Building OpenThinker: a reasoning model that beat DeepSeek!

​Reasoning has unlocked a whole new paradigm of improving AI capabilities. When DeepSeek announced and open-sourced their R1 reasoning model and the distilled 7B and 32B variants, they did not release the corresponding dataset that was used to train these models. OpenThoughts is a community effort led by Bespoke Labs to curate the best reasoning dataset. As part of this, we have released several datasets, and the corresponding distilled models, with the latest being OpenThoughts2-1.2M dataset and the OpenThinker3-7B model which beats DeepSeek-R1-7B in several reasoning benchmarks. The OpenThinker models have been downloaded more than 500k times and the datasets have been #1 on HuggingFace trending datasets.

In this talk, we will dive into the details of the learnings we have had in curating OpenThoughts3, and also share details on what worked and what didn't. We will discuss how to adapt this recipe for your own domain, and briefly touch upon what the future of reasoning looks like.

​About the Speaker:
Mahesh Sathiamoorthy
x: https: https://x.com/madiator
LinkedIn: https://linkedin.com/in/smaheswaran 
Current: Co-founder and CEO at BespokeLabs.AI 
Prev: Staff Software Engineer at Google DeepMind