AvataRL: Adaptive Data Spaces and Sample Efficient Language Modelling

• ​How can we keep training AI models when high-quality pretraining data starts to run out?

• ​⁠What we can learn from AvataRL - using a smaller mentor model to guide what the main model should learn next.

• ​Training a model that not only learns from data but learns how to learn by steering its own practice sessions.

• ​Why uniform training is wasteful and how adaptive data selection helps make every batch count.

• ​How the mentor chooses examples with the most learning signal and avoids wasting steps on what the model already knows.

• ​⁠Balancing batches to stay diverse and avoid repetitive data, leading to more robust learning outcomes.

​About the speaker: Abhishek Mishra (aka tokenbender), Researcher at eXperiment Labs https://x.com/tokenbender || https://tokenbender.com

​Preread: https://tokenbender.com/post.html?id=avatarl

​To attend online, please use the link below:
https://meet.google.com/awd-hfrj-xur?hs=122&authuser=0