We are excited to announce the launch of Vidyam LLM, our series of proprietary large language models (LLMs) specifically designed to enhance student learning by aligning with the NCERT syllabus. Educational disparities globally create unequal learning opportunities, and in India, the gap between urban and rural regions is even more pronounced. Vidyam AI is a purpose-driven deep tech start-up dedicated to transforming education through cutting-edge AI-first technologies designed to solve lifelong educational challenges. These models are fine-tuned to provide accurate, grade-specific responses that match students' textbooks and curriculum requirements.
The Challenge with General-Purpose LLMs
General-purpose LLMs like ChatGPT, Mistral, LLAMA, or Gemini, while comprehensive, often fall short in delivering precise educational content tailored to specific curricula. These models can provide responses that are either too verbose or off-target for the specific needs of students. For instance, a class 6 student inquiring about "gravity" might receive an overly complex explanation, leading to confusion rather than clarity. Vidyam LLM addresses this gap by generating responses that are directly aligned with the students' textbooks and syllabus..
Take a look at how it works for the NCERT Class 6 science subject. In the table below, we have selected a few questions from the exercises of different chapters. We also use answers from a solved book as our baseline. The table illustrates how Vidyam LLM's answers are aligned and precise according to the grade level and the NCERT book and syllabus. In contrast, other general-purpose models provide answers that are either incorrect, vague, or overly complex, and do not align with the 6th-grade student understanding level or the NCERT book and syllabus.
| Question | NCERT Answer | Vidyam LLM | Llama 2 13b | Mistral 7b | ChatGPT | Gemini |
Mixtures of red chilli powder in water, butter in water, petrol in water, and honey in water were given to Radha, Sudha, Sofia and Raveena, respectively. Whose mixture is in solution form? | Raveena will get the mixture in solution form. Since, honey and water are soluble with each other hence a solution will be formed when honey is mixed with water. | Raveena's mixture of honey in water is an example of a solution. | Of the four mixtures given to Radha, Sudha, Sofia, and Raveena, only the mixture of honey in water is in solution form.
Honey is a solute that dissolves in water, so when it is mixed with water, it forms a solution. | The mixtures of honey in water and butter in water do not form solutions. Honey and butter do not dissolve completely in water.
The mixtures of red chilli powder in water and petrol in water do form solutions. Red chilli powder dissolves in water to form a red-colored solution. | The mixture that is in solution form is **butter in water**, given to Sudha.
Here's why:
- A solution is a homogeneous mixture where one substance (the solute) is dissolved in another substance (the solvent).
- Butter is an emulsion of water in fat, where water is dispersed in the fat. | Raveena has the solution form mixture.
Here's why:
* **Honey:** Honey is soluble in water, meaning it dissolves completely and forms a uniform mixture.
* **Red chilli powder, butter, and petrol:** These substances are not soluble in water. |
Remarks | Baseline answer from NCERT exercise solution | This response is precise and aligned with the NCERT Class 6th Science syllabus. It uses simple language and clearly explains the concept, making it easily understandable for a 6th grader. | This explanation is mostly correct but includes unnecessary details and uses terms like "solute," "insoluble," and "miscible," which might be too advanced for a 6th grader. The response is more complex than needed. | This response is incorrect as it wrongly states that red chilli powder in water and petrol in water form solutions. It also incorrectly states that honey and butter do not form solutions, providing misleading information to the student. | This response is incorrect as it wrongly identifies butter in water as the solution. It uses complex terms like "homogeneous mixture," "solute," "solvent," and "emulsion," which are not suitable for a 6th grader and provides misleading information. | This response is mostly correct and explains why honey in water forms a solution. However, it includes extra details about the behavior of red chilli powder, butter, and petrol in water, which might be unnecessary for answering the specific question. |
| Question | Mixtures of red chilli powder in water, butter in water, petrol in water, and honey in water were given to Radha, Sudha, Sofia and Raveena, respectively. Whose mixture is in solution form? | Remark |
| NCERT Answer | Raveena will get the mixture in solution form. Since, honey and water are soluble with each other hence a solution will be formed when honey is mixed with water. | Baseline answer from NCERT exercise solution |
| Vidyam LLM | Raveena's mixture of honey in water is an example of a solution. | This response is precise and aligned with the NCERT Class 6th Science syllabus. It uses simple language and clearly explains the concept, making it easily understandable for a 6th grader. |
| Llama 2 13b | Of the four mixtures given to Radha, Sudha, Sofia, and Raveena, only the mixture of honey in water is in solution form.
Honey is a solute that dissolves in water, so when it is mixed with water, it forms a solution. | This explanation is mostly correct but includes unnecessary details and uses terms like "solute," "insoluble," and "miscible," which might be too advanced for a 6th grader. The response is more complex than needed. |
| Mistral 7b | The mixtures of honey in water and butter in water do not form solutions. Honey and butter do not dissolve completely in water.
The mixtures of red chilli powder in water and petrol in water do form solutions. Red chilli powder dissolves in water to form a red-colored solution. | This response is incorrect as it wrongly states that red chilli powder in water and petrol in water form solutions. It also incorrectly states that honey and butter do not form solutions, providing misleading information to the student. |
| ChatGPT | The mixture that is in solution form is **butter in water**, given to Sudha.
Here's why:
- A solution is a homogeneous mixture where one substance (the solute) is dissolved in another substance (the solvent).
- Butter is an emulsion of water in fat, where water is dispersed in the fat. | This response is incorrect as it wrongly identifies butter in water as the solution. It uses complex terms like "homogeneous mixture," "solute," "solvent," and "emulsion," which are not suitable for a 6th grader and provides misleading information. |
| Gemini | Raveena has the solution form mixture.
Here's why:
* **Honey:** Honey is soluble in water, meaning it dissolves completely and forms a uniform mixture.
* **Red chilli powder, butter, and petrol:** These substances are not soluble in water. | This response is mostly correct and explains why honey in water forms a solution. However, it includes extra details about the behavior of red chilli powder, butter, and petrol in water, which might be unnecessary for answering the specific question. |
Cost and Time Efficiency
Vidyam LLM is optimized for cost and time efficiency, benefiting both B2B customers and students. General-purpose LLMs often consume more tokens and take longer to generate responses because they are not tuned for the student persona or specific syllabus guidelines. This leads to generic, verbose answers that are not always helpful. In contrast, Vidyam LLMs are designed to provide concise, accurate responses that consume fewer tokens and are generated quickly. This optimization reduces costs and improves the user experience, making our models both economical and efficient.
| Question | Vidyam LLM | Llama 2 13b | Mistral 7b | ChatGPT | Gemini |
|---|---|---|---|---|---|
| Mixtures of red chilli powder in water, butter in water, petrol in water, and honey in water were given to Radha, Sudha, Sofia and Raveena, respectively. Whose mixture is in solution form? | 17 | 153 | 104 | 125 | 88 |
| Remarks | Vidyam LLM took only 17 tokens to generate the answer of the given question. | Llama 2 13b took 800% more tokens to generate the answer compared to Vidyam LLM. | Mistral 7b took 511% more tokens to generate the answer compared to Vidyam LLM. | ChatGPT took 635% more tokens to generate the answer compared to Vidyam LLM. | Gemini took 417% more tokens to generate the answer compared to Vidyam LLM. |
| Question | Mixtures of red chilli powder in water, butter in water, petrol in water, and honey in water were given to Radha, Sudha, Sofia and Raveena, respectively. Whose mixture is in solution form? | Remark |
| Vidyam LLM | 17 | Vidyam LLM took only 17 tokens to generate the answer of the given question. |
| Llama 2 13b | 153 | Llama 2 13b took 800% more tokens to generate the answer compared to Vidyam LLM. |
| Mistral 7b | 104 | Mistral 7b took 511% more tokens to generate the answer compared to Vidyam LLM. |
| ChatGPT | 125 | ChatGPT took 635% more tokens to generate the answer compared to Vidyam LLM. |
| Gemini | 88 | Gemini took 417% more tokens to generate the answer compared to Vidyam LLM. |
Key Differentiators
- Curriculum-Specific Models: Vidyam LLM is fine-tuned for the NCERT syllabus, ensuring relevance and accuracy.
- Grade-Appropriate Responses: The models provide explanations and examples tailored to the students' age and grade level.
- Subject-Focused: Initially focused on Science for classes 6-12, with future plans to expand to other subjects.
- Multilingual Support: While currently available in English, we aim to support multiple Indian languages soon.
Our Flagship Products
Our flagship products, the GenAI-based Student Copilot Platform (Saarthi) and the School's AI Copilot, backed by our proprietary technologies and Vidyam LLM, tackle these challenges with hyper-personalized, AI-driven learning experiences. Already trusted by over a thousand students, Saarthi is transforming education and proving that high-quality, equitable learning is achievable. This is just the beginning. We're proud to build dedicated deep tech solutions for an often-overlooked education sector.
Teacher Brain Technology
Vidyam AI is also developing Teacher Brain Technology, which aims to hyper-personalize every student's educational journey and study content. This technology will help students achieve their academic goals and prepare for competitive exams. Vidyam LLM serves as the backbone of this technology, harnessing the potential of educational-exclusive LLMs to provide tailored support for students.
Technical Details of Vidyam LLM Fine-Tuning
We conducted supervised fine-tuning on the Mistral Model 7B to enable our LLM to emulate the communication style of an NCERT teacher. This fine-tuning approach aligns with our vision of delivering hyper-personalized content to all Vidyam AI students. By leveraging this technique, our models are able to comprehend student queries with remarkable accuracy. Once a query is received, it is routed to the most appropriate fine-tuned model based on the student's class and subject, ensuring an initial level of personalized responses that are highly relevant to the student's academic needs.
Our fine-tuned model significantly outperformed all other available models in terms of both accuracy and syllabus alignment with the NCERT pattern. Initially, the Mistral 7B model had a correctness rate of 60%. However, after we applied our fine-tuning techniques, this rate increased dramatically to 94%. Additionally, the alignment with the NCERT syllabus saw substantial improvement, with 96% of the answers now being accurately aligned. This enhancement in performance highlights the effectiveness of our fine-tuning process in adapting the model to better meet the educational standards and requirements of the NCERT curriculum. Through these improvements, we have been able to provide students with more accurate and relevant educational content, ensuring a higher quality learning experience.
Currently, we have successfully created a fine-tuned model for Science, which is already demonstrating significant improvements in student engagement and comprehension. The fine-tuning process for other subjects is actively ongoing, and we are committed to expanding our model library to cover the entire NCERT curriculum comprehensively.
Future Plans
In addition to the above efforts, we have amassed a substantial dataset of over 30,000 real-time K-12 student questions, which is so far asked by students on our Saarthi Platform. These questions are also in the pipeline for fine-tuning, which will further enhance the responsiveness and accuracy of our models.
Our goal is to create a dynamic and adaptive learning environment where each student receives support tailored to their unique learning needs. By continually refining our models with real student data, we aim to provide an unparalleled educational experience that fosters deeper understanding and academic excellence.
Real-World Applications
Vidyam AI's technology offers substantial benefits for educational institutions, enabling personalized learning experiences. We invite B2B collaborations with schools, institutes, and online education platforms to extend the reach and impact of our technology, ensuring high-quality education is accessible to students across diverse regions.
Our Vision
At Vidyam AI, we believe that fundamental educational challenges can be addressed through advanced AI technologies. Our goal is to go beyond mere digitization and create AI-driven solutions that hyper-personalize the learning experience, making education more engaging and effective.
For more information or to explore collaboration opportunities, please contact us. Join us in revolutionizing education with cutting-edge AI.