Creating a Reasoning Dataset with No Budget.

Author(s): Akhil Theerthala

Originally published on Towards AI.

Building a dataset. Image generated using Imagen.

In the pursuit of understanding reasoning models, the first major roadblock anyone would encounter is finding the right problem to solve. Despite the general understanding of the training regime, we need to solve a problem end-to-end to understand the hidden intricacies of developing a reasoning model. That’s how I started this journey anyway.

The following series of articles is a log of everything I had to learn to tune a reasoning model into solving a problem I’d usually like to avoid. This discussion aims to summarise how I chose a problem to solve and how I generated the dataset.

What problem to choose?a. What are a few possible areas that are interesting?b. Which of these areas has been explored the least?c. Why choose the least-explored areas?What kind of datasets are available on the World Wide Web?a. What artefacts are missing?Steps in Generating the Dataa. Getting the Answers: Plutus-v2 and the Llama-3.2–1B model.b. Why should we do this? Why should we use the outputs as answers?c. The next steps — The crux of the whole dataset.d. What are the resources I need? Where can I get that compute for cheap or even free?Areas of Improvement.Next StepsReferences

Well, there… Read the full blog for free on Medium.

Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor.

Published via Towards AI