The challenge
As training data becomes increasingly saturated, the need for high-quality, complex STEM datasets has never been greater. One of our clients, a leading tech and AI company operating in the social media space, sought a diverse and complex dataset to refine their cutting-edge chatbot renowned for offering unfiltered answers with advanced reasoning, coding and visual processing capabilities. To elevate the chatbot's proficiency, with a specific focus on enhancing its ability to perform logical inference, quantitative analysis and causal reasoning in scientific contexts, they required a highly complex multimodal STEM dataset that provides richer contextual information beyond traditional prompt-response pairs (PRPs).
The TELUS Digital solution
TELUS Digital’s off-the-shelf STEM dataset was customized specifically to train and fine-tune their AI model for scientific and mathematical reasoning. The dataset includes over 30,000 high-quality multimodal PRPs in English, meticulously curated to cover various sub-topics across physics, mathematics, chemistry and biology. Sourced from a global network of subject-matter experts, including researchers and academics, the PRPs are designed to reflect real-world scientific challenges, ensuring relevance and usability.
The dataset features both descriptive and non-descriptive problems ranging in difficulty from “hard” to “very hard”, which ensured that the AI model was exposed to a broad spectrum of complex challenges. The dataset comprises:
- 4,485 physics PRPs, including 1,914 featuring images, with 21.2% very hard and 78.8% hard questions.
- 9,606 math PRPs, including 2,428 featuring images, with 27.8% very hard and 72.2% hard questions.
- 8,341 chemistry PRPs, including 2,393 featuring images, with 25% very hard and 75% hard questions.
- 7,771 biology PRPs, including 2,460 featuring images, with 34.2% very hard and 65.8% hard questions.
The results
By licensing our pre-built dataset, the client eliminated the need for time-consuming data creation, enabling swift integration into their training pipeline. This significantly reduced project timelines and facilitated faster iteration cycles, accelerating their development process. The multimodal dataset, with its rich integration of text and visual data, empowered the client's chatbot model with enhanced image deciphering capabilities, markedly improving its ability to solve challenging scientific problems. Furthermore, the dataset served as a valuable evaluation step in their model development, allowing the customer to measure the AI model's progress in handling complex multi-step reasoning tasks, particularly those that have proven challenging for even the most advanced state-of-the-art models.
