The artificial intelligence (AI)-powered digital assistant platform Merlyn Mind has released a set of large language models (LLMs) designed for the education sector and made available under an open-source licence.
According to Merlyn, its LLMs will let educators and students to interact with generative models based on user-selected curriculum, resulting in a more enriching educational experience.
The LLMs are part of the company\’s generative AI platform for education, and they are able to engage with certain libraries of instructional materials.
To yet, no true LLMs focusing on education have been declared. General-purpose LLMs (most interface with OpenAI) are used by certain educational services, but they have the same problems we\’ve been talking about (hallucinations, lack of ironclad safety, privacy issues, etc.),\’ says Satya Nitta, CEO and co-founder of Merlyn Mind. But our generative AI platform and LLMs are the first to be designed specifically for educational purposes.
Nitta claims that conventional LLMs are educated using massive quantities of online data to provide conclusions and suggestions. These answers may not meet curriculum standards. In contrast, Merlyn\’s LLMs don\’t utilise the wider internet at all, instead relying purely on academic corpora selected by users or institutions.
\”As education institutions, school leaders, and teachers make thoughtful strategic choices on the content and curriculum they use to best help students, Merlyn\’s AI platform is built for this reality with a solution that draws from the school\’s chosen corpora to overcome hallucinations and inaccuracies with a generative AI experience,\” Nitta said.
The Merlyn voice assistant provides access to the generative AI platform with an emphasis on education for both teachers and pupils. Users may ask Merlyn questions aloud in the classroom, or request that it create quizzes and exercises depending on what is being discussed.
The platform also provides tools for educators to create curriculum-aligned materials including presentations, lesson plans, and assessments.
Getting rid of delusions to reveal the truth in the classroom.
Current state-of-the-art LLMs, as pointed out by Nitta of Merlyn, often produce erroneous answers, sometimes known as hallucinations. Even though it\’s a significant step forward, OpenAI\’s GPT-4 still has hallucinations around 20% of the time.
He stressed the need of correct and exact answers in the classroom, as user prompts must draw from certain material sources. The organisation uses a variety of methods to assure consistent and accurate replies and to reduce the likelihood of hallucinations occurring.
The LLM starts by locating the most relevant sections from the material utilised by the school district or instructor for instruction when a user makes a request, such as asking a question or providing a command to create evaluations. The data is subsequently fed into the language model.
The model creates answers based just on the input, without consulting any of its stored knowledge. The answer is double-checked by a different language model to make sure it accurately reflects the original query.
According to Merlyn, the main model has been fine-tuned such that it confesses failure rather than producing a bogus answer when it cannot create a high-quality response.
\”Hallucination-free responses, with attribution to the source material are commensurate with the need to preserve the sanctity of information during teaching and learning,\” stated Nitta. Our method has shown that we hallucinate fewer than 3% of the time, and we are on track to achieve our target of 100% accuracy without hallucinations.
Safety, efficiency, and effectiveness
According to the firm, it follows strict privacy standards that meet the legal, regulatory, and ethical needs of educational institutions. Among them are the General Data Protection Regulation (GDPR), the Children\’s Online Privacy Protection Act (COPPA), and the United States\’ own Family Educational Rights and Privacy Act (FERPA). Merlyn promises unequivocally that customer data is never shared with other parties.
Conversational experiences and transcripts are screened for personally identifiable information (PII) and any found is removed. Text transcripts of spoken recordings are deleted either six months after creation or ninety days after the client contract is terminated, according to Nitta. De-identified data produced from text transcripts is all that we save and utilise to enhance our services and for other legitimate reasons.
The business claims that their LLMs designed specifically for use in schools are more compact and efficient than competing general-purpose devices. While standard general-purpose models often include over 175 billion parameters, Merlyn\’s ranges from 6 billion to 40 billion.
Nitta further pointed out that, in comparison to general-purpose models, LLMs are much more effective in both training and operation (inferencing).
Merlyn\’s LLMs have an average delay of roughly 90 ms per [produced] word, but bigger models often take 250 ms or more. \”This is a huge benefit if an LLM or multiple LLMs need to be used sequentially to answer a user query,\” he said. Compared to Merlyn\’s LLMs, \”using a 175-billion-parameter [model] three times in succession can lead to unreasonable long latencies, poor user experience, and much less efficient use of computing resources,\” the authors write.
Possibilities for LLMs to advance in academics
According to Nitta, generative AI can drastically improve classroom instruction. However, it must be utilised properly, with security and precision as top priorities.
We invite the developer community to take use of these models by downloading them and using them to ensure the robustness of their LLM answers. In addition to our voice assistant, Merlyn is also accessible via a chatbot interface that provides multimodal (including aligned picture) responses, and we have been asked to make Merlyn accessible through an API. We are releasing several of our learning management systems (LLMs) as open source for the benefit of users with a technical background.
In the same vein as previous AI developments, he said, the most significant solutions in certain areas, like education, come when teams create AI technology on purpose.
Specifically, \”these platforms and solutions will be imbued with a deep awareness of domain-specific workflows and needs and will understand specific contexts and domain-specific data,\” as Nitta put it. To to the author, \”when these conditions are met, generative AI will utterly transform industries and segments, ushering in untold gains in productivity and enabling humans to reach our highest potential.\”