders, and teachers make thoughtful strategic choices on the content and curriculum they use to best help students achieve defined learning objectives. Our AI platform is built for this reality with a solution that draws from the school’s chosen corpus to overcome hallucinations and inaccuracies with a generative AI experience purpose-built for the unique needs of hyper-local education content.  
2. Low hallucinations: The occurrence of hallucinations or inaccuracies in state-of-the-art LLMs is unacceptably high for use in education. Education is a domain where a premium is placed on highly accurate responses to user prompts. We use multiple techniques to ensure that our approach and platform provide responses that are as hallucination-free as possible.
3. Efficacy: Backed by educational research, our models are designed to boost the capabilities of classroom leaders using principles such as Socratic dialogue, share-of-voice, and differentiated instruction. 
4. Privacy and safety: Our platform is equipped with models and filters tuned to check the safety of inputs and outputs from the language models to assure age-appropriate outputs. We understand the importance of protecting students’ and staff’s personal information and the responsibilities we have under the Family Educational Rights and Privacy Act (FERPA), the Children’s Online Privacy Protection Act (COPPA), and applicable U.S. state student data privacy laws. We implement technical, administrative, and physical information security safeguards to help protect the confidentiality, availability, and integrity of personal information in its care. 
5. Efficiency: Our LLM platform is designed to lower costs, latency, and environmental impact. It achieves this by drawing from multiple LLMs simultaneously, each tuned to a specific task or set of tasks. 

We also will power Merlyn, our AI assistant now in use in thousands of classrooms, with this generative AI platform. 

“Our mission has always been to improve the lives of teachers and students, and ultimately learning outcomes, by bringing the latest advances in AI to education,” says Dr. Satya Nitta, co-founder and CEO of Merlyn Mind. ”This platform is a major progression toward that goal. We’re enabling anyone in the broader edtech ecosystem to build applications that are safe, private, and tuned to the hyper-local content of their users.

“We believe in the power of the open source movement and the need to work together to get generative AI right for education. We are delighted to release three of our models to the community for use without restrictions. The models are trained on improving safety during LLM operation in education and trained on education-specific tasks like question answering, and assessment generation. Generative AI will transform education far more significantly than most other fields, and accessing its power in a safe and curriculum-aligned manner is crucial for improving learning outcomes.” 

Rob Hutter, founder of edtech venture fund Learn Capital and the Executive Chairman of Merlyn Mind, agrees. “The arrival of large language models has profound implications for teaching and learning, and it’s crucial that the education sector gains access to distinct models that are tailored to the high-stakes needs of educating people with accuracy, safety, and effectiveness. Merlyn is a rare company doing deep tech in education, and they are capable of building an LLM platform for the space while contributing some of their key models to open source for the entire education community to use.”

The future of AI

We envision that we will live in a world where every learner, educator, and professional will have their own AI assistants, which will be capable of remembering their preferences and learning patterns as well as performing related actions within the tools and resources they use. Our Merlyn assistant is a step toward realizing this vision. 

To build useful and generally intelligent systems, we are working at the leading edge of AI and with the AI research community to develop models and platforms that will plan and reason, have long-term memory, have better models of human cognition, have a better understanding of the world, and will perform actions on behalf of their users. We believe that over time, generative AI will transform the world as significantly as the birth of the internet itself has. 

However, like other AI advances, the most meaningful solutions in a vertical domain (like education) will come when teams develop AI in a purpose-built way. These platforms and solutions will be imbued with a deep awareness of domain-specific workflows and needs and will understand specific contexts as well as domain-specific data. When these conditions are met, generative AI will utterly transform industries and segments, ushering in untold gains in productivity and enable humans to reach their highest potential.

To learn more about how we built our open-source large language models, visit Merlyn Mind’s AI Labs.

And if you'd like to try them for yourself, you'll find them on HuggingFace.

Merlyn Mind Appropriateness model

Merlyn Mind Corpus-qa model

Merlyn Mind Teacher-assistant model

‍