Goodfire's Predictive Data Debugging Forecasts Post-Training Behavior Changes Before Spending Compute

allen_ai · x · 2026-09-10

Ai2 and Goodfire detail a new post-training research result: teams usually discover unwanted behavior changes only after training finishes, then guess which of hundreds of thousands of examples caused them. Goodfire built "predictive data debugging" on Ai2's open stack — estimating which behaviors preference training will strengthen or suppress before committing compute. The open stack includes the Dolci preference dataset for Olmo 3, intermediate checkpoints with reproducible recipes, and OLMES evals for measuring capability shifts.

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

Original post →

More from coding & agent

coding & agent channel →