A Hitchhiker’s Guide to Good Research

​Subtitle: Lessons from Post-Training Multilingual and Multimodal Models

​Details:

• ​⁠In this talk, we will be covering good research principles, using concrete stories from Aya and Aya vision models.

• ​⁠⁠How ideas, data, infrastructure, and evaluation interact in practice when training and shipping models.

• ​How to reason about post-training pipelines, data mixtures, model capacity, and honest evaluation design.

• ​How to turn messy experiments into crisp papers with strong framing, baselines, and ablations.

• ​Practical principles for doing good research, from choosing the right problems to running disciplined experiments and collaborating well.

​About the speaker:
Viraat Aryabumi,
Member of Technical Staff at Cohere, previously built large-scale ad systems at Amazon, MSc in AI from the University of Edinburgh.
/viraataryabumi || /viraataryabumi

​
To attend online:

• ​⁠Add to calendar: https://bit.ly/3MlI6sw

• ​Gmeet link: meet.google.com/tgo-vetw-txw