Ideas & Writeups

A data-generation plan for the long-context mid-training stage: what to source, how to synthesize it (OLMo-style bootstrap vs. prompt-driven), the staged curriculum, and the multilingual bet.

A generate-then-judge recipe for multilingual synthetic instruction data on OLMo-3-7B, benchmarked against sourced, translated, and our own earlier synthetic baselines under three independent judges. Fullscreen slide deck — arrow keys or swipe to navigate.