Reproduce “Training Foundation Models with a Single Algorithm” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yuda Song, Alvin Zhang, Fahim Tajwar, Daman Arora, Jeff Schneider, Daniel Khashabi, Ruslan Salakhutdinov, J. Andrew Bagnell and 2 more
The authors cast pretraining, midtraining, supervised fine-tuning, and suitable reinforcement-learning updates as maximum-likelihood training with a changing data mixture. In reasoning experiments, unified schedules reach higher accuracy and coverage at larger tested compute budgets, though staged training can lead at intermediate budgets. After coding updates, a 0.76B model with 5% midtraining-data mixing scores 58.47% pass@16 versus 52.12%, limited to this continuation setup.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.