Great to be part of this project led by the amazing @hamishivi.bsky.social. The most fun (in retrospect) thing is to observe how the results start to shift as we scale up the candidate pool, evaluation suite, and selection size :) And eventually we find a simple method does the best!

How well do data-selection methods work for instruction-tuning at scale?

Turns out, when you look at large, varied data pools, lots of recent methods lag behind simple baselines, and a simple embedding-based method (RDS) does best!

More below ⬇️ (1/8)

This is a great effort for the migration, thanks for putting it together! Can I be added to the list?