Reproduce “Bi-Phase Training: Learning Efficiently From Scratch” 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
Abel Gurung, Joseph Campbell
Bi-Phase Training (BPT) reduces trainable parameters in language-model pretraining by keeping most base weights fixed, periodically merging low-rank updates, and persistently rescaling rows and columns. In experiments on OLMo-2-1B and Qwen-2.5-1.5B, its evaluation-loss curves track full training closely over 10 billion tokens, while ReLoRA finishes higher. These results cover two tested models, not all models or tasks.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.