Ai2's Open Stack — Dolci Data, Checkpoints, OLMES — Enabled Goodfire's Behavior Prediction

allen_ai · x · 2026-09-10

Goodfire wanted to move post-training debugging earlier, and Ai2's open stack made it possible: the Dolci dataset provides Olmo 3's preference data, Olmo ships intermediate checkpoints plus reproducible recipes, and OLMES measures capability changes. The thread note underscores how open data and checkpoints enable third-party interpretability research that closed stacks can't support.

Related event: Ai2 and Goodfire unveil predictive data debugging for pre-training behavior forecasting(7 posts)→

Original post →

More from Research

Research channel →